Bauhaus-style composition of red and yellow circles, text-like bars and a black diagonal line.

Illustration supplied by the author.

Harvard researchers compared AI-generated scientific reviews with unusually good human experts. The results make our favorite quality-control assumption rather uncomfortable.

There is one reassuring phrase that appears in almost every serious conversation about artificial intelligence.

Of course, a human must review the output.

Naturally.

AI hallucinates. AI makes mistakes. AI can be confidently wrong.

Humans, fortunately, possess the extraordinary ability to be confidently wrong while holding a university degree.

And so we have established a comforting hierarchy.

The machine produces.

The human checks.

The human is the final authority because, well, human.

A new Harvard working paper suggests that we might need to reconsider the arrangement.

Not because AI has become infallible.

But because the people supposedly protecting us from its mistakes may be making considerably more of their own.

An Awkward Experiment at Harvard

Gary King and Steven Worthington of Harvard University's Institute for Quantitative Social Science examined 124 academic manuscripts submitted to rapidPeer, a service providing expert feedback before papers are sent to journals.

This is where the study gets interesting.

The human benchmark wasn't a collection of exhausted academics performing unpaid peer review between teaching, meetings, and wondering why they ever pursued a research career.

The reviewers were carefully selected specialists.

They were well compensated.

They were instructed to read the manuscripts thoroughly and provide substantial, constructive feedback.

And their work was considered unusually good.

Among 74 authors who evaluated the reviews they received, 88% rated them as high or very high quality.

In other words, the researchers weren't comparing a modern AI system with the academic equivalent of a one-star customer support department.

They were comparing it with human experts operating under unusually favorable conditions.

The dataset contained 134 human reviews covering 124 manuscripts, collected between 2019 and 2026.

The researchers then generated AI reviews using Anthropic's Claude Opus 4.8, including a version equipped with a purpose-built review prompt and web search.

To assess the results, they used a different company's model, OpenAI's GPT-5.5, as a blind evaluator.

They also asked three former editors of leading scholarly journals to judge a randomly selected subset of ten manuscript-review pairs.

And the results were rather unfortunate for the traditional assumption of human superiority.

The Machine Won. Quite Decisively.

The AI judge preferred the tuned AI-generated reviews in all 132 comparisons where it reached a consistent decision.

The human editors also preferred the AI reviews in 26 of their 30 judgments.

That's 87%.

And this is particularly important: the figure comes from actual human editors evaluating reviews without knowing which were written by people and which by AI.

The advantage wasn't confined to grammar or pleasant formatting.

AI reviews scored better across the evaluated categories, including methods, weaknesses, theoretical contributions, framing, communication, tables, and sources.

There was also an accuracy audit.

The researchers checked individual claims made by the reviewers against the manuscripts.

The reported rate of incorrect claims was:

  • 5.5% for AI reviews.
  • 15.9% for human reviews.

Approximately three times as high for the humans.

Now, before anyone begins preparing a ceremonial retirement package for academic peer reviewers, an important qualification: these error judgments were primarily made by another AI model. The researchers performed additional validation, including human checks, but this was not a comprehensive independent human audit of every claim.

Also, more than a third of AI reviews still contained at least one flagged factual error.

So no, the machines haven't achieved perfection.

They have merely produced an awkward result for the proposition that humans are automatically the safer option.

Perhaps the Humans Just Needed to Sound More Like AI

My favorite part of the study concerns writing style.

One obvious objection to AI-generated reviews is that they are often extremely polished.

Neatly structured.

Comprehensive.

Confident.

Full of helpful observations delivered in that unmistakable tone of a consultant who has never missed a deadline or experienced existential despair.

Perhaps the AI judge simply preferred AI writing.

Perhaps the substance wasn't better at all.

Perhaps the humans lost because their reviews sounded too human.

The researchers tested this.

They took the human reviews and had Claude rewrite them in an AI-like style while preserving their substance.

Then they repeated the evaluation.

The human reviews still lost.

Their average score barely changed.

The AI-generated reviews remained comfortably ahead.

Which produces a wonderful conclusion.

Apparently, giving a human expert the writing style of an AI does not automatically give them the analytical performance of one.

Who would have thought?

It is also a reminder that polished writing and substantive quality are different things. The study attempted to separate them, and the advantage survived.

Of course, the experiment doesn't eliminate every possible bias in AI judging.

But it does make the simplest explanation, that the model merely preferred its own kind of prose, considerably less convincing.

There Is a Difference Between a Better Review and Better Science

This is where we should become appropriately skeptical.

The paper, dated October 28, 2026, is still a working paper and has not itself completed peer review.

Which is rather amusing, given the subject.

More importantly, the study measured the quality of constructive manuscript reviews.

It did not establish that AI is better at deciding which papers deserve publication.

It did not demonstrate that authors who follow AI feedback produce better final manuscripts.

And the human-editor experiment involved only ten papers, even though the full study used 124.

There is also a problem that deserves much more attention than the usual debate over whether AI can catch a statistical mistake.

Science is not merely the production of manuscripts that satisfy a checklist.

It involves recognizing unusual ideas.

Questioning accepted assumptions.

Taking unfamiliar methods seriously.

And occasionally realizing that an apparently absurd argument is interesting precisely because it challenges what everybody already believes.

The Harvard authors themselves warn that widespread reliance on similar AI systems could make research more homogeneous.

A model might improve individual papers while encouraging entire disciplines to become more conventional.

Better organized.

Better referenced.

More internally consistent.

And perhaps a little less willing to produce the kind of work that changes how a field thinks.

That is a genuine risk.

The ideal scientific reviewer is not necessarily the one who most efficiently explains how to make a paper resemble other successful papers.

Sometimes the best contribution is recognizing why a paper should be allowed to look different.

The Human in the Loop Is Not a Quality Guarantee

This brings us back to the original assumption.

AI output requires human review because humans are better at detecting errors.

That sounds responsible.

It sounds prudent.

It sounds like something a committee could unanimously approve after several meetings.

But it is not a universal empirical truth.

Human reviewers have biases.

They miss errors.

They become tired.

They misunderstand methods outside their expertise.

They sometimes give superficial feedback.

And they may confidently criticize claims that the manuscript never actually made.

The Harvard study reminds us that expertise is not the same thing as infallibility.

Nor is being human a quality-control mechanism.

If the goal is to detect factual inconsistencies, examine methodology, or identify weaknesses in a manuscript, it may be sensible to let an appropriately configured AI system perform much of that work.

Then use human expertise where it contributes something distinctive.

Judgment about scientific importance.

Responsibility for editorial decisions.

Understanding disciplinary context.

Recognizing unusual but potentially valuable ideas.

And evaluating situations in which a machine's confidence exceeds the evidence available.

This does not mean human judgment always wins in those domains. That would be another assumption requiring testing.

It means the workflow should be designed around demonstrated strengths and limitations rather than a sentimental belief that human supervision automatically improves everything.

Perhaps the future is not AI producing and humans checking.

Perhaps it is AI checking humans, humans checking AI, and both being subject to independent verification when the stakes justify it.

A considerably less flattering arrangement.

But possibly a more useful one.

Who Should Review Whom?

There is a larger lesson here.

We've become accustomed to describing AI as a brilliant but unreliable assistant.

Something capable of impressive work that must always be supervised by a more trustworthy human.

And sometimes that description is accurate.

But it also reflects a historical arrangement rather than a permanent law of intelligence.

The assumption was reasonable when AI systems routinely made elementary mistakes that competent professionals could detect in seconds.

The question is whether the arrangement should remain unchanged when the measured capabilities change.

Because the purpose of review is not to preserve the dignity of the reviewer.

It is to improve the quality of the work.

If a machine catches more errors, provides more useful feedback, and does so consistently under the conditions being tested, making a human the mandatory final checker may not automatically improve the outcome.

It could even make it worse.

The answer is not to eliminate accountability or place blind faith in the next impressive model benchmark.

It's to stop treating human judgment as an unquestionable gold standard.

Human expertise is valuable.

It should remain valuable for what it actually contributes.

Not because it belongs to a species we happen to be members of.

And perhaps the most amusing detail in all of this is that the researchers tested whether rewriting the human reviews in AI style would help.

It didn't.

Apparently, the problem wasn't how the humans sounded.

It was what they had to say.

So perhaps the next time someone insists that a human must review the AI because humans are better at spotting mistakes, we should ask a slightly inconvenient question:

Who reviewed that assumption?

Study notes

The 5.5% and 15.9% figures are the proportions of assessed claims flagged as contradicting the manuscript, not overall review-error rates. The accuracy audit used the matched AI review without external tools; unverifiable claims remained in the denominator. The 26-of-30 editor judgments covered ten manuscript-review pairs, and the editors did not receive the manuscripts themselves. The authors disclose that they conceived and run rapidPeer. The paper is a working paper.

Sources

Keep reading

Explore the writing archive →

Stay in touch

More essays and updates on ABVX Substack.

Read and subscribe on Substack →