The Nobel Problem: How AI Is Breaking Scientific Attribution

The Nobel Prize for Chemistry in 2024 went to David Baker, Demis Hassabis, and John Jumper for protein structure prediction. Hassabis and Jumper built AlphaFold. Baker built a complementary system — RoseTTAFold and its successors — using different methods. The prize acknowledged both the AI breakthrough and the long preceding decades of structural biology that made it possible.

The Nobel Committee made a reasonable choice in a difficult situation. But the very existence of the difficulty — who gets credit when a large team, a complex model, and years of training data produce a scientific breakthrough — is now a permanent feature of the scientific landscape, and the Nobel Prize is the easy case.

Why Attribution Matters

Scientific credit is not just prestige. It is the currency through which the research enterprise allocates resources. Citations determine impact factors, which determine journal prestige, which determines where researchers publish, which determines tenure cases, which determines who gets to run labs, which determines what questions get studied. Grant funding follows citation metrics. Hiring decisions are based on publication records. The entire system of incentives that shapes scientific careers is built on the assumption that individual researchers (and, to a lesser extent, their institutions) are the primary unit of scientific production.

AI-assisted research challenges this assumption at multiple levels.

At the most basic level: when an AI model identifies a compound, predicts a structure, or proposes a synthesis route, and a research group then validates it experimentally and publishes the result, what is the appropriate attribution? The current answer is that the human researchers are listed as authors, the AI model is mentioned in the methods section, and the paper is credited to the institution. This mirrors the treatment of other research instruments — nobody lists the mass spectrometer as an author.

But a mass spectrometer does not generate hypotheses. When the AlphaFold model correctly predicted the structure of a protein that then turned out to be a drug target, the scientifically significant step — the one that required something like scientific insight — was performed by the model, not by the researchers who ran it. The researchers made important contributions: they chose the protein to study, validated the prediction, and understood what the structure implied for biology. But the structural prediction itself was model-generated.

The Ghost Author Problem

A more immediate and practical issue: ghost authorship of scientific papers by AI language models.

In 2023, several papers submitted to and published in peer-reviewed journals contained text that was clearly LLM-generated — not assisted but generated, with the characteristic sentence patterns, the overuse of specific transition words, the occasional hallucination of references. Some of these were caught by editors using detection tools (which are themselves imperfect). Many were not.

The scientific community’s response has been norm-setting: journals have adopted policies ranging from “AI cannot be listed as an author” (essentially universal) to “AI assistance must be declared” (increasingly common) to “AI-generated text is prohibited” (a minority, and unenforceable). The International Committee of Medical Journal Editors issued guidance that AI cannot be an author and that human authors are responsible for AI-generated content. This is sensible and inadequate.

The inadequacy comes from the verification problem. A declaration that “sections of this manuscript were drafted with AI assistance, subsequently reviewed and edited by the authors” is unverifiable. There is no reliable technical test for AI assistance versus AI generation. The detection tools that exist have high false positive rates on non-native English writing, which creates an obvious equity concern, and have been shown to be evadable by simple paraphrasing. The policy relies entirely on researcher honesty, which historically has been an insufficient constraint when career incentives point the other way.

The Authorship Inflation Problem

A related but distinct issue: authorship lists in AI-assisted research are getting longer.

This is partly because large computational projects genuinely involve large teams — the AlphaFold 2 paper had 24 authors, the paper describing the single-cell atlas of the human body had 250. It is partly because including more people on a paper has become a social norm in large collaborative projects, which has its own distorting effects on how citation-based metrics measure individual contribution.

But there’s a specific AI-driven dynamic: the researchers who built and trained the models whose predictions enabled the discovery feel they have a legitimate claim to credit on papers that use those predictions. This is not unreasonable — they contributed something scientifically significant. It is also not how authorship has traditionally worked, which treats model-builders as analogous to equipment manufacturers rather than collaborators.

Several journals have introduced contribution statements as a partial response — standardized descriptions of what each author actually did (data collection, model development, experimental validation, writing, etc.). These are better than nothing and less informative than they appear, because the categories are broad and the process of assigning them is unaudited.

Who Benefits from the Status Quo

The current attribution system, with its ambiguity and unenforceable disclosure requirements, is not neutral. It systematically benefits large, well-resourced groups who can deploy powerful AI tools while keeping the resulting publications on their curriculum vitae with minimal transparency about the source of the scientific insight.

It benefits groups at prestigious institutions, whose papers get more downloads and citations regardless of content quality. It benefits researchers who got early access to powerful models through industry collaborations, which are heavily concentrated in a few universities.

It disadvantages researchers who are honest about the degree of AI contribution — who write methods sections that accurately describe model-generated hypotheses and robotic experimental execution — because reviewers apply a discount to papers that seem less intellectually intensive, even if the results are equally valid.

And it disadvantages early-career researchers in a specific way: the currency of an academic career is publications with your name prominently placed, demonstrating independent scientific judgment. If the intellectual heavy lifting is increasingly performed by AI models deployed by the senior researchers who control access to those models, early-career researchers may be doing the experimental validation work (still essential, but less visible) while the model-level contribution goes to the PI’s CV.

What Would Fix It

The structural fix requires changing what academic credit tracks. Citation counts and author position are proxies for scientific contribution that made sense when papers were primarily intellectual products of the researchers named on them. The proxies are failing.

A credible alternative would track specific contributions — not just “data analysis” in a contribution statement, but which specific analysis, on which specific data, using which specific tool. This is technically feasible; it is a richer version of what data science pipelines already do with reproducibility. It requires journals to collect and display this information, funders to consider it in grant evaluation, and institutions to use it in tenure decisions. None of these institutions have yet moved.

The Nobel Committee’s solution — wait until a discovery has been validated, understood, and clearly attributed, then honor the humans most responsible — works at decadal time scales for paradigm-shifting contributions. It says nothing about the distribution of credit for the thousand smaller AI-enabled discoveries happening every year that collectively form the infrastructure of the next paradigm shift.

Those discoveries are happening. The credit for them is, largely, going to whoever published the paper, which is increasingly whoever had access to the model that made the discovery possible. That is not a sustainable equilibrium.