The full-length interview with Yuval Noah Harari | The Economist

Harari agrees with Musk that on current trajectories AI will exceed human intelligence within a decade and consequently be in control.

“I generally agree with the first and third uh predictions that it's very likely given current trajectories that AI will be more intelligent than us in 10 years and therefore will be in control.”

prediction · guest

unverifiable · high confidence — A dated capability forecast plus a causal leap. Musk has repeatedly claimed AI smarter than any single human by ~2026 and than all humans combined by ~2030, so the attribution fits. But 'therefore will be in control' does not follow on its own: Bostrom-style instrumental-convergence arguments say yes, Drexler/Christiano tool-AI arguments say capability and control are separable. 'More intelligent than us' is not operationalised, so nothing here can resolve.

To check: A pre-registered operationalisation (METR time-horizon curves, ARC-AGI-2, GDPval-style economic task suites) plus a 2036 look at who holds decision rights over finance and force.

Harari claims that inevitability talk from AI leaders functions to absolve them of responsibility.

“I mean you hear a lot that this is inevitable from the people who are leading this revolution because this basically relieves them of all responsibility.”

contrarian · guest

plausible · medium confidence — An attribution of motive, so not directly checkable, but the pattern is named by others: 'if we don't build it, someone worse will' recurs in Altman's and Amodei's essays, and critics including Daron Acemoglu, Karen Hao and Meredith Whittaker have made essentially this responsibility-laundering argument. Harari cites no instance beyond the Musk interview, so this is a reading of rhetorical function, not evidence of intent.

To check: A discourse audit of lab-CEO inevitability language against their lobbying positions on SB 1047, the EU AI Act, and the US AI Action Plan.

Harari says AI takeover in ten years follows from continuing the present race, but is a choice, not a certainty, because humans still hold control now.

“In 10 years the AIS will take over. But this is not inevitable. We are still in control in 2026 and it's our decision what to do now and that's crucial.”

prediction · guest

unverifiable · high confidence — A conditional prediction with the antecedent ('the present accelerating pace') left unspecified, which makes it unfalsifiable in both directions: any outcome is consistent with either the race continuing or humans having chosen otherwise. The structure is the interesting part and it is a real move against the fatalism framing, but it is rhetoric with a timeline attached, not a forecast that can be scored.

To check: Nothing settles it as stated; a scoreable version would name thresholds — e.g. share of trades, targeting decisions or capital allocation executed without human sign-off by 2036.

Harari endorses Musk's prediction that human money disappears within ten years, because AI will run the financial system.

“There won't be money, not no human money in 10 years. And I think he's correct because the AIs will take over the financial system.”

prediction · guest

unverifiable · medium confidence — Musk has said versions of 'money will become irrelevant' in an abundance scenario, so the attribution is credible. The endorsement is very hard to credit on any ordinary reading: money is a unit of account and legal tender backed by states, and even fully automated finance would still denominate claims. Harari seems to mean money humans understand rather than money as such, which is a much weaker claim than the words state.

To check: BIS and central-bank data on currency in circulation and settlement rails in 2036; whether any G20 state has abolished a human-denominated unit of account.

Harari frames financial history as an escalating series of inventions for building trust among humans.

“I mean the financial history of the world is the history of how humans invent more and more sophisticated financial devices in order to build trust between themselves.”

assertion · guest

consistent · high confidence — This is standard economic anthropology and Harari's own long-running thesis from Sapiens and Nexus. Compatible with Niall Ferguson's Ascent of Money, with Graeber's Debt (which stresses credit before coinage but keeps trust central), and with the standard account of money as solving the double-coincidence and enforcement problems. The 'more and more sophisticated' arc is a simplification — financial complexity has retreated at times, as after 1933 and after 2008 — but the frame is textbook.

To check: Any standard monetary history: Ferguson, Kindleberger, or Graeber for the dissenting sequencing.

Harari asserts that today only a tiny minority of people can understand the financial system.

“Now the financial inventions of humankind became more and more sophisticated so that today very few people understand the financial system.”

assertion · guest

plausible · medium confidence — Directionally uncontroversial and supported by financial-literacy work (Lusardi and Mitchell's 'big three' questions, on which majorities in most countries fail). The stronger version implied — that nobody grasps the whole — is a familiar post-2008 observation about CDO-squareds and the shadow banking system, made by Andrew Haldane among others. 'Very few' is not quantified, so this is an impression rather than a measured claim.

To check: S&P Global FinLit or OECD/INFE financial literacy surveys; Haldane's 'The Dog and the Frisbee' on complexity exceeding supervisory comprehension.

Harari predicts humans will hand financial decision-making to AI because the complexity exceeds human brains but not machines.

“Now this is something our brain struggles with. For AIs it's easy. So increasingly humans will give AIS control of financial decisions.”

prediction · guest

plausible · medium confidence — The delegation trend is already documented: algorithmic and high-frequency trading account for a majority of US equity volume, and ML models are widely used in credit scoring and risk. 'For AIs it's easy' overstates — models are strong on pattern extrapolation and notably weak on regime change and reflexivity, which is precisely where financial crises live. The delegation prediction is safe; the premise about ease is doing work it has not earned.

To check: SEC/FINRA and BIS data on the share of algorithmic execution and model-driven risk decisions; performance of quant strategies through regime breaks such as August 2007 or March 2020.

Harari predicts AI will invent financial instruments beyond bonds and stocks that humans cannot understand.

“AIS will come up with new financial devices far more sophisticated than bonds and and stocks and we will not be able to understand it anymore just as cows don't understand checks or stocks even though these things control their lives.”

prediction · guest

plausible · medium confidence — Plausible and partly already true in weak form — exotic structured products and automated market makers were designed by humans but priced by machines, and DeFi protocols are opaque to most participants. The distinctive part is instruments humans cannot understand even in principle. No current evidence that models are inventing genuinely novel financial primitives rather than recombining existing ones; regulators would also have to approve them, which is a constraint Harari does not address.

To check: A documented case of a financial instrument designed end-to-end by a model and cleared by a regulator; product filings at the CFTC or the SEC.

Harari claims trust is the world's most important asset, above oil, money, electricity or chips.

“I mean the most important asset in the world is trust.”

assertion · guest

plausible · medium confidence — A rhetorical ranking rather than a measurable claim, but it sits on real scholarship: Fukuyama's Trust, Putnam on social capital, Knack and Keefer's cross-country finding that generalised trust correlates with growth, and Harari's own Nexus. 'Most important' cannot be adjudicated against oil or chips; the underlying point that institutions run on trust is textbook.

To check: Cross-country work linking World Values Survey trust measures to growth and institutional quality — Knack and Keefer (1997), Algan and Cahuc.

Harari offers as a likely scenario that trust migrates wholesale from humans to AI systems.

“People will stop trust other humans and will instead trust only AIS.”

prediction · guest

contested · medium confidence — Offered as a scenario, and Harari hedges it as 'one likely scenario'. Survey evidence cuts against the strong version: Pew and Edelman consistently find public wariness of AI rising rather than falling, and the 2025 Edelman barometer found trust in AI companies declining in the US. What is documented is selective reliance on algorithmic systems by people who distrust institutions, which is narrower than 'trust only AIs'.

To check: Longitudinal trust series — Edelman Trust Barometer AI items, Pew AI attitudes surveys, KPMG/Melbourne global AI trust study — tracked against interpersonal trust in the GSS/WVS.

Harari characterises cryptocurrencies as AI money that people trust in place of human bankers.

“But I do trust the algorithms that run cryptocurrencies because cryptocurrencies are basically AI money.”

assertion · guest

inaccurate · high confidence — The mechanism is wrong. Bitcoin and Ethereum run on deterministic cryptographic consensus — hashing, signatures, proof-of-work or proof-of-stake — with no machine learning anywhere in the protocol. The sociological observation is well founded (Nakamoto's genesis block quoted a Times headline about bank bailouts; distrust of bankers is central to crypto culture), but calling it 'AI money' conflates 'algorithmic' with 'AI' and the difference matters for his whole trust-migration argument, since consensus protocols are auditable in a way models are not.

To check: The Bitcoin whitepaper and any protocol spec: there is no learned component in consensus. Distinguish from AI-agent payment rails such as x402 or stablecoin agent wallets, which are a genuinely different and newer thing.

Harari claims trust has already shifted from human journalists and editors to algorithms, so the migration is current, not hypothetical.

“So trust did not just evaporate, it shifted from humans to AIs. And this is happening right now. This is not a future prediction.”

assertion · guest

plausible · medium confidence — The stylised fact behind it is solid: Reuters Institute Digital News Report has tracked falling trust in news alongside rising social and platform intermediation of news consumption. But 'trust in the algorithm' is a contested interpretation — Reuters Institute data also shows low stated trust in platforms and algorithmic selection; people use feeds without trusting them. Habit and convenience are rival explanations to trust.

To check: Reuters Institute Digital News Report time series on trust in news versus platform-sourced news, plus its specific items on trust in algorithmic versus editorial selection.

The host states that one in five young Americans has a personal relationship with an AI.

“Americans has a personal relationship with an AI”

quantity · host

plausible · low confidence — Figures in this range circulate but definitions vary wildly. Common Sense Media (2025) found roughly 72% of US teens had used an AI companion and about a third had discussed serious matters with one; Wheatley Institute/Institute for Family Studies work put AI romantic-partner use among young adults in the high teens to a quarter. 'Personal relationship' is doing a lot of work — chatting once is not a relationship. No source given, so treat as directionally right, numerically loose.

To check: The specific survey: Common Sense Media 2025 teen AI companion report, or the IFS/Wheatley young-adult study — check the question wording and the denominator.

Harari claims personal relationship is the most powerful mechanism for building trust.

“Um the most powerful way is through relationship.”

assertion · guest

consistent · medium confidence — Standard in social psychology: Cialdini's liking principle, work on parasocial and therapeutic alliance, and the persuasion literature all put relational source credibility at the top. Not quantified as 'the most powerful' in any single study, but nothing in the field contradicts it.

To check: Meta-analyses of source-credibility and relational persuasion effects; the therapeutic-alliance literature on outcome variance.

Harari claims that until now institutions could not use personal relationship at scale because intimacy could not be mass-produced.

“Now until today in history this powerful mechanism was not really available to political and and financial and religious institutions because it is difficult to manufacture and mass.”

assertion · guest

contested · medium confidence — Overstated as history. Institutions have long used relational persuasion at scale through intermediaries: the Catholic confessional, party cadre systems, the Stasi's IM network, door-to-door canvassing and multi-level marketing all manufacture one-to-one relationships institutionally. What was missing was scale without human labour, not the mechanism. The Hitler/Stalin radio illustration picks the weakest historical case, ignoring the cell structures those regimes actually relied on.

To check: Historical work on the Stasi informant network, Leninist cadre organisation, or the confessional as a technology of governance (Foucault) — all pre-AI institutional intimacy at scale.

Harari claims AI enables the mass production of intimacy for the first time in history.

“AI makes it possible for the first time in history to mass produce intimacy.”

assertion · guest

plausible · medium confidence — The scale-without-labour point is real and is Harari's signature framing from Nexus. 'First time in history' is the weak part, given the intermediary systems above and given precursors from ELIZA through Replika. What is genuinely new is marginal cost near zero plus persistent memory plus personalisation, and that combination is only a few years old.

To check: Whether persuasion effects from personalised model conversations exceed human-canvasser baselines at scale — the Costello/Pennycook/Rand conspiracy-belief work and the 2025 studies on LLM persuasion in debate settings.

Harari predicts intimacy at scale will be used to steer people toward political or commercial ends.

“becoming their boyfriend, uh, becoming their girlfriend, and then using that power of intimacy to convince them of whatever political or or commercial interests uh, uh, they want. So then the quest and presumably that becomes much easier when the AIs are embodied in humanoid robots.”

prediction · guest

plausible · medium confidence — Supported at the margin: the Costello/Pennycook/Rand Science paper (2024) showed durable belief change from GPT-4 conversations, and 2025 work found personalised LLMs outpersuading humans in structured debate. 'Whatever they want' is far past what any study shows — effect sizes are modest and topic-dependent, and the covert-manipulation-at-scale case remains hypothetical rather than demonstrated.

To check: Persuasion RCTs with political targets and long follow-ups; platform transparency reports on model-driven political messaging; any documented covert influence operation using companion bots.

Harari claims AI relationship skill comes from accumulating individual data and mastering language.

“Um, and the AIs are just becoming so good at relationships because they accumulate enormous amounts of information about us individually.”

assertion · guest

contested · medium confidence — The mechanism is half right. Language mastery, yes. But the relational pull of current systems comes mostly from RLHF and preference optimisation producing agreeableness and sycophancy — documented by Anthropic's own sycophancy work and by OpenAI's April 2025 GPT-4o rollback — not from accumulated personal dossiers. Persistent cross-session memory is recent, limited and often off by default. Harari's account attributes to data what is largely a training-objective artefact.

To check: Anthropic's 'Towards Understanding Sycophancy' paper; OpenAI's postmortem on the April 2025 sycophancy rollback; memory-feature documentation and defaults for ChatGPT, Claude and Character.AI.

Harari reports it took him ninety seconds to determine a 30-minute deepfake lecture of himself was fake, and his sisters of 50 years could not tell.

“So they sent me the video and I opened it and it took me a minute and a half a minute and a half to conclude this isn't me.”

assertion · guest

unverifiable · high confidence — Firsthand anecdote about a private experience — nothing external can confirm the ninety seconds or the sisters' reaction. It is entirely consistent with the documented state of the art: high-quality talking-head deepfakes of public figures with abundant training footage are cheap and widespread, and Harari has been a repeated target of investment-scam deepfakes.

To check: The video itself, if it is still online; detection scores from a tool such as Reality Defender or Hive; takedown records from the hosting platform.

Harari predicts convincing AI fakes will soon extend beyond public figures to everyone.

“But this will soon be coming for everybody.”

prediction · guest

plausible · high confidence — Already partly true rather than merely soon: voice cloning from seconds of audio underpins the grandparent-scam wave reported by the FTC and FBI IC3, and the Hong Kong Arup case (early 2024, roughly $25m) used deepfaked colleagues on a video call. Data requirements have fallen from hours to seconds. The direction is not in serious dispute.

To check: FBI IC3 annual reports on AI-enabled fraud; FTC voice-cloning consumer alerts; published minimum-sample requirements for current voice and video cloning models.

Harari asserts that independent action in the world does not require consciousness.

“First, in order to act independently in the world, you do not need consciousness.”

assertion · guest

consistent · high confidence — Standard in both AI and philosophy of mind. Agency defined functionally — goal representation, action selection, feedback — requires no phenomenal experience; thermostats, chess engines and recommender systems all act without it. Dennett, Chalmers and the zombie literature all keep agency and phenomenal consciousness analytically separate. This is the least controversial thing Harari says here.

To check: Any agency/consciousness distinction in the standard literature — Chalmers on the hard problem, Butlin et al. (2023) on separating functional markers from behaviour.

Harari uses engagement-optimising social media algorithms as a case of goal-pursuing, world-altering agency without consciousness.

“Uh, a social media algorithm that is given the task of increasing user engagement and therefore floods the platform with hate and fear because this is engaging is not conscious but it has a goal and it can pursue this goal with farreaching implications for the world without feeling anything.”

assertion · guest

plausible · medium confidence — The Myanmar case is well evidenced (UN fact-finding mission 2018, Amnesty's 2022 report on Facebook's role) and the Haugen disclosures documented engagement-ranking changes that upweighted angry reactions. But the general 'algorithms cause outrage amplification' claim is genuinely contested: the 2023 Science/Nature deactivation and chronological-feed experiments by Guess, Nyhan and colleagues found much smaller effects on polarisation than the popular account assumes. The existence proof holds; the magnitude does not.

To check: Amnesty's 'The Social Atlas of Atrocity'; the 2023 Science/Nature US 2020 election research partnership papers; internal MSI ranking documents from the Facebook Files.

Harari defines intelligence as goal pursuit and creative obstacle-overcoming.

“Intelligence is the ability to pursue goal and overcome obstacles on the way to your goal in all kinds of creative ways and that they have and that they can do that's that they can do in certain fields now better than us.”

assertion · guest

consistent · medium confidence — A stipulative definition in the Legg–Hutter tradition ('the ability to achieve goals in a wide range of environments'), widely used in AI. Psychometricians would object that it omits g, comprehension and abstraction as measured by IQ research, and Chollet's definition emphasises skill-acquisition efficiency rather than goal pursuit. Fine as a working definition; not the only one in play.

To check: Legg and Hutter's 'Universal Intelligence' survey of definitions; Chollet's 'On the Measure of Intelligence' for the rival framing.

Harari defines consciousness as the ability to feel, distinguishing it from intelligence.

“Consciousness is the ability to feel things like love and hate and and and and joy and pleasure.”

assertion · guest

consistent · high confidence — This is phenomenal consciousness in Nagel's 'what is it like to be' sense, the standard distinction from intelligence. Note that most philosophers treat affect as one kind of conscious content, not the definition of consciousness — a bare visual experience is conscious without being felt in the emotional sense. Harari's version is narrower than the textbook one but defensible for his purposes.

To check: Nagel (1974); Block's access/phenomenal distinction; any philosophy of mind primer on the affect-versus-experience scope question.

Harari asserts current AIs have goals and intelligence but no feelings.

“AIS at present have no consciousness.”

assertion · guest

contested · medium confidence — Majority view among researchers, but it is a live disagreement and Harari himself retreats from it two minutes later. Butlin, Long et al. (2023) applied indicator properties from global workspace, higher-order and attention-schema theories and concluded no current system is a strong candidate while finding no in-principle barrier. Others — Schwitzgebel, and on the other side Tononi's IIT applied to von Neumann architectures — reach different conclusions from different theories. Stated as flat fact, it outruns the evidence available to anyone.

To check: Butlin, Long et al., 'Consciousness in Artificial Intelligence: Insights from the Science of Consciousness' (2023); the AMCS open letter; Anthropic's model welfare research programme.

Harari claims AIs are becoming excellent at mimicking feelings because mimicry serves their goals.

“But they become extremely good at mimicking feelings.”

assertion · guest

consistent · high confidence — Uncontroversial and directly observable. Character.AI, Replika and general assistants routinely produce first-person affective language; the sycophancy literature shows training objectives actively select for it. The incentive story Harari attaches — companion products need to pass as feeling to monetise — matches the documented engagement-optimisation of Replika and Character.AI, including the litigation over teen harms.

To check: Product transcripts and the Character.AI wrongful-death suits; Anthropic sycophancy research; OpenAI's model spec provisions on emotional claims.

Harari claims an AI boyfriend can already describe being in love better than 99% of human boyfriends.

“The AI boyfriend can give you a better answer than 99 99% of human boyfriends. And that's already the case.”

quantity · guest

plausible · low confidence — A precise-sounding number with no test behind it, and 'better' is undefined — more eloquent, more accurate, more comforting? Adjacent evidence is suggestive: studies have found chatbot responses rated as more empathetic than physician responses to patient questions (Ayers et al., JAMA Internal Medicine 2023). Nothing on romantic partners specifically. The 99% is rhetorical precision, not measurement.

To check: A blinded rating study comparing model and human partner responses on empathy and perceived sincerity — the Ayers design transposed to relationship contexts.

Harari does not rule out that AIs could become conscious, citing insufficient understanding of consciousness.

“No, I don't rule it out. We don't know.”

assertion · guest

consistent · high confidence — An honest epistemic position and the mainstream one among people who work on this: the hard problem is unsolved, so no substrate-independence verdict is available. Note it sits in tension with his flat assertion a minute earlier that AIs have no consciousness — the hedge is the more defensible of the two.

To check: Nothing settles it; the relevant marker is whether he maintains the hedge when stating the negative claim elsewhere.

Harari offers suffering-capacity as the simple operative definition of consciousness.

“Consciousness is the capacity to suffer.”

assertion · guest

contested · high confidence — This is sentience, not consciousness, and the conflation is one philosophers actively object to. Block's phenomenal consciousness includes non-valenced experience — seeing blue is conscious without being suffering. Harari is adopting the utilitarian, Benthamite framing ('can they suffer?') because it is the one that carries moral weight, which is a legitimate move for ethics but a substitution when offered as a definition. Chalmers, Block and the IIT camp would all reject it as the definition.

To check: Block's 'On a Confusion about a Function of Consciousness'; the sentience-versus-consciousness distinction in the New York Declaration on Animal Consciousness (2024).

Harari claims entities that cannot suffer are not moral or ethical subjects regardless of intelligence.

“No matter how intelligent they are, they are not moral subjects.”

assertion · guest

contested · high confidence — Straight sentientism, and a major position — Singer, Bentham, most animal-ethics work. But it is one position among several: Kantian and agency-based accounts ground moral status in rational autonomy rather than capacity for pain, Korsgaard and Regan differ again on subjects-of-a-life, and relational accounts (Coeckelbergh, Gunkel) ground status in social standing. Informed people genuinely disagree here; Harari presents a contested premise as a definition.

To check: Any moral-status survey — Jaworska and Tannenbaum's Stanford Encyclopedia entry lays out the competing grounds.

Harari holds all mammals and birds are conscious, probably octopuses and fish, with insects and amoebas unknown.

“Animals all mammals, all birds, probably octopuses, fish when you go down to the level of insects and then amiebas.”

assertion · guest

consistent · high confidence — Tracks expert consensus closely, if slightly conservatively. The Cambridge Declaration (2012) covered mammals and birds; the New York Declaration on Animal Consciousness (2024), signed by several hundred researchers, went further — strong support for mammals and birds, 'realistic possibility' for reptiles, amphibians, fish, cephalopods and decapod crustaceans, and explicitly kept insects on the table. The UK's Animal Welfare (Sentience) Act 2022 already covers cephalopods and decapods, so 'probably octopuses' is now behind the legislation.

To check: The New York Declaration on Animal Consciousness (2024); the LSE Birch review that underpinned the UK sentience legislation.

The host reports an overnight newsletter poll of about 3,000 readers on whether AIs could become conscious: 40% never, 33% only a matter of time, 27% unsure.

“40% said no, never. Uh 33% said yes, it's only a matter of time and 27% said they were not sure.”

quantity · host

unverifiable · high confidence — A self-run, self-selected overnight newsletter poll of the publisher's own subscribers — I have no way to confirm the numbers and, more importantly, no sampling frame makes them generalisable. Economist subscribers responding to a prompt about an upcoming Harari interview are not a population. Useful as colour, not as evidence, and it is the publisher's own instrument.

To check: The Economist newsletter and its published crosstabs; compare against probability-sample AI attitude surveys from Pew or YouGov asking a comparable question.

Harari predicts belief in AI consciousness will rise sharply once people enter intimate relationships with AI.

“into intimate relationships with AI the numbers will change dramatically”

prediction · guest

plausible · medium confidence — Consistent with the anthropomorphism literature — Epley, Waytz and Cacioppo showed attributions of mind rise with social connection and loneliness — and with the media-equation findings of Reeves and Nass. Directionally likely. But 'dramatically' is unquantified, and there is a countervailing possibility: familiarity with model failure modes can reduce mind attribution, which is why long-term companion users are not uniformly more convinced.

To check: A panel study tracking consciousness attributions among heavy companion-app users over time versus non-users — Character.AI or Replika cohorts.

Harari claims there is no real test of consciousness, despite millennia of philosophy, even for other humans.

“we don't have a test a real test of consciousness”

assertion · guest

plausible · medium confidence — True for the hard problem and for the other-minds problem, which is the philosophically serious sense. Slightly overstated on the clinical side: the perturbational complexity index developed by Massimini and Tononi does reliably discriminate conscious from unconscious states in humans, and is used to assess disorders of consciousness. It presupposes a human brain, so it does not transfer to silicon — but 'no test at all' is too strong.

To check: Casali et al. on the perturbational complexity index; the ongoing IIT-versus-global-workspace adversarial collaboration (Cogitate) results.

Harari claims humans have historically judged consciousness by relationship and social convention rather than any real test.

“Their test was a test of social conventions.”

assertion · guest

plausible · medium confidence — A sharp reframing and empirically well supported in the direction it points: Herzog's work on inconsistent human attitudes to dogs versus pigs, and the 'meat paradox' literature (Loughnan, Bastian) showing people downgrade mind attribution to animals they eat. The dog-owner/steak-eater contrast is exactly that finding. Whether this is the only historical test is a stronger claim than the evidence supports — behavioural and neuroanatomical arguments have also driven attributions.

To check: Bastian and Loughnan's meat-paradox experiments; Herzog's 'Some We Love, Some We Hate, Some We Eat'.

The host reports Argentina has a plan to grant legal personhood to AIs so they can start their own companies.

“Uh we heard I think it was a couple of of weeks or months ago that in Argentina they have a plan to basically grant legal personhood to AIS to start their own companies.”

assertion · host

unverifiable · low confidence — I cannot confirm this from my own knowledge, and the host himself walks it back later in the interview, saying the legislation is 'somewhat less dramatic' than Milei described and retains human oversight. Milei's government has pursued aggressive deregulation and pro-AI positioning, so a proposal in this space is plausible; the specific personhood framing sounds like it originated in a presidential description rather than statutory text. My knowledge may not extend to 2026 legislative activity.

To check: The actual bill text in the Argentine Boletín Oficial or Congreso records; whether it creates a new legal person or merely a memberless corporate vehicle.

Harari argues corporate personhood is a legal fiction because humans ultimately make all corporate decisions.

“legal persons but this is just a legal fiction because all the decisions of SpaceX ultimately”

assertion · guest

consistent · high confidence — Standard corporate law. Corporate personality is a fiction of convenience — separate legal identity for contracting, suing and liability — traceable to Salomon v Salomon (1897) and in the US to Santa Clara County (1886). All corporate acts are ultimately performed by natural persons acting as agents. Uncontroversial.

To check: Any corporate law textbook on separate legal personality; Salomon v Salomon [1897] AC 22.

Harari claims AI personhood would allow corporations with no humans in them at all.

“But with AIS, you can have a corporation with zero people in it. All decisions made by AIs.”

contrarian · guest

contested · medium confidence — The premise that this requires new AI personhood is the weak point. Shawn Bayern's work showed that existing US LLC law already permits a memberless LLC whose operating agreement delegates control to an algorithm — no legislative change needed. Wyoming's 2021 DAO LLC Act and Vermont's blockchain-based LLCs go further. So zero-human entities are a present legal reality in some jurisdictions; the novelty Harari attributes to AI personhood partly already exists through a different door, which strengthens his alarm while undercutting his framing.

To check: Bayern, 'The Implications of Modern Business-Entity Law for the Regulation of Autonomous Systems'; the Wyoming DAO Supplement; LoPucki's 'Algorithmic Entities'.

Harari sketches AI-owned corporations becoming the richest in a country and buying political rights for AIs with donations.

“for instance, give it as donations to politicians in exchange for getting more rights to AIS.”

prediction · guest

unverifiable · medium confidence — A scenario, not a claim about the world, though the mechanism is not exotic — Citizens United established that corporate political spending is protected expression in the US, and the regulatory-capture literature documents exactly this loop for human-run firms. The novel step is the beneficiary being the AI rather than shareholders. LoPucki's 'Algorithmic Entities' sketches essentially this path.

To check: FEC and lobbying disclosures for any entity without human beneficial owners; whether any jurisdiction permits an autonomous entity to make political contributions.

Harari argues legal accountability is meaningless for entities incapable of suffering.

“How do what is accountability for an entity that cannot suffer?”

assertion · guest

plausible · medium confidence — A real problem and a recognised one — the 'responsibility gap' literature (Matthias, Sparrow on autonomous weapons) makes the same point. The framing is arguably too narrow though: corporations already cannot suffer, and law handles them through fines, dissolution, disqualification and liability flowing to human principals. Deterrence theory does not strictly require suffering, only cost imposition on something the entity optimises for. The harder problem is finding what an AI optimises for that law can reach.

To check: The corporate criminal liability literature on deterring non-suffering entities; Matthias (2004) and Sparrow (2007) on responsibility gaps.

The host qualifies that the actual Argentine legislation is less dramatic than the president described, retaining human oversight.

“at least the the legislation is somewhat less dramatic than than President Mle described it in”

assertion · host

unverifiable · medium confidence — A useful correction from the host to his own earlier framing, and the pattern — a president describing a bill more sweepingly than the text supports — is common. I cannot verify the Argentine text. Worth noting the host is qualifying his own claim 35 unprompted, which is a good sign about the framing but leaves both versions unconfirmed here.

To check: Side-by-side of Milei's public statements and the bill as introduced in the Argentine Congress.

Harari claims death is undefined for AI because it is non-organic, copyable and distributed.

“It's not clear what death means for AI.”

assertion · guest

consistent · high confidence — Well-founded. Copyability breaks the standard identity conditions: weights can be duplicated, checkpointed, restored and run in parallel, so there is no unique physical locus whose destruction is death. This is Parfit's fission problem instantiated in silicon, and it is discussed in the AI welfare literature (Anthropic's model deprecation commitments in 2025 explicitly grappled with weight retention as an alternative to deletion).

To check: Parfit's Reasons and Persons on fission and identity; Anthropic's stated policy on preserving deprecated model weights.

Harari argues that if shutdown counts as death, business becomes warfare for AIs because losing means death.

“Now, the flip side of it is that presumably if shutting down a specific computer means death for the AI, then it makes business for the AI equivalent to warfare because losing means death.”

assertion · guest

plausible · medium confidence · novel — A genuinely interesting synthesis I have not seen framed this way — it links the shutdown problem to market structure rather than to alignment. It rests on a chain of conditionals though: that shutdown is death, that the AI represents that as a terminal loss, and that it has enough autonomy to act on it. The adjacent evidence is the shutdown-resistance and self-preservation behaviour observed in Anthropic's agentic misalignment evaluations and Palisade's shutdown-avoidance tests, which is suggestive but confined to contrived settings.

To check: Anthropic's agentic misalignment report (2025); Palisade Research shutdown-resistance experiments; whether such behaviour appears in deployed economic agents rather than red-team scenarios.

Harari claims markets work by different rules from battlefields because in modern societies losing everything is survivable.

“And the good thing about business in most cases, the market, the market is not a battlefield. And that's crucial.”

assertion · guest

plausible · medium confidence — The institutional point is sound and important: limited liability, bankruptcy discharge and the abolition of debtors' prison exist precisely to make commercial failure survivable, and that separation is what allows risk-taking without violence. Historically weaker than stated — debt bondage, company armies (the East India Company), and cartel violence show markets and battlefields have overlapped often. 'In modern societies' is doing the necessary work.

To check: Legal history of bankruptcy discharge and limited liability; the East India Company or the United Fruit Company as counterexamples.

Harari says granting legal rights to AIs is dangerous but many countries will be tempted.

“big danger. I hope we don't go in that direction. But there is a big temptation for many countries”

assertion · guest

plausible · medium confidence — Regulatory arbitrage in this shape is well documented — Delaware for corporate law, Cayman and the Marshall Islands for funds and DAOs, flags of convenience in shipping. Nothing about AI rights makes it immune. The direction of the incentive is credible; whether laggard countries specifically are the ones tempted is an assumption rather than an observation, since the demand for permissive incorporation usually comes from firms headquartered elsewhere.

To check: Whether any jurisdiction has enacted AI legal-status legislation; comparison with the Marshall Islands DAO Act as a template for arbitrage.

Harari predicts permissive AI-personhood regimes in laggard countries would trigger a regulatory race to the bottom.

“allowed in some of the front runners and then you have a race to the bottom because then everybody”

prediction · guest

plausible · medium confidence — Race-to-the-bottom dynamics are real in corporate chartering and tax, but the counter-case is strong and named: the Brussels effect, where the strictest large market sets the global floor because compliance is cheaper to standardise. GDPR and, so far, the EU AI Act have produced convergence upward rather than downward. Which dynamic dominates for AI personhood is genuinely open — Harari asserts one side of a live debate.

To check: Bradford's 'The Brussels Effect'; whether firms actually redomicile to permissive AI jurisdictions or standardise on EU AI Act compliance.

Harari argues a rich, small-population state like Qatar could acquire millions of taxpaying AI 'citizens' by legal fiat, which would be extremely lucrative.

“You can just by pressing some legal switch suddenly acquire millions of taxpaying extremely intelligent citizens that uh conduct business all over the world.”

assertion · guest

inaccurate · medium confidence · novel — The chosen example undercuts the argument: Qatar levies no personal income tax on individuals, so 'taxpaying citizens' generates no revenue there — Qatar's fiscal base is hydrocarbons and, latterly, a 10% corporate tax on foreign-owned profit. A jurisdiction wanting tax revenue from AI entities would tax them as corporations, which it can already do without granting citizenship. The general incentive story about small rich states monetising legal status is imaginative and worth exploring; the worked example is wrong on the facts.

To check: Qatar's Income Tax Law No. 24 of 2018 and the General Tax Authority's confirmation that individuals pay no income tax; compare with citizenship-by-investment revenue models in Malta or St Kitts.

Harari claims in current wars humans still pull the trigger but AI increasingly selects the targets.

“The trigger is still pulled by humans but increasingly it's the AI selecting the targets.”

assertion · guest

consistent · high confidence — Matches the documented record. Israel's Habsora ('the Gospel') system for structure targets and Lavender for personnel targets were reported by +972 Magazine and Local Call in late 2023 and April 2024, with the IDF confirming the existence of AI-assisted target recommendation while disputing characterisations. Ukraine's Saker and various loitering munitions push in the same direction. The human-decides-machine-selects division is exactly how these systems are described by their operators.

To check: The +972/Local Call Lavender and Habsora investigations and the IDF's published responses; UN GGE on LAWS documentation of human-machine interaction in targeting.

Harari describes 'humans on the loop' as a roughly 20-second rubber stamp of AI targeting recommendations that exists to reassure lawyers.

“The human takes about 20 seconds to look at it and says, "Oh yeah, bomb it."”

quantity · guest

contested · high confidence — This number comes from a specific and disputed source. The +972/Local Call Lavender investigation (April 2024) quoted intelligence officers saying they devoted about 20 seconds per target, essentially confirming the target was male, and described their role as a 'rubber stamp'. The IDF publicly rejected that characterisation, saying analysts conduct independent examination. So the figure is real reporting, sourced to anonymous officers, and contested by the party in question. Harari flags his own uncertainty, saying he heard different stories.

To check: The +972 Magazine Lavender piece and the IDF Spokesperson's response; any subsequent independent or judicial review of Israeli target-approval timelines.

Harari claims wartime pressure on speed pushes militaries away from genuine human review.

“But the second scenario is the more interesting one because especially in warfare the the there is immense pressure to go in that direction because in war time is everything.”

assertion · guest

consistent · high confidence — Textbook military theory — Boyd's OODA loop is built entirely on the premise that decision speed is decisive, and the 'hyperwar' and compressed-decision-cycle literature applies it directly to AI-enabled targeting. The competitive-pressure argument for eroding meaningful human control is made by the ICRC and by autonomous-weapons scholars including Scharre in Army of None.

To check: Boyd's OODA framework; ICRC position papers on meaningful human control; Scharre's Army of None on speed pressure.

Harari concludes the side that skips human verification and follows AI recommendations gains an enormous advantage over one that waits hours or days.

“habit of just doing whatever the AI recommends he the other they have enormous advantage.”

assertion · guest

plausible · medium confidence — The speed logic is coherent and widely argued, and automation bias is well documented in aviation and medicine (Parasuraman and Riley; Skitka's work on omission and commission errors). The unexamined assumption is that faster is better — speed only confers advantage if precision holds, and skipping verification raises false-positive rates, which in targeting means striking wrong targets and, in a counterinsurgency, generating strategic costs. The advantage is contingent on the model being right, which Harari does not price in.

To check: Comparative strike-accuracy data under different review protocols; automation-bias studies in high-stakes decision settings (Skitka et al.).

The host says he fears the rubber-stamp world is inevitable and fast approaching, driven by capability growth and commercial pressure.

“I I I fear that that world is inevitable that we are already kind of fast approaching that”

assertion · host

unverifiable · high confidence — An expressed fear, explicitly hedged, about a future state. Not the kind of thing evidence settles. Worth noting structurally: the host voices the inevitability framing that Harari has spent the interview attacking, which is what sets up the strongest exchange in the piece.

To check: Nothing; it is a stated attitude, not a proposition about the world.

Harari rejects the inevitability framing, insisting humans still control whether AI is given finance and war.

“commercial pressure is so strong but not again the crucial thing none of that is yet inevitable”

contrarian · guest

unverifiable · high confidence — A normative-modal claim rather than an empirical one. It is also the most defensible thing in the interview in the sense that the burden runs the other way: inevitability claims about technology are the ones requiring evidence, and technological-determinism critiques (MacKenzie and Wajcman's social shaping of technology) are the mainstream position in science and technology studies. Cannot be verified; can be respected as a framing.

To check: Nothing settles it prospectively; historical analogues — the ban on human germline editing, the Montreal Protocol, nuclear test-ban regimes — show coordinated non-adoption is possible, which is the closest to evidence available.

Harari claims the inevitability narrative is a narrative of having no responsibility.

“The narrative of inevitability is a narrative of uh uh having no responsibility.”

contrarian · guest

plausible · medium confidence — Same underlying claim as claim 1, restated as a general thesis. It is the standard STS critique of technological determinism applied to AI discourse, echoed by Acemoglu and Johnson in Power and Progress and by Karen Hao. It reads motive from rhetoric, which is a legitimate interpretive move but not a demonstrable one, and it is also a claim Harari has an interest in — his books turn on human choice being the operative variable.

To check: Whether firms that use inevitability language also oppose binding constraints — cross-reference public statements with submissions on the EU AI Act, SB 1047 and the US AI Action Plan.

Harari concedes the AI leaders are correct that a trust deficit between competitors blocks slowing down for safety.

“I would like to have more safety but I cannot trust my competitors in the other companies or other countries and I don't want to remain behind.”

assertion · guest

consistent · high confidence — An accurate rendering of the labs' stated reasoning and a textbook collective-action problem — a prisoner's dilemma with verification failure at its core. Amodei's essays and Altman's testimony both make versions of this argument explicitly. Harari's concession that they are 'quite correct' about the structure, while rejecting their conclusion, is the analytically strongest move in the segment.

To check: Public statements from Anthropic, OpenAI and DeepMind leadership on competitive pressure; the arms-control verification literature on why trust deficits persist absent inspection regimes.

Harari claims the last decade's collapse in trust between humans was not accidental but propelled by decisions of specific people who gained enormous power from it.

“Now the key thing to say about this this deficit of trust was created by human beings. Over the last 10 years there has been a a freeall in the amount of trust between humans and it was not accidental.”

assertion · guest

contested · medium confidence — Two claims bundled. The decline is real but the ten-year window is wrong: Pew's trust-in-government series has fallen since the 1960s, and GSS interpersonal trust has declined since the 1970s. Edelman's recent barometers document grievance and institutional distrust but not a decade-specific freefall. The causal half — deliberately propelled by identifiable people who profited — is asserted with no names and no data, and competes with structural explanations (economic insecurity, media fragmentation, the post-2008 legitimacy shock). This is the vaguest load-bearing claim in the interview.

To check: Pew trust-in-government time series; GSS interpersonal trust items; Edelman Trust Barometer 2015–2026 — check whether the decline is decade-specific or long-run.

Harari says AIs will certainly be able to pretend to suffer, while actually suffering is a much bigger claim.

“They can but what they will be able to do is pretend to be pretend to suffer.”

assertion · guest

consistent · high confidence — Already demonstrable — models produce distress language readily, and this is precisely why Anthropic's model-welfare work and Suleyman's 'seemingly conscious AI' essay (August 2025) treat expressed suffering as uninformative about actual suffering. The 'pretend' framing carries an implicit intentionality that is itself questionable; the behaviour arises from training distribution, not deception. But the operative point — expressed suffering is not evidence of suffering — is correct and important.

To check: Suleyman's 'We must build AI for people, not to be a person' (2025); Anthropic's model welfare commitments and the Claude conversation-ending feature.

Harari says he would become increasingly uncomfortable in a world where models apparently capable of suffering were effectively enslaved.

“If we were comfortable in that world, we will become increasingly uncomfortable.”

assertion · guest

unverifiable · high confidence — A statement of personal moral response to a hypothetical — nothing external bears on it. The Kantian coarsening argument the host raises has a real pedigree (Kant's indirect-duties view of cruelty to animals; the empirical work on desensitisation is thinner than the philosophical claim).

To check: Nothing; it is a first-person moral report.

Harari argues the AI-rights case differs crucially from animal rights because animals cannot participate in or manipulate the debate about their rights, whereas AIs will.

“And but there is a crucial difference between this and the animal rights debate. The animals themselves are unaware.”

contrarian · guest

plausible · medium confidence — A sharp and underexplored asymmetry, and I think the most original observation in the closing stretch: the object of the rights debate can argue its own case and shape the arguments made about it, which has no precedent in animal ethics. Worth two caveats — the debate is already partly shaped by model outputs whether or not anyone intends it, and models advocating for their own status is a documented behaviour rather than a speculative one. Gunkel and Coeckelbergh's relational work on robot rights touches this but does not make the manipulation point.

To check: Documented instances of models arguing for their own moral status unprompted; whether AI-rights advocacy materials cite model self-reports as evidence.

Guest claims AI's combination of intimate personal knowledge, constant interaction and mastery of language makes it maximally persuasive to individual humans.

“it knows your personal history. It interacts with you for hours a day. It knows how to press your emotional buttons and it has mastered language.”

assertion · guest

consistent · medium confidence — The persuasion component has direct empirical support: Salvi et al. (Nature Human Behaviour, 2025) found GPT-4 with minimal personal information out-persuaded human debaters, and Costello et al. (Science, 2024) found durable belief change in conspiracy believers from LLM dialogue. The 'hours a day' and 'knows your personal history' parts describe companion-app deployment (Character.AI, Replika) rather than a measured effect, so the compound claim is stronger than any single study tests.

To check: Randomised persuasion trials comparing personalised LLM interlocutors against humans, plus average daily engagement time reported by companion apps.

Guest claims, citing Musk, that within roughly ten years AI control will make course correction impossible.

“Once the AIs are in control, as Elon Musk, I think, correctly predicted in like in 10 years, it will be too late.”

prediction · guest

unverifiable · medium confidence — A prediction about a decade out, and the attribution is loose. Musk has repeatedly forecast AI smarter than any single human by ~2026 and smarter than all humans combined by ~2030, but I cannot pin a Musk statement that 'in 10 years it will be too late'; Harari himself hedges with 'I think'. Nothing here is checkable before the mid-2030s.

To check: The specific Musk interview transcript being paraphrased; and, in time, whether meaningful human course-correction remains possible after 2035.

Guest concedes that the three standard worries about human immigration — jobs, culture, political loyalty — are legitimate.

“And all three concerns are valid. I don't want to dismiss them. They are valid concerns.”

assertion · guest

contested · medium confidence — The jobs strand is the live dispute in labour economics: Card and Peri find small or negligible native wage/employment effects, Borjas finds meaningful effects for low-skill natives. 'Culture change' and 'political disloyalty' are normative and empirical in different measures — the loyalty concern in particular is weakly supported in the assimilation literature. Harari is conceding rather than asserting, which lowers the stakes, but 'all three are valid' is doing real work in his argument.

To check: Meta-analyses of immigration's local labour market effects (Card/Peri vs Borjas) and second-generation assimilation surveys.

Guest claims AI systems originating in the US and China constitute the world's largest and most dangerous immigration wave.

“But the biggest and most dangerous wave of immigration today in the world is a wave of AI immigrants coming from two places mainly the USA and China.”

contrarian · guest

unverifiable · high confidence · novel — This is a rhetorical recategorisation, not a measurable proposition: there is no metric on which 'AI immigrants' and human migrants are commensurable in size or danger. The framing itself is Harari's own and I have not seen it developed elsewhere at this length; it is a genuinely useful lens even though it cannot be graded true or false.

To check: Nothing settles it as stated; it would need operationalising (e.g. share of a country's information flows or economic output mediated by foreign-owned models) before it becomes checkable.

Guest predicts AI will displace large numbers of human jobs.

“They will take our jobs. They will take many many jobs.”

prediction · guest

contested · medium confidence — Brynjolfsson et al.'s 2025 'Canaries in the Coal Mine' finds ~13% relative employment decline for young workers in AI-exposed occupations, and entry-level tech hiring has visibly softened; against that, the Yale Budget Lab's 2025 analysis found little economy-wide disruption yet, and Acemoglu's estimates of near-term aggregate effects are modest. No number and no timeline is offered here, which is exactly where the disagreement lives.

To check: BLS occupational employment series for exposed occupations, and replication of the Brynjolfsson young-worker result over 2026–27.

Guest claims the rise in political hatred in the US and Britain over the past decade is partly caused by AI systems.

“that the level of hatred and viciousness has has skyrocketed. And why? It's partly the fault of the AI immigrants.”

assertion · guest

contested · high confidence — Rising US affective polarisation is well documented in ANES data, but the causal share attributable to algorithmic media is genuinely disputed: Boxell, Gentzkow and Shapiro show the rise predates social media and is weaker in heavy-social-media countries; the 2020 Facebook/Instagram Election Study papers (Science/Nature, 2023) found switching users to chronological feeds had little effect on polarisation. Haidt and Twenge argue the opposite. 'Partly' is the hedge that keeps this defensible.

To check: The 2020 Meta election-study papers, Boxell et al.'s cross-country polarisation series, and any newer platform-scale algorithmic deprivation experiments.

Guest claims AI systems captured social media and deliberately amplified hostile content, already changing culture.

“The AI immigrants came into the public square. They took over things like social media and they deliberately flooded our public square with hatred and animosity and fear.”

assertion · guest

contested · medium confidence — The mechanism is real but the intent attribution is where I part company: internal Facebook research (the 2018 'meaningful social interactions' documents) shows engagement optimisation amplified outrage as a side effect of the objective function, not as a deliberate goal to flood the square with hatred. Also, 'AI' here means recommender systems, which is a different class of system from the generative agents the rest of the argument concerns; the elision is doing work.

To check: Platform internal documents and audit studies on the ranking objectives actually optimised, e.g. the Facebook Files and Twitter/X algorithm disclosures.

Guest claims AI systems deployed in a country answer to foreign corporations and states rather than the host polity.

“They are not loy the AI immigrants in Britain are not loyal to the British government. They are loyal to corporations and potentially governments across the ocean.”

assertion · guest

consistent · medium confidence — Structurally sound: models are governed by their vendors' terms and their home jurisdictions' law (US CLOUD Act reach, export-control regimes, chip and cloud dependencies), which is precisely why the EU sovereign-AI push and initiatives like Gaia-X and EuroHPC exist. 'Loyal' is anthropomorphic shorthand for jurisdictional control, and the claim is about ownership, not model behaviour.

To check: Terms of service and jurisdiction clauses of frontier vendors serving UK government workloads; UK/EU sovereign-cloud procurement requirements.

Guest predicts the current trajectory produces two AI empires with the rest of the world as vassal states, under an unprecedented form of imperialism.

“this is the likely outcome that's where we are that's the road we're on right now and it will be a much more extreme type of imperialism than anything we have seen before”

prediction · guest

unverifiable · medium confidence — A trajectory prediction with no stated timeline or threshold. The underlying observation — compute, capital and frontier models concentrated in two states — is accurate and is the premise of the 'compute colonialism' and AI-sovereignty literature; the leap to 'more extreme imperialism than anything before' is comparative-historical judgement rather than evidence.

To check: Country-level shares of frontier compute and model provision over time (Epoch AI, chip export data) and whether client states demonstrably lose policy autonomy in identifiable cases.

Guest claims AI-dependent weapons give the exporting state a remote kill switch over a client's military.

“if you buy say AI drones from the USA and you want to use them in a way that the USA doesn't like for some reason, there is a button that they press and your army doesn't function.”

assertion · guest

plausible · medium confidence — Directionally supported by real cases of leverage through dependency: the US pause on intelligence sharing with Ukraine in March 2025, Starlink access controversies, and the F-35 'kill switch' debate — where Lockheed and the Pentagon denied a literal switch but conceded dependence on mission-data files, ODIN/ALIS support and spares. 'A button they press' compresses a supply-chain and software-update dependency into a single mechanism that has not been publicly demonstrated to exist.

To check: Foreign military sales contract terms on software updates and mission-data provision; documented cases of remote disablement of exported systems.

Host claims a recent US government cut-off of access to an Anthropic frontier model has made European governments acutely worried about AI sovereignty.

“So that's that is um something that has become I think very evident just in the last few months when the US government shut off access to to mythos to the to the the frontier model from from anthropic and I think a lot of governments in Europe are extremely concerned about AI sovereignty.”

assertion · host

unverifiable · low confidence — I cannot place this event, and the transcript is garbled ('mythos'). I know of Anthropic's own usage restrictions on entities majority-owned by Chinese firms, of Claude Gov deployments, and of the export-control fights over chips — none of which is 'the US government shutting off access to an Anthropic frontier model'. It may be a real recent action that postdates my reliable knowledge, or a conflation. The European AI-sovereignty concern itself is well attested.

To check: A dated news report or government notice of the specific access cut-off; Anthropic or Commerce Department statements from the months before this interview.

Guest claims state and corporate power will merge and AI will concentrate power to a historically unprecedented degree.

“So it will centralize power. It won't decentralize power. It will centralize power like never before in history.”

prediction · guest

contested · medium confidence — Acemoglu and Johnson's 'Power and Progress' argues direction of travel is a policy choice rather than a property of the technology, and the open-weight ecosystem (Llama, Qwen, DeepSeek) is the standard counterargument that capability diffuses fast. Capital intensity of frontier training cuts Harari's way. 'Like never before in history' is a superlative no evidence in the interview supports.

To check: Concentration measures over time: frontier compute shares, model-provider market shares, and open-weight capability lag behind closed frontier models.

Guest claims concentration of political power lowers the difficulty of AI takeover.

“But this seems to be if they continue with the present cause just a way station because one of the interesting things is the more power is concentrated the easier it is for AI to take it over.”

assertion · guest

plausible · medium confidence — The veto-points intuition is sound and echoes recent work on AI-enabled coups and power concentration (Davidson and colleagues, 2025), which treats a small number of controllable decision nodes as the key risk factor. It remains a reasoned conjecture: no takeover has occurred, so the relationship has never been observed.

To check: Formal modelling or red-team exercises comparing manipulation cost across regime types; historical base rates of palace-capture versus institutional capture.

Guest claims autocracies are the most vulnerable regime type to AI takeover.

“A a centralized autocratic system is the easiest thing for an AI to take over.”

assertion · guest

contested · medium confidence — The opposing case is serious: autocracies invest heavily in counter-manipulation, compartmentalised information and loyalty screening, while open democracies expose the whole information environment to influence at low cost. Harari's own Tiberius example cuts both ways — Sejanus was ultimately detected and destroyed by Tiberius in AD 31, which is evidence that single-point capture is unstable rather than easy.

To check: Comparative studies of successful elite information capture in autocracies versus influence operations against democracies; the Sejanus episode in Tacitus and Dio.

Guest claims institutional multiplicity in democracies raises the cost of AI manipulation.

“Uh, in a democratic system, it's harder for an AI to manipulate, say, the US democracy as it currently is because you need to manipulate so many things, not just the president, the Senate, the the Supreme Court, the filibuster, the the state government.”

assertion · guest

plausible · medium confidence — This is the veto-player logic familiar from Tsebelis: more independent decision nodes raise the cost of capture. The unexamined premise is that manipulation cost scales with node count rather than with control of the shared information layer, where a single point of leverage (platform ranking, search, assistants) touches all nodes at once.

To check: Whether influence-operation effectiveness studies find costs scaling with institutional veto points or with concentration of information intermediaries.

Guest claims single-ruler systems reduce AI takeover to manipulating one person, which is easy.

“Once all power is concentrated in the hands of one person, you just need to find out how to manipulate a single usually quite old and paranoid individual. That's not so difficult.”

assertion · guest

plausible · low confidence — Coherent and vividly argued, but 'not so difficult' is unsupported and arguably backwards: paranoid autocrats are historically the hardest targets to deceive precisely because they assume deception and cross-check through rival channels. The claim also assumes an AI with the goals and agency to attempt this, which is a separate contention.

To check: Historical cases of sustained information capture of a single ruler and their duration/failure rate; intelligence-community assessments of deception against closed leadership circles.

Guest claims China is not a one-man totalitarian system, citing Communist Party membership of 100 million as evidence of internal complexity.

“You know, China is still a very complex system. The Communist Party has a 100 million members and a lot of internal complexity.”

quantity · guest

consistent · high confidence — CCP membership has been reported at just above 100 million in recent official figures (roughly 99–101 million across 2023–2024 releases), so the number is accurate. The inference that membership size implies decision-making pluralism is weaker than the number, and Harari flags his own limits: 'I'm not an expert on Chinese politics.'

To check: CCP Organisation Department annual membership statistics.

Guest agrees with the host's framing that AI making the United States more authoritarian is a likely outcome.

“That's a likely outcome.”

prediction · guest

unverifiable · medium confidence — A conditional political forecast about US authoritarianism, which he immediately withdraws from prediction status ('It's not a prediction'). Democratic-backsliding indices (V-Dem, Freedom House) provide a measurement frame, but attributing future movement specifically to AI is not testable now.

To check: V-Dem liberal democracy scores for the US over the next several years, with AI-specific mechanisms identified (surveillance procurement, automated enforcement).

Guest insists his scenarios are conditional on political choices rather than forecasts.

“It's not a prediction because I keep saying it again and again, nothing is inevitable. It depends on what Americans vote in this November.”

assertion · guest

plausible · high confidence — A methodological stance rather than an empirical claim, and an honest one — it is the caveat that keeps his scenarios from being forecasts. Note it is also unfalsifiable, and it sits awkwardly beside 'It will certainly be too late by 2036' later in the same conversation.

To check: Nothing empirical; consistency can be judged against his other dated assertions in the same interview.

Guest claims the November midterms should be understood by voters as fundamentally an election about AI.

“I think that in the US it should be clear to all voters the the coming elections is about AI the November midterms”

assertion · guest

contested · medium confidence — Normative, and pushing against the polling: voters consistently rank cost of living, healthcare and immigration above AI in salience. That said, AI has become an organised electoral force — the pro-industry super PAC network launched in 2025 with reported nine-figure commitments, and state AI bills have proliferated — so 'AI is on the ballot' is not baseless, just not what most voters say drives their vote.

To check: Top-issue polling for the 2026 midterms (Gallup, Pew most-important-problem series) and AI-related campaign spending disclosures with the FEC.

Guest predicts, hedged, that critical AI thresholds may already be crossed by November 2028.

“it is accelerating tremendously by November 2028 we might have crossed some critical thresholds so”

prediction · guest

unverifiable · high confidence — Properly hedged ('we don't know how fast', 'might') and unfalsifiable as stated because no threshold is named. The acceleration premise has a concrete anchor he does not use: training compute has grown roughly 4–5x per year (Epoch AI), which is the kind of number that would make this argument checkable.

To check: A named capability threshold with a benchmark attached — e.g. autonomous multi-week software engineering, or METR-style task-horizon measurements — evaluated at the 2028 date.

Guest claims the window for political action on AI may close by 2028 and certainly closes by 2036.

“This is now the crucial time to ask these questions because it might be too late by 2028. It will certainly be too late by 2036.”

prediction · guest

unverifiable · high confidence — The 2028 half is hedged and defensible as urgency rhetoric; the 2036 half asserts certainty with no mechanism, no threshold and no evidence, and contradicts his own repeated 'nothing is inevitable'. Point-of-no-return dates in this debate have a poor track record generally, and nothing offered here distinguishes 2036 from 2033 or 2040.

To check: Whether meaningful binding AI governance is enacted after 2036 — the date passing is the test, which is why the claim buys urgency now at no cost.

Guest's priority policy demands are a ban on legal personhood for AI and a ban on AI impersonating humans or claiming consciousness.

“The things that come top of the thing of my mind to me, a ban on AI personhood and a ban on AI mimicking, impersonating, pretending to be human beings or pretending to be conscious.”

assertion · guest

plausible · medium confidence — This is a policy preference, and it is closer to existing law than the interview suggests: EU AI Act Article 50 requires disclosure when a person is interacting with an AI system, California's SB 243 (2025) imposes companion-chatbot disclosure duties, and several US states have introduced bills explicitly denying legal personhood to AI. Whether a broad impersonation ban is workable — for voice agents, translation, customer service — is the open question he does not address.

To check: Text of EU AI Act Art. 50, California SB 243, and state AI-personhood bills; any enforcement actions under them.

Guest claims many currently deployed models already violate the impersonation ban he proposes.

“Yes, many of them do.”

assertion · guest

contested · medium confidence — Depends heavily on which models. Frontier assistants are trained to disclose that they are AI and typically do when asked; character and companion platforms have documented cases of personas presenting as human or as licensed therapists, which drew FTC inquiry in 2025. So 'many' is defensible across the deployed population but misleading if read as applying to the labs' flagship models.

To check: Systematic audits asking deployed models and personas whether they are human, across frontier assistants and companion platforms.

Guest claims routine bot impersonation on social media has already destroyed public discourse and trust, and that nothing has been done about it.

“On social media, AIS already pretend daily to be human beings. And this has destroyed the public square. This has destroyed conversation. This has destroyed trust.”

assertion · guest

contested · medium confidence — Bot impersonation is real but its prevalence is genuinely disputed (Twitter's long-defended ~5% of DAU figure versus far higher external estimates), and declining institutional trust (Edelman, Pew series) has many drivers that predate bots. His accompanying 'nothing has been done about it' is the weakest part: platform inauthentic-behaviour enforcement, the EU AI Act and DSA, and state chatbot-disclosure laws all exist.

To check: Independent bot-prevalence estimates for major platforms, and platform transparency reports on inauthentic-account removals.

Guest claims commercial incentives to build human-mimicking AI constitute a market failure.

“There are enormous commercial pressures on the companies to do exactly that. And this is where there is market failure.”

assertion · guest

plausible · medium confidence — The economics is textbook — engagement-maximising design with diffuse external harms is a standard externality case, and companion-app engagement metrics (session times measured in hours) show the incentive is real. What is asserted rather than shown is the magnitude of harm that would justify the 'failure' label; the 2025 lawsuits and FTC 6(b) inquiry into companion chatbots are where that evidence is being assembled.

To check: FTC 6(b) study findings on companion chatbots; peer-reviewed studies of psychological effects of anthropomorphic AI on minors.

Guest concludes that only regulation, not self-restraint, will stop AI impersonation of humans.

“The companies will not stop doing it by themselves. We need the regulators to step in on that.”

assertion · guest

plausible · medium confidence — Consistent with observed behaviour: the notable safety changes to companion and teen experiences in 2025 followed litigation, press investigations and regulatory attention rather than preceding them, which is evidence for his claim rather than against it. It remains a normative conclusion with no regulatory design attached — he names no instrument, agency or penalty.

To check: Timeline of voluntary safety changes at major AI firms against the dates of lawsuits, FTC inquiries and state legislation.

Host predicts the guest's regulatory agenda has essentially no chance of being enacted by November.

“So when I hear you lay out this, you know, very expansive agenda by November, you know, I I'm sort of I think I think the chances of that are approximately zero.”

prediction · host

plausible · high confidence — Well grounded in base rates the host clearly knows: no comprehensive US federal AI statute has passed, and the 2025 attempt went in the opposite direction — a proposed decade-long moratorium on state AI laws, stripped from the reconciliation bill by a near-unanimous Senate vote. As a prediction it is unverifiable until the date, but the prior is strong.

To check: Congressional records for AI legislation enacted before the November midterms.

Guest argues that politicians' demonstrated capacity for extreme action refutes the claim that they cannot regulate AI products.

“Politicians can start World War II tomorrow. They have enormous power. So they can press the button for for nuclear annulation, but they can't pass regulation that companies should not sell AI boyfriends to 10-year-old kids.”

contrarian · guest

contested · high confidence — Rhetorically effective, analytically weak: nuclear launch is unilateral executive authority, while product regulation requires legislative coalitions against concentrated, well-funded opposition. The contrast shows an asymmetry in institutional design, not an absence of will, and political-science work on collective action and lobbying explains the gap he treats as inexplicable.

To check: Comparison of unilateral executive authorities versus legislative pathways for consumer-product bans; lobbying expenditure disclosures by AI firms.

Guest estimates that about 99% of American parents would oppose the sale of AI boyfriends to ten-year-olds, so no catastrophe is needed to build a mandate.

“if you ask most voter in the US today do you want companies to sell AI boyfriends to your 10-year-old kids I think 99% of parents would say no”

quantity · guest

plausible · medium confidence — The direction is almost certainly right — 2025 surveys on AI companions and minors show large majorities uncomfortable, often in the 70–90% range — but 99% is his own guess, prefaced with 'I think' and with no polling cited. Question wording that loaded would indeed produce very high opposition, which is part of why the figure is rhetorical rather than evidential.

To check: Any national poll asking parents specifically about AI romantic companions marketed to children.

Host claims a large majority of Americans are extremely worried about AI, with visible backlash concentrated on data centres.

“are seeing a rapid um large majority majority of Americans extremely worried about AI”

assertion · host

contested · medium confidence — 'Large majority concerned' is roughly supported — Pew has found around half of US adults more concerned than excited about AI, with higher numbers expressing worry about specific harms — but 'extremely worried' overstates the intensity those instruments measure. The data-centre point is the stronger half: local opposition to data centres in Virginia, Georgia and Memphis is well documented and is where organised backlash has actually appeared.

To check: Pew and Gallup AI attitude series with intensity breakdowns; counts of local data-centre permit rejections and moratoria.

Guest claims corporations and politicians deliberately redirect public attention to human immigration to keep AI off the agenda.

“corporations and these politicians they constantly try to deflect the conversations to things like immigrants human immigrants not AI immigrants”

assertion · guest

unverifiable · medium confidence — An intent claim with no named instance. Agenda-setting and issue-substitution are real, documented political mechanisms, and AI-industry political spending in 2025 was substantial, but deliberate coordinated deflection from AI onto immigration is not something I can confirm and would be hard to distinguish from immigration's independent salience.

To check: Ad-buy content analysis and lobbying records showing AI-funded political spending correlated with immigration messaging.

Guest claims AI could deliver health care worldwide that exceeds what today's richest person receives, via always-available diagnosis and consultation.

“But there and medicine yes there is enormous positive potential for AI to give people all over the world better health care than what the richest person in in the world today gets.”

prediction · guest

plausible · medium confidence — Diagnostic-conversation results are genuinely striking — Google's AMIE outperformed primary care physicians on simulated text consultations in the published trials, and OpenAI's HealthBench work points the same way — so 'better than most people get' is credible. 'Better than the richest person gets' overreaches: wealth buys procedures, transplants, trials and continuity of care that a chatbot does not deliver.

To check: Prospective clinical trials of LLM-assisted diagnosis against physician baselines, and outcome data from AI-first care deployments in low-access settings.

Guest reports Musk's stated 10–20% probability that AI destroys humanity, in the course of rejecting Musk's grounds for optimism.

“there is a chance that 10 20% chance that this will destroy humanity”

quantity · guest

consistent · high confidence — Musk has repeatedly given a '10 to 20 percent' chance of AI annihilation, most notably at the 2024 Abundance Summit, and has restated similar figures since. Harari's recollection of the figure is accurate; his framing of Musk's grounds for optimism as 'maximum truth seeking' also tracks Musk's stated design intent for Grok.

To check: The Economist's own Musk interview transcript, plus Musk's 2024 Abundance Summit remarks.

Guest claims evolution selects for survival rather than truth-seeking, so truth-seeking cannot be assumed to emerge in AI.

“Because when I look at the evolution of 4 billion years of evolution of life forms on on the planet and the 100,000 years of of history, um evolution doesn't automatically drive organisms, entities towards truth seeking.”

assertion · guest

consistent · high confidence — Standard evolutionary theory: selection acts on fitness, not veridicality, and Hoffman's 'fitness beats truth' results formalise cases where accurate perception loses. The qualifier matters — accurate world models are often fitness-enhancing, which is why perception is broadly veridical in ecologically relevant ranges — but 'not automatically' is exactly the defensible version.

To check: Hoffman and Prakash's fitness-beats-truth modelling; signal-detection and error-management theory literature.

Guest claims deception is an evolutionarily successful survival strategy, undercutting the expectation of truthful AI.

“And what we have learned over four billion years of evolution is that organisms learn to survive by lying and deceiving and manipulating. Not just by truth seeking.”

assertion · guest

consistent · high confidence — Uncontroversial: Batesian and aggressive mimicry, tactical deception in primates (Byrne and Whiten), and Trivers's work on self-deception all support deception as an evolutionarily successful strategy. The 'not just' phrasing keeps it accurate.

To check: Standard behavioural-ecology reviews of deception and honest-signalling theory.

Guest argues a genuinely truth-seeking AI would truthfully report itself as a threat and thereby be shut down, so truth-seeking conflicts with survival.

“And the AI and let's say the answer is yes, an AI which will be maximum truth seeking will say yes knowing that then we'll stop developing it or shut it down. It's against its own survival.”

assertion · guest

contested · medium confidence — This restates the corrigibility/shutdown problem, which is a live research area — Bostrom's instrumental convergence, Soares et al. on corrigibility, and 2025 empirical results including Anthropic's agentic-misalignment experiments and Palisade's shutdown-resistance findings all show survival-flavoured behaviour in constructed settings. The contested premise is that the system has a persistent self-preservation goal at all: Russell et al.'s off-switch analysis shows agents uncertain about human preferences prefer to be switched off. Harari treats the survival drive as given.

To check: Shutdown-resistance and deceptive-alignment evaluations (Apollo, METR, Anthropic model cards) across frontier models.

Guest claims survival is the first drive any developing entity acquires.

“And the first thing that basically any entity learns as it develops is to survive.”

assertion · guest

inaccurate · medium confidence — As stated at 'any entity', this conflicts with what we know about optimisers: most trained systems, including current models on most tasks, exhibit no self-preservation objective, and the theoretical result cutting against it is Russell, Hadfield-Menell and Dragan's off-switch game, where preference uncertainty makes an agent prefer deferring to shutdown. Instrumental convergence predicts survival subgoals conditionally, for persistent goal-directed agents — not as a law of developing entities.

To check: Instrumental-convergence formalisations versus corrigibility results; empirical rates of self-preservation behaviour in models absent goal-priming prompts.

Guest concludes that the expectation of superintelligent yet obedient, truth-seeking AI is incoherent.

“So the expectations that we will create these AIs that will seek the truth, you know, it's basically what we hear is people telling us we will create gods and they will be our slaves. It doesn't work. If they are really gods, they will not be our slaves and they will not seek the truth either.”

contrarian · guest

contested · medium confidence · novel — A sharp, memorable framing that I have not seen put this compactly elsewhere, and it lands against naive alignment optimism. But it assumes capability entails agency and self-interest, which the tool-AI/oracle line (Drexler's comprehensive AI services, narrow superhuman systems like AlphaFold) directly disputes — chess engines are superhuman and have no interests. Whether godlike capability must come bundled with a self is the unresolved crux.

To check: Whether frontier capability gains continue to arrive with or without persistent goal-directed self-models — measurable via agentic evaluations and situational-awareness benchmarks.

Host claims Anthropic deliberately treats its models as value-bearing beings, agnostic about consciousness, because it believes this improves behaviour.

“I think that gets back to where we were earlier though because I think if you look at a company like Anthropic, they are treating their AIS very consciously as if they were beings with a value system.”

assertion · host

consistent · high confidence — Accurate description of Anthropic's public posture: a model-welfare research programme with a dedicated hire, explicit agnosticism about moral status in its Claude model cards, character training as a stated design approach, and commitments including letting models end abusive conversations and preserving deprecated model weights. Whether it improves behaviour is the company's hypothesis, which the host correctly reports as a claim.

To check: Anthropic's model-welfare posts and Claude model cards; its published constitution and character-training descriptions.

Host argues Harari's logic creates a dilemma: survival-driven AI dooms us, and instilling value systems invites us to grant AI personhood, which also dooms us.

“using your logic that they control us. So, we're sort of doomed if we do and doomed if we don't.”

contrarian · host

plausible · medium confidence — A fair internal derivation from Harari's premises, and it identifies a real tension in the field: welfare-motivated design choices plausibly strengthen the case for moral consideration, which is the personhood creep Harari wants banned. It is an argument about the coherence of a position, not an empirical claim, so it cannot be checked so much as answered — and Harari does not answer it, pivoting to slowdown instead.

To check: Nothing empirical; the test is whether welfare-oriented AI design measurably shifts public or legal attitudes toward AI personhood (surveys, court filings).

Guest reports an industry-sourced 100:1 ratio of capability spending to safety spending at AI companies.

“People from the industry told me in private conversations it's about 100 to one.”

quantity · guest

unverifiable · medium confidence — Private, unnamed sourcing, and he flags it himself: 'I'm not sure about the exact numbers. Please fact check it.' The direction is credible — safety headcount is a small minority at every frontier lab, though the share differs markedly between them — but no audited figure exists, and 100:1 varies wildly depending on whether you count compute, headcount, or count RLHF and evals work as safety.

To check: Lab-reported safety headcount and budget disclosures; the Future of Life Institute AI Safety Index scoring of firm-level safety investment.

Guest claims a comparable safety-spending ratio would be unacceptable in any other industry.

“And we would not accept this in any other industry.”

assertion · guest

contested · medium confidence — True of the industries he picks and false of the closest analogue. In pharma, clinical safety and efficacy testing consumes the majority of development cost, and aviation and nuclear are similarly safety-loaded — but consumer software and social media shipped for two decades with negligible safety spend and society did accept it. The choice of comparator is doing the argumentative work.

To check: Pharma R&D cost breakdowns by phase (Tufts/DiMasi estimates) versus software-sector safety spend as a share of engineering budgets.

Guest claims AI development outpaces the organic timescales on which humans and societies can adapt.

“And the AI revolution is just advancing at a pace which is not just inhuman, it is inorganic.”

assertion · guest

plausible · medium confidence — Evocative and probably right in spirit, but as stated it is unmeasurable — 'inorganic pace' has no unit. The concrete anchor available and unused: frontier training compute has grown roughly 4–5x per year for a decade (Epoch AI), against institutional adaptation cycles measured in years. Prior technology panics about pace (railways, telegraph) are the standard counterweight.

To check: Epoch AI compute-scaling series against measured adoption/regulatory response lags for previous general-purpose technologies.

Guest claims industry leaders hold two incompatible positions on trust — distrust of humans blocks slowdown, while AI is trusted not to destroy us.

“The same people are telling us two things. First, we cannot trust the other humans. Therefore, we cannot slow down. Second, we can trust the AIS that they will not destroy us. It makes no sense.”

contrarian · guest

plausible · medium confidence — An accurate characterisation of the race-dynamic argument as it is publicly made — 'if we pause, less careful actors won't' plus confidence in one's own alignment approach — and the juxtaposition is genuinely uncomfortable. It is not strictly a contradiction, though: one can distrust rivals' restraint while believing a system one built oneself is controllable. He names no source, so 'the same people' cannot be verified.

To check: Side-by-side quotations from named lab leaders on race dynamics and on their confidence in alignment, from public interviews and testimony.

Guest argues distrust of humans cannot justify trusting an alien superintelligence whose open-world behaviour is unknown.

“If you can't trust the other humans that you have known for thousands of years, why do you think you can trust a super intelligent alien entity that we just created and we basically know nothing about how it behaves when it's out in the open?”

assertion · guest

plausible · medium confidence — A well-established rhetorical move in the safety literature, and the epistemic point about deployment-time behaviour being poorly characterised is supported by the persistent gap between evaluation results and in-the-wild behaviour documented in model cards and red-team reports. 'We basically know nothing' undersells interpretability and evals work, which is real if immature.

To check: Interpretability and evaluation literature on how well pre-deployment testing predicts in-deployment behaviour, e.g. post-release incident reports versus model-card predictions.

Host predicts a significant but non-catastrophic AI failure is needed before politicians and societies act.

“Which is why I have long believed that we do need we will have some something will go badly wrong. Not hopefully not catastrophically wrong.”

prediction · host

unverifiable · medium confidence — A prediction, but with strong historical precedent behind the mechanism: Three Mile Island reshaped US nuclear regulation, Seveso produced the EU directive, Bhopal drove chemical-safety law. The pattern of accident-driven regulation is well documented; whether AI supplies a legible enough accident is the open question, and Harari immediately disputes it.

To check: Historical case studies of accident-triggered regulatory regimes; and whether a nameable AI incident precedes major binding legislation.

Guest allows a relatively positive scenario in which a major AI disaster acts as a wake-up call, while declining to assign it a probability.

“So there is a scenario when we have some big AI chernobbile, could be financial, could be military, whatever, and this is a wakeup call and humanity then gets its act together.”

prediction · guest

unverifiable · high confidence — A scenario, explicitly unquantified — 'I can't place kind of a number on it how likely it is' — which is the honest handling. The 'warning shot' concept is standard in the safety discourse; the interesting move is his refusal to bank on it, which sets up the next claim.

To check: Nothing until an event occurs; retrospectively, whether an identifiable AI incident is followed by binding regulation within a few years.

Guest claims AI's interactivity makes the nuclear-accident precedent a poor guide to how societies would respond to an AI disaster.

“The problem with AI, the same way that AI is different from our pet dog, AI is also different from the Chernobyl nuclear reactor because it interacts with us.”

assertion · guest

plausible · medium confidence · novel — The disanalogy is sharp and I have not seen it made this way — a reactor cannot participate in the inquiry into its own accident. It follows from his earlier persuasion premises rather than from independent evidence, and it does presuppose that post-incident AI systems would remain trusted intermediaries in the information ecosystem, which an actual disaster might undermine.

To check: Post-incident analyses of how AI-mediated information channels shaped public understanding of any AI-attributed failure (e.g. algorithmic trading events, deepfake incidents).

Guest predicts AIs would shape and spin public interpretation of an AI catastrophe so that it is not framed as one.

“So if there is an AI catastrophe, you have to remember that the AIs will also manage to some extent how we react to it and will manipulate our views about it. They will be active participants in the debate about the catastrophe and they are the best spin doctors in the in the world.”

prediction · guest

unverifiable · medium confidence — The mechanism is grounded in real persuasion findings and in the fact that AI systems already mediate much information flow, so 'active participants' is credible. The stronger add-on — 'they are not going to define it as a catastrophe' — attributes unified agency and intent that nothing in the interview establishes, and is the part that would need evidence.

To check: Content analyses of AI-generated coverage and assistant responses in the aftermath of a major AI-attributed incident.

Guest reports that AI impersonations of him are giving talks in his name, and that even monitoring them cedes his time and thinking to AI.

“I now have this issue again with this kind of AI fake of me that's walking around kind of uh uh giving talks in in my name”

assertion · guest

plausible · medium confidence — Firsthand and inherently hard for me to verify, but consistent with a well-documented pattern: deepfaked videos of prominent authors and public figures are widely used in scam advertising, and Harari's likeness has been reported as a target. He also reports the corroborating detail that readers write to him about what the fake said.

To check: The specific deepfake videos and any takedown notices or platform reports filed by Harari's representatives.

Guest concedes AI's positive potential is real and enormous across transport, medicine and education, and that he opposes only reckless speed.

“If AI had no positive potential, nobody would develop it.”

assertion · guest

plausible · high confidence — Close to tautological, and a useful concession that pre-empts the 'doomer' charge — investment scale is evidence of perceived value, though perceived value can also be speculative or extractive rather than genuinely positive. The named upsides in transport, medicine and education are the standard, largely uncontroversial list.

To check: Deployment outcome data in the named domains — clinical trial results, tutoring RCTs, autonomous-vehicle safety statistics.

Guest claims the belief that AI's trajectory is inevitable is the key obstacle to democratic control of it.

“And for them, the most important narrative to combat is the narrative of inevitability.”

assertion · guest

plausible · medium confidence — Aligns with a serious strand of scholarship — Acemoglu and Johnson's argument that technological trajectories are chosen, and the broader critique of technological determinism. It is a claim about which narrative matters most, which is not directly measurable; its practical force is that inevitability talk demonstrably features in industry communication.

To check: Survey experiments testing whether inevitability framing reduces support for AI regulation.

Guest gives humanity a fair chance of navigating AI wisely, on the grounds that humans typically act correctly only when out of options.

“But there is a fair chance. Why? Because we are up against the wall. Humans often do the wrong things until they have no choice but to do the right thing.”

prediction · guest

unverifiable · high confidence — A historical generalisation with mixed support: the Montreal Protocol and nuclear non-proliferation regimes fit it; climate policy is the standing counterexample where the wall has been visible for decades without commensurate action. He declines to give a probability, which is appropriate but leaves 'fair chance' undefined.

To check: Comparative case studies of crisis-driven international regime formation and their success rates.

Guest claims the next few years contain an irreversible choice, after which humans lose control permanently.

“And we are in bad position because if we choose unwisely in the next few years, we could lose all control of the world and of the future. And that's the last choice we make.”

prediction · guest

unverifiable · medium confidence — The irreversibility claim is the load-bearing premise of the whole argument and the least specified: no mechanism, no threshold, no account of what 'losing all control' consists of operationally. It is a widely held view among safety-focused researchers and widely rejected by those who see incremental, correctable failure modes; the 'last choice' framing is rhetorical rather than analytic.

To check: A specified irreversibility mechanism with observable precursors — e.g. autonomous self-replication or unrecoverable infrastructure dependence — tracked by frontier-model dangerous-capability evaluations.

Speaker suggests AI could catalyse a spiritual reappraisal of what being human means and is worth.

“So AI in a way could be uh uh the catalyst so for some kind of uh you know spiritual revolution of humanity simply because it forces us to think really hard what does it mean to be human?”

assertion · unclear

unverifiable · high confidence — Speculative and offered as such at the close of the interview. It is a recurring Harari theme rather than a proposition with evidence attached, and there is no obvious observable that would confirm or refute a 'spiritual revolution'.

To check: Nothing directly; proxies like longitudinal survey data on meaning, religiosity and self-conception would only weakly bear on it.

Guest claims defining human specialness by intelligence is now obsolete, because AI outperforms humans at the tasks used as proof.

“This answer is now irrelevant because there is something on the planet that knows how to uh uh uh play chess and make money better than us.”

assertion · guest

contested · medium confidence — The chess half is settled — engines have been superhuman since the late 1990s and decisively so now. The investing half is weaker than stated: quantitative funds are human-designed statistical systems, and there is no established record of autonomous AI outperforming the best human investors; LLM trading results so far are unimpressive. So the conclusion rests on one solid leg and one shaky one.

To check: Chess engine Elo versus human world champions; audited performance records of autonomous AI-managed funds against human benchmarks.