Scientists React to AI Breakthrough With Shock, Vindication and Regret

In this article

A Monday That Broke Three Careers
The Hypothesis at the Center of the Storm
How Scientists React to AI Overturning ETH: Three Answers
Who Reviews the Machine?
Who Owns the Proof?
A Major Revolution in the Field
What Scientists React to AI Results Teach Future Researchers
What This Changes for Everyone Else
References

A Monday That Broke Three Careers

The Announcement Lands

Ask how scientists react to AI breakthroughs, and you usually get polished abstractions. On Monday, three of the people whose careers are woven into one of computer science’s central conjectures answered the question personally, and the answers were shock, vindication and regret.

OpenAI and Anthropic each announced that an internal model had disproved the exponential time hypothesis, the long-standing assumption that some problems are so intrinsically hard that the only way to solve them is to try nearly every possible answer. Hundreds of published studies lean on that assumption. So do entire research programs — including three belonging to the men who created, refined and doubted the hypothesis over twenty-five years.

The interview with AFP captures a profession watching its most private skill, mathematical judgment, arrive somewhere else. Paturi spoke of surprise. Williams, a well-documented skeptic, spoke of vindication. Impagliazzo, the conjecture’s co-author, spoke of regret: he would have preferred a human refuter whose hand he could shake.

What makes the moment different from earlier AI milestones is who is reacting. When a chess engine wins, the players are spectators. When a model overturns a conjecture that a living scientist spent decades defending, scientists react to AI as authors of the affected literature — with grief, relief and hard questions about what comes next.

The reactions came in an AFP interview that is quickly becoming one of the most-read documents in theoretical computer science. Its value is not only what the three scientists said, but the order in which they said it: first the surprise of a builder, then the relief of a skeptic, and finally the regret of a man watching his name attached to a conjecture a machine may have outgrown. Read closely, how scientists react to AI results here doubles as a map of the discipline’s fault lines.

None of this happened quietly. The announcements landed as part of a wider wave of lab output in mathematics, and the response spread through mailing lists, group chats and conference hallways before the first blog post was drafted. What follows unpacks the three reactions, the review the three scientists are now running together, and the questions about ownership and access they insist the community must answer for itself — and why how scientists react to AI news now shapes it.

A Quarter-Century, Undone in Days

Here is the measure of how quickly the ground moved in October 2026 — the week in which scientists react to AI became a question about themselves, and a quarter-century of settled assumption was undone inside days:

One conjecture, two announcements, four days

Years since Impagliazzo and Paturi proposed ETH 25

Days from OpenAI’s announcement to this interview 4

Frontier labs claiming an independent refutation 2

The Hypothesis at the Center of the Storm

The 1999 Paper That Started It

Russell Impagliazzo and Ramamohan Paturi of the University of California San Diego proposed the exponential time hypothesis around twenty-five years ago, growing it out of their 1999 study of k-SAT. The hypothesis is not one dramatic claim but a precise bet about the shape of hardness: certain constraint problems cannot be beaten by more than exponential factors, no matter how clever the algorithm.

The Skeptic in the Room

From the beginning, the field also harbored its doubters. Ryan Williams of MIT, whose own lower-bound results reshaped the field, has spent years publicly doubting that ETH is true — which is exactly why watching scientists react to AI vindicate his skepticism carries an edge he does not hide.

What ETH Actually Says

The conjecture is anchored to 3-SAT, the canonical problem of satisfying a Boolean formula written as clauses of three variables. ETH says every algorithm for 3-SAT must spend roughly 2 to the power of a constant times the number of variables. A single algorithm growing as 2^o(n) — slower than any true exponential — would refute it outright.

Term	Meaning	Why it matters here
3-SAT	Satisfying a Boolean formula in three-variable clauses	The benchmark problem every hardness claim is calibrated against
ETH	3-SAT needs 2^(c·n) time for some constant c above zero	Proposed by Impagliazzo and Paturi; a quarter-century of research rests on it
2^o(n)	Growth slower than every true exponential function	A 2^o(n) algorithm for 3-SAT would refute ETH in one stroke
Lower bound	A proof that no algorithm can beat some difficulty limit	Ryan Williams’s research area, and the tradition the AI claims challenge

If both announcements survive scrutiny, the conjecture falls — and with it, the default assumption embedded in hundreds of papers across theoretical computer science and biology that treat ETH as bedrock.

The hypothesis’s quiet power is easy to underestimate. ETH belongs to a family of assumptions — alongside P versus NP and the random oracle model — that researchers use as scaffolding: when you cannot prove a problem is hard, you borrow hardness from a conjecture everyone agrees to treat as safe. Papers across cryptography, fine-grained complexity and even biology cite ETH as the reason their optimisation targets are unbeatable. Undo the scaffold, and every borrowed hardness claim must be re-examined — which is part of why watching scientists react to AI touch this particular conjecture feels like watching an audit of the field’s foundations.

There is also a human geometry to the storm. Impagliazzo and Paturi have collaborated since their student years, and Williams built his reputation partly by attacking the kind of barriers they formalised. The three have argued for two decades over whether ETH-style assumptions are the right way to model hardness. Being handed the answer by a commercial model, rather than by each other, is a twist none of them appears to have scripted.

How Scientists React to AI Overturning ETH: Three Answers

Surprise: Ramamohan Paturi

Paturi’s first reaction was surprise, but he refused to narrate the end of his own field. ETH has shaped a great deal of research over three decades, he told AFP, yet the research built around it would not simply disappear. Breakthroughs extend the frontier, he argued — and shaping that frontier remains within human capacity and creativity.

Vindication: Ryan Williams

Williams put it bluntly: he never fully believed the hypothesis, he is on the record as a prominent skeptic, and so he felt vindicated. It is a reminder that even among colleagues, how scientists react to AI results tracks their prior bets. The man who wagered against ETH for two decades now watches the machine cover his position.

Regret: Russell Impagliazzo

Impagliazzo, who shares authorship of the conjecture with Paturi, was candid about being less cheerful than his co-author. He would have much preferred the refutation to come from a human who could explain their thought process — someone whose hand he could shake. For him, the method of a discovery is inseparable from its meaning.

Scientist	Ties to ETH	Reaction	In their words
Ramamohan Paturi, UC San Diego	Co-proposed ETH around 25 years ago	Surprise	ETH has shaped a great deal of research over the past three decades
Ryan Williams, MIT	Longtime public skeptic of ETH	Vindication	I did not fully believe the hypothesis… so I felt vindicated
Russell Impagliazzo, UC San Diego	Co-proposed ETH around 25 years ago	Regret	I would have much preferred it to be discovered by a human who could explain their thought process

Set side by side, the three answers show why generalisations about how scientists react to AI breakthroughs so often miss the point. Reaction is not a property of the technology; it is a property of the relationship between the researcher and the result. The builder, the doubter and the author had every reason to feel differently, and they did — in public, on the same day, in the same interview.

Who Reviews the Machine?

The Unlikely Review Panel

The people now organizing the independent review of the AI labs’ claims are the hypothesis’s co-author, its co-architect and its loudest critic. Impagliazzo, Paturi and Williams have agreed to share the work of deciding whether a machine has genuinely overturned the idea one of them invented and another spent a career doubting. Putting the most motivated, most qualified and most adversarial readers on one panel is a template, arguably, for how scientists react to AI claims worth checking — and if the machine’s construction survives that, its truth will be something no press release could have manufactured.

What Reviewers Will Check

Refutation of a hypothesis this strong needs only one counterexample — in this community’s telling, an explicit 2^o(n)-time algorithm for 3-SAT — but that counterexample must be airtight. The labs have published lengthy preprints and code, along with machine-checkable formalizations intended to let skeptics verify every step without trusting the models.

Why This Review Matters

The outcome will set the standard for every AI-made scientific claim that follows. Our companion explainer, Exponential Time Hypothesis Refuted? What Frontier AI Labs’ Claim Means, walks through the competing constructions in detail, and the wider sweep of results is covered in OpenAI’s announcements across mathematics, computer science and physics.

How scientists react to AI claims under this kind of scrutiny is itself part of the experiment: the discipline is testing whether machine output deserves the same slow, adversarial review it gives humans — and whether that review can move at the speed of the announcements.

Expect the review to move slowly on internet time. The community has every incentive to take weeks or months: the cost of rubber-stamping a wrong refutation is far higher than the cost of delay. And the review will not stop at the main idea — the machine-checked formalisation lets critics attack each lemma individually, which turns verification into something closer to code review than to reading a paper. That shift is one of the clearest signals yet of how scientists react to AI claims becoming routine: not with awe, but with checklists.

Who Owns the Proof?

Announcements Before Peer Review

All three scientists raised a shared worry: these results come from models owned by companies, and the companies decide what gets published and when. A press release is not a refereed paper. Paturi put it plainly — a company’s announcement cannot substitute for the community’s assessment.

Access to the Tools That Produce Proof

Publicly releasing a proof is important, Paturi argued, but it does not provide equal access to the tools that produced it. Concentrated access to powerful research systems raises serious questions about who can participate, who gets early access and who shapes the research agenda. Watching scientists react to AI ownership of results is watching a community notice its instruments moving behind commercial glass.

The Thought-Process Gap

When a human solves a problem, Impagliazzo observed, they share not just the solution but the thought process that led to it — often more valuable than the answer itself. Access to the model itself is the closest analogue we have. Williams kept it shorter: yes, it is worrisome, and it is on the AI companies to be responsible about their power.

Dimension	The company path	The community path
Timing	Labs choose what and when to announce	Reviewers set the pace of acceptance
Access	Powerful models sit behind corporate accounts	Papers, artifacts and formal proofs stay public
Legitimacy	An announcement cannot certify a result	Independent assessment is the gold standard
Pedagogy	The answer ships; the reasoning may not	The thought process is part of the contribution

There is also a practical asymmetry inside the worry. Papers are free; frontier models are not. A biologist in Nairobi and a theorist in Nantes can read the same preprint today, but only institutions that can pay for access can reproduce the discovery process itself — the prompts, the search strategies, the model’s false starts. When Paturi says the tools matter as much as the results, he is describing a slow-moving change in who gets to be a producer of science, not just a consumer of it. This is the access question hiding behind how scientists react to AI ownership of results.

A Major Revolution in the Field

Glamorous Work Versus Unglamorous Work

Until now, the same researchers mostly did both halves of their craft: the glamorous part, solving, and the less glamorous part, verifying and explaining. Impagliazzo’s fear is that AI now takes the first half, leaving humans the second — until next year, when AI progresses further and takes that too. He said plainly that he has no idea what research will look like next year.

Impagliazzo’s two-halves framing deserves more attention than it has received. Research has always rewarded solving, and hiring, tenure and prizes all track the glamorous half. If verification and explanation become the human share, the profession will have to decide whether that half is a service job or a calling — and whether it pays accordingly. It is an uncomfortable question precisely because the answer determines who stays, and how scientists react to AI-driven change for a generation will follow from it.

A Field Worth Joining Anyway

Williams reads the same scene as opportunity: we are starting to experience a major revolution in the field, and it is absolutely worth going into. Some people may want to leave; others may find the moment is finally interesting enough to arrive. Both responses are rational bets on where human attention will be most valuable.

Williams’s enthusiasm is not naivety either. A field in a revolution gets new problems faster than it gets new researchers, and the open questions generated by a machine-refuted ETH — what is actually hard now, and how fine-grained must our hardness assumptions become — are exactly the kind that fill dissertations. The revolution he welcomes is, among other things, a jobs program for skepticism.

Judgment as the Remaining Skill

Paturi’s advice to students is the most tactical of the three: learn to use these tools while developing your own judgment. If machines generate more candidate theorems than any human can read, the scarce expertise becomes choosing which problems matter and which proofs deserve trust — skills adjacent to the ones the field already rewards. Recent work on reverse problem generation shows why: machines excel at producing problems and solutions, but the question of which ones are worth asking is still stubbornly human. Watching scientists react to AI is how the field teaches itself what to do next.

What Scientists React to AI Results Teach Future Researchers

Students Are Watching the Reaction, Not Just the Result

Ask what scientists react to AI change by watching their mentors, and the honest answer is ambivalence on a human scale. Graduate students in 2026 can see a conjecture loved for twenty-five years fall in four days, and see its creators express pride, relief and mourning in the same interview. That emotional data shapes careers more than any policy memo.

The advice implied by the reactions is consistent even when the emotions differ: use the tools, but keep your judgment on a separate leash. Paturi tells students to build judgment alongside fluency; Williams tells them the field is worth joining; Impagliazzo, by mourning the missing thought process, tells them to study reasoning, not just answers. Students who hear how scientists react to AI from all three mentors get a curriculum none of them designed.

The Frontier Still Needs Humans

Paturi insisted the breakthroughs extend the frontier rather than abolish it, and he encouraged young students to join. The task of moving a frontier — deciding what questions deserve a decade — has not been automated away in any way the interview suggested. Even the machine’s result needs three world experts to certify what it means.

Trust But Verify, With Teeth

The independent review organized by Impagliazzo, Paturi and Williams is itself a lesson: expertise does not vanish the moment a machine outperforms it on one benchmark; it mutates into custodianship. A study on whether AI can pass for human in follow-up conversation underlines the same asymmetry — generating convincing output and owning its meaning are different jobs.

How scientists react to AI systems is therefore not a sideshoot of this story. It is the mechanism by which the result becomes knowledge: models propose, and the community, for now, decides.

There is also a morale point worth stating plainly: the scientists most affected by this result are not protesting the refutation. They are interrogating it, teaching it and organising its review. Whatever the future holds for human discovery, this is what professional maturity looks like in transit — and it is a better model for young researchers than either techno-utopianism or siege mentality.

What This Changes for Everyone Else

No Panic, and No Prize — Today

Impagliazzo was direct about the everyday stakes: there will be no immediate consequences for almost all people, as he put it, except his friends and himself. The result does not represent a security threat. Deployed cryptography does not fall because ETH falls, and no laptop suddenly solves everything.

Better Tools in Many Sciences, Eventually

The realistic upside arrives slowly. A first breakthrough is usually followed by improvements and generalizations, Impagliazzo said, so a genuine refutation of ETH might eventually yield better data analysis tools across many branches of science. Williams agreed the result is theoretical but opens the possibility of much more efficient computing across an enormous variety of scenarios.

Watch the Crypto Question Anyway

Both scientists flagged the same caveat: future extensions of this line of work might require stronger encryption. That is the sentence security teams should file, not out of fear, but because the next result may come on the same four-day clock.

One more consequence deserves naming: the sociology of the field has changed. Conjectures will now be attacked continuously by tireless, well-funded machine search, not occasionally by human visitors. Researchers will choose which open problems to keep close knowing a lab model may reach them first, and the community’s rituals — open letters, review panels, public repositories — are the new antibodies. How scientists react to AI pressure on famous problems will decide whether mathematics keeps its slow culture or acquires the lab’s clock.

For now, the lasting lesson of how scientists react to AI breakthroughs is measured in three words from three experts: shock, vindication, regret — and the discipline they chose to apply to a machine.