Hinton vs LeCun: How to Think When Experts Completely Disagree

By techandmindsetlab
May 2026
~11 min read

Geoffrey Hinton, who shared the Nobel Prize in Physics in 2024 for foundational work on neural networks, estimates a 10-20% probability that AI destroys humanity. He left Google specifically to be able to speak freely about AI existential risk. He has described a scenario where “rich people will use AI to replace workers creating massive unemployment,” and his probability estimates for catastrophic outcomes are not rhetorical — he has stated them precisely, repeatedly, and with explicit acknowledgment of uncertainty.

Yann LeCun, who holds the Turing Award (the field’s highest honor) for work that is technically complementary to Hinton’s, calls existential AI risk “complete B.S.” and “preposterous.” He works at Meta, has built some of the most capable AI systems in existence, and does not think AI in its current or foreseeable trajectory poses anything like the risks Hinton describes.

Gary Marcus — a cognitive scientist and long-time AI critic who predicted limitations that current AI systems have indeed exhibited — has argued that hallucinations in language models are architectural defects, not fixable bugs, stating: “the only way to kill hallucinations is to not run the system.” His critique is different from both Hinton and LeCun: not existential risk, but fundamental technical unreliability in current architectures.

These are not fringe voices or public intellectuals with impressive titles but narrow expertise. These are the people who built the technology. And they disagree, completely, about its most fundamental implications.

The question worth examining is not who is right. It is: what does this level of disagreement tell us, and how should we reason in its presence?

The First Instinct to Resist: Authority Averaging

The most common response to expert disagreement is a form of epistemic averaging: find a moderate position between the extremes, assign credibility to each expert roughly proportional to their credentials, and land somewhere in the middle. This feels rigorous because it appears to take multiple views seriously. It is not rigorous. It is a heuristic that systematically produces miscalibrated conclusions.

Consider what authority averaging does here: it takes Hinton’s 10-20% and LeCun’s roughly 0% and produces some intermediate estimate, weighted by how much you like each person’s credentials. But averaging over expert disagreement assumes that the disagreement reflects different assessments of shared evidence — that both experts have the same information and are simply weighting it differently. If that’s true, averaging is somewhat defensible. If the disagreement reflects different priors, different conceptual frameworks, or different underlying questions, averaging mixes incompatible quantities and produces numerical soup.

The more important skill is not to average but to understand the structure of the disagreement: what are they actually disagreeing about, why might each of them be systematically biased in their respective directions, and what would constitute evidence that would move each of them toward the other’s position?

Disaggregating the Disagreement

When you examine what Hinton and LeCun are actually arguing, the disagreement is not a single claim about a single fact. It consists of several distinct empirical and conceptual sub-disagreements, each of which can be assessed partially independently.

Disagreement 1: Capability Trajectory

Hinton believes current AI architectures — neural networks, transformers — are on a trajectory toward general intelligence that is faster than most people appreciate, and that this trajectory creates risks. LeCun believes current architectures are fundamentally limited in ways that preclude the kind of autonomous, self-directed reasoning required for existential risk scenarios. He has been developing alternative architectural proposals (energy-based models, world model architectures) partly because he thinks the current dominant approach has specific, fundamental limitations.

This is an empirical disagreement about architecture and capability scaling — the kind of question that evidence can partially address. The track record is mixed: scaling has produced capabilities that both researchers have found surprising, but it has also run into specific limitations (compositional reasoning, robust world modeling, reliable grounding) that LeCun’s critique predicted. Neither side has a clean predictive win here.

Disagreement 2: The Risk Pathway

Hinton’s risk scenario involves something like: sufficiently capable AI systems pursuing goals that aren’t perfectly aligned with human welfare, at a scale and speed that outpaces human intervention. LeCun’s objection is roughly: current AI systems don’t have goals in any meaningful sense; they generate next tokens based on statistical patterns; attributing goal-directedness and strategic behavior to these systems is a category error driven by anthropomorphism rather than technical analysis.

Gary Marcus’s critique cuts differently: the problem is not alignment of goal-directed systems but unreliability of statistically-driven systems. Hallucinations aren’t a bug to fix — they’re what you get when you build systems that model surface-level text patterns without grounded world models. His concern is less catastrophic risk and more systematic misinformation at scale from systems that are confidently wrong.

These are genuinely different risk models pointing at genuinely different failure modes. You cannot resolve the disagreement by picking one; you can hold all three as partially correct about different aspects of a complex technology.

Disagreement 3: The Incentive Structure Question

An honest analysis of expert disagreement must include incentive analysis — not as a way to dismiss views, but as a way to understand systematic bias. Both Hinton and LeCun have conflicts of interest in opposite directions.

Hinton’s position (resigned from Google, speaking freely about risks) generates visibility, speaking invitations, and a certain kind of influence that comes from being the credentialed scientist who rings the alarm. His institutional incentives, post-Google, are partially aligned with being taken seriously as a risk voice. This does not make him wrong, but it is a factor to note.

LeCun works at Meta, which is deploying AI at massive scale and would face significant reputational and regulatory consequences if existential risk concerns were widely accepted as credible. His professional context creates pressure — not necessarily conscious — toward views that support continued aggressive AI development. This also does not make him wrong.

The existence of opposing incentive structures doesn’t cancel them out or leave us where we started. It tells us to look carefully at the claims each expert makes that happen to align with their institutional interests, and to weight those claims with appropriate additional skepticism.

Base Rates and Reference Classes

One framework that cuts through expert disagreement more cleanly than credential-weighting is base rate analysis: what does the historical record of comparable situations tell us about the probability of the outcome in question?

For Hinton’s risk scenario, the relevant reference class is “powerful new technologies that could cause civilizational harm.” The historical record here is genuinely mixed: nuclear weapons did not cause civilizational collapse despite many predictions that they would, though the near-misses (Cuban Missile Crisis, multiple false alarm incidents) suggest the outcome was not assured. Engineered pandemics have not yet caused civilizational harm. Neither has industrial pollution, despite causing enormous damage. The track record of humanity surviving civilizationally threatening technologies is decent, though the sample size is small and the technological systems in question were not self-improving.

For LeCun’s position, the relevant reference class is “new technologies that critics predicted would be civilizationally harmful but were not.” This is a longer list: nuclear power, GMOs, the internet, social media (arguable), and earlier AI waves. Technologies that were predicted to cause catastrophic harm and did not outnumber those that did, by a large margin.

Base rates don’t resolve the disagreement — the “this time might genuinely be different” argument is available to Hinton, and it’s not trivially dismissible. But they should anchor your priors before the expert disagreement comes into play, not after.

Track Records and Calibration

Another underused framework: rather than asking what expert X believes, ask how well-calibrated expert X’s past predictions have been.

Gary Marcus has made specific, falsifiable predictions about the limitations of deep learning architectures — about systematic failures in compositional reasoning, robustness to distribution shift, and reliable grounding. A number of these predictions have proven partially correct (hallucinations and compositional reasoning failures remain serious unsolved problems) and some have proven too pessimistic (the capabilities of scaled models exceeded his predictions in several domains). His calibration is imperfect but trackable.

Hinton’s track record on AI capabilities is extraordinary — he was right about neural networks before most of his field was. His track record on AI safety and societal impact is shorter and harder to evaluate; his risk concerns are relatively recent, and the predictions haven’t had time to resolve. LeCun’s architectural predictions about limitations of current approaches have a mixed record similar to Marcus’s. He was right that current systems lack certain types of reasoning; wrong about how far they could scale on the dimensions they’re good at.

Track record analysis does not give you a winner. It gives you calibration information: which claims from which experts have historically proven accurate, and in which domains.

What Expert Disagreement Itself Signals

There is a meta-level observation worth making: the persistence of fundamental disagreement among people who have spent their careers on a problem is itself epistemically informative. It suggests that the problem does not have the kind of clear, checkable answers that would resolve disagreement among careful reasoners with access to the same evidence.

This is different from, say, disagreement about climate change (where the scientific consensus is clear and the “disagreement” among experts is often manufactured or represents a small minority dissenting from a large consensus). In AI risk, the disagreement is among core practitioners, is not resolvable by pointing to a clear evidentiary record, and involves fundamental uncertainties about the behavior of systems that don’t yet exist in the relevant form.

That is a signal to hold your own views loosely, update them as evidence accumulates, and be skeptical of anyone — including the experts — who presents their position as settled. The appropriate epistemic state here is calibrated uncertainty with explicit acknowledgment of what would change your mind.

Forming Your Own View: A Framework Summary

Putting these tools together:

1. Disaggregate the disagreement. What specific empirical and conceptual sub-claims are in dispute? Can you assess each independently?
2. Apply incentive analysis. What are each expert’s institutional and reputational incentives? Which of their claims align suspiciously well with those incentives?
3. Check base rates. Before weighting expert opinion, anchor on what the historical record of comparable situations suggests.
4. Evaluate track records. How well-calibrated have each expert’s past predictions been in the relevant domains?
5. Notice what would change your mind. For each expert’s position, ask: what evidence would cause them to update significantly? If the answer is “nothing,” treat that as a flag about epistemic quality, not about the correctness of the view.
6. Maintain calibrated uncertainty. If you’ve done the above and remain genuinely uncertain, that uncertainty is accurate and should be reported as such — rather than collapsed into false confidence in one direction.

Questions Worth Sitting With

• Is fundamental expert disagreement itself a reliable signal about the epistemic quality of a field — suggesting that our tools for generating and evaluating evidence in that field are inadequate? Or does it sometimes simply reflect that the questions are genuinely hard, not that the field is epistemically immature?
• Hinton and LeCun were intellectual collaborators for decades and built much of the same technical foundation. Their divergence on risk suggests that technical expertise in building AI does not determine one’s conclusions about AI safety. What factors actually do determine these conclusions, if not technical knowledge?
• If the relevant predictions (about AI capabilities, AI risk scenarios, AI societal impact) have resolution timescales of 10-30 years, can any of the current expert disagreements be resolved before the decisions that depend on them need to be made?

Apply the Framework to One AI Claim

Take one AI claim you’ve encountered recently — in the news, from a colleague, in a vendor pitch, from a researcher — and run it through the framework above. Apply the five steps: disaggregate the specific claims, check for incentive alignment, look for base rate anchors, evaluate the track record of the person making the claim, and identify what evidence would change your mind. Then share the claim and your analysis in the comments. Not to reach agreement, but to see what the framework surfaces that uncritical acceptance or reflexive skepticism would miss. The quality of your reasoning about AI will matter more over the next decade than which expert you happen to agree with today.