Comparative synthesis
Cross-profile comparison
Working synthesis. This page compares six public records using the same questions. It is an interpretation of the profile reports, not a claim that the six people share one position or that the reports establish private beliefs.
Scope
The comparison covers Dario Amodei, Sam Altman, Demis Hassabis, Yoshua Bengio, Ilya Sutskever, and Geoffrey Hinton. The source windows differ because the strongest attributable evidence begins at different times. The common cutoff is September 2026.
At a glance
| Figure | Most stable through-line | Important movement in the record | Latest located intervention threshold |
|---|---|---|---|
| Dario Amodei | Empirical safety research tied to increasingly capable systems | From tractable failure modes and evidence caution to mandatory testing and explicit frontier pacing | Pace frontier capability growth when current incidents and capability trends make safety lag consequential |
| Sam Altman | Large upside plus non-zero catastrophic risk; no permanent halt | Shorter timelines, iterative deployment, regulatory retrenchment, then a September 2026 pacing endorsement | Do not push much further without better monitorability, alignment, and independent evaluation |
| Demis Hassabis | Extraordinary scientific benefit conditional on control and stewardship | From technical safeguards and self-governance to staged deployment, standards, and an option to coordinate a slowdown | Preserve the option of coordinated slowdown if frontier risks warrant it |
| Yoshua Bengio | Human-centred, verifiable, democratic AI governance | A self-acknowledged 2023 change after ChatGPT/GPT-4; stronger licensing, safety cases, and Scientist AI | No frontier training or deployment without a safety case persuasive to independent experts |
| Ilya Sutskever | Capability optimism coupled to progressively lower tolerance for uncontrolled power | Conditional release, superalignment, safety-before-productisation, then incremental deployment and power limits | Safety and control should advance at least as fast as capability; deployment can be used to learn, cautiously |
| Geoffrey Hinton | Concrete safety concerns plus growing catastrophic-risk precaution | Timeline compression, stronger regulation, conditional restriction of open weights, and a superintelligence controllability threshold | Do not develop superintelligence without evidence of safety and controllability plus broad legitimacy |
A shared baseline: benefits do not erase risk
All six records retain a public commitment to substantial benefits from AI, especially scientific, medical, or productivity gains. None of the profiles supports the crude summary that “AI should stop” without qualification. Their disagreement is about which systems should face exceptional controls, how much evidence is enough, and whether the frontier has crossed the relevant threshold.
A second shared pattern is capability sensitivity. The profiles distinguish ordinary or lower-capability systems from frontier systems that may have access to tools, autonomy, biological or cyber capabilities, or the ability to influence institutions at scale. That is why openness, licensing, audits, and deployment restrictions appear in conditional rather than blanket form.
The deepest difference is the evidence threshold
Amodei’s record places unusual weight on observable failure modes and empirical demonstrations before prescribing strong rules, even while allowing that uncertainty can justify preparation. His 2026 position is presented as an evidence-triggered escalation.
Bengio and Hinton make the stronger precautionary case that waiting for conclusive evidence is unreasonable when the potential harm is irreversible. Bengio’s profile documents a clear self-described change after the acceleration of 2022–23; Hinton’s profile documents a sharp timeline revision and a growing burden-of-proof principle.
Altman, Hassabis, and Sutskever occupy intermediate but distinct positions. All three have emphasized learning through deployment or testing, but all three also accept capability thresholds and stronger safeguards as stakes rise. Their public records therefore complicate the binary between “deploy and learn” and “pause until solved.”
Deployment and the role of real-world learning
Altman’s strongest mid-period argument is that gradual release lets developers and society learn from real use rather than attempt to perfect systems in secret. Sutskever’s 2025 revision similarly gives incremental deployment an epistemic role, after an earlier safety-before-productisation stance. Hassabis’s record also supports delayed or staged release, with more control over the weights of the most powerful models.
Amodei’s framework is less about ordinary product iteration and more about testing dangerous capabilities, independent evaluation, and whether safety methods generalise as systems become more agentic. Bengio’s Scientist AI proposal tries to change the object being built: systems designed to understand and predict without the same autonomous goal pursuit. Hinton’s record is the least tolerant of making the public frontier the experiment when controllability has not been demonstrated.
Governance is not one axis
The profiles show at least four governance choices that are often collapsed together:
- Who evaluates? Internal teams, independent experts, third-party auditors, or public regulators.
- What triggers intervention? Compute, capabilities, observed dangerous behaviour, a safety case, or a defined superintelligence threshold.
- What is restricted? Deployment, tool access, model weights, source code, compute, or the rate of capability improvement.
- What is the institutional level? Company, national regulator, international standards body, or a treaty-like authority.
Bengio and Hinton generally place the heaviest burden on frontier developers and the strongest role on statutory or international oversight. Amodei and Hassabis emphasize standards, testing, and independent assessment. Altman’s preferred institutional design shifts more visibly over time: from oversight and an IAEA-like authority, through scepticism about a new licensing agency, to international standards and direct pacing. Sutskever’s record is strongest on technical alignment, controlled release, coordination, and infrastructure security rather than on one fixed regulator.
Openness is conditional across the record
The six profiles do not map neatly onto “open” and “closed.” Altman’s personal record includes a clear 2025 criticism of OpenAI’s open-source stance, while still retaining capability thresholds. Bengio and Hinton support open inquiry and safety knowledge but distinguish that from releasing dangerous frontier weights that cannot be recalled or patched. Sutskever’s GPT-2 and Safe Superintelligence records support staged or closed release for the most capable systems. Amodei and Hassabis likewise treat frontier access as a safety variable rather than an absolute principle.
The useful comparison is therefore not who is “for openness,” but which forms of access each person treats as safe at which capability level, and whether a release can be reversed after a vulnerability is discovered.
The 2023–26 convergence—and its limits
The field’s public language changes materially after 2022–23. Catastrophic risk becomes more explicit; capability thresholds, safety cases, external testing, interpretability, monitoring, and international coordination become more operational. By September 2026, Amodei explicitly argues for pacing, Altman says he agrees that the frontier needs pacing, Hassabis says Amodei is pointing in the right direction, and Bengio’s conditional-pacing position is already established. Hinton’s latest record is stronger still at the superintelligence threshold.
This is a convergence in the vocabulary of precaution, not proof of a common policy. The disagreements remain consequential: Amodei’s evidentiary discipline differs from Bengio’s precautionary burden of proof; Altman’s and Sutskever’s deployment-as-learning logic differs from Hinton’s tolerance for experimentation; and “pacing” can mean anything from slower capability growth to a conditional prohibition on a defined capability destination.
What this pilot can test
The value of the Space is not only the six conclusions. It is the reusable method: choose a public figure, define a date window, privilege primary sources, mark attribution, distinguish assertion from fact, record changes and tensions, and make the evidence trail inspectable. If the format works for this pilot, additional figures can be added as peer pages without changing the structure.