It seems that Pangram and other AI-detection tools are increasingly being used to reveal to what extent texts have been written by AI as opposed to human beings. I find the usage of such tools problematic for many reasons; here, I want to focus on scientific texts and point out two such reasons: that such tools do not differentiate between two different ways of using AI for writing and that such tools only focus on writing and not on other parts of the scientific process. I will explain why I find each shortcoming problematic for taking a normative position toward scientific papers on the basis of an AI-detection tool having identified them as AI-written. I finish with a discussion of how to acknowledge how AI has been used in scientific work.

My main concern is with the inference that is often drawn from the information provided by an AI-detection tool. A reader sees a high detection score, concludes that the text is “slop”, and stops assessing it with an open mind. Let me call this the AI-slop reflex. The point I want to make is that a finding that the final wording of a paper is largely AI-written does not by itself show that the paper lacks substantial human intellectual input, that the paper is of low quality, or that it should be rejected or flagged as “slop”.

For the sake of argument, I will assume that the AI-detection tool is fully accurate. Questions about false positives and the reliability of different tools are important, but my argument does not depend on them. Even a perfect detector of AI-written text would not reveal who developed the ideas, made the scientific judgments, checked the analysis, or took responsibility for the paper.

Two different ways of using AI for writing a text

Let us assume that we have a text that an AI-detection tool has identified as 90–100% AI-written. The premise among many of those favoring the use of this kind of tool seems to be that AI-written texts are of low quality and that AI-written texts should consequently be purged, rejected, or flagged as “slop”. My argument is that AI-written texts can be of high quality and that they should not be purged, rejected, or flagged as “slop” simply for having been identified as AI-written.

Let us differentiate between two types of AI-written texts: those written with full AI use and those written with interactive AI use. The former cases involve a human just asking AI to put together a paper, and AI does so. The AI-detection tool flags this as AI-written. The latter cases involve a human working interactively with AI to produce a paper, where the human provides the research question, the outline of the study, key points of the overall argument, tentative text segments, and key references, and through an intense combination of effort, and through hundreds if not thousands of iterations, a paper is produced. The AI-detection tool flags this as AI-written.

My question is: Should these two papers be assessed in the same way? The AI-detection tool does not differentiate between them. For someone who thinks scientific credit and quality are related to active human intellectual involvement, the interactive case seems worth taking seriously as a potential scientific contribution, while the full AI-use case does not.

Of course, the number of iterations is not in itself decisive. A person can spend a great deal of time producing a weak paper. The central questions are whether humans came up with the central research questions, made the important scientific choices, understood and checked the AI’s suggestions, verified the claims and references, and can explain and defend the final paper. An AI-detection tool applied to the final text cannot answer these questions.

The point is that reliance on AI-detection tools gives a simplistic number that does not reveal the process behind the AI writing or the degree of human involvement. Nor does it indicate quality. It reveals something about the likely source of the final wording, but the source of the wording is not the same thing as the source of the intellectual contribution.

This distinction is familiar from collaboration among humans. One co-author may write almost the entire text, while the other co-authors develop the research question, formulate the theory, assemble the data, conduct the empirical analysis, or interpret the results. We would not conclude that the person who wrote the sentences made the entire scientific contribution. Human authorship is not normally allocated according to who produced each word.

One can add that a third type of paper, the paper written fully by a human, can be of low quality. There is human “slop”, just as there is AI “slop”. By relying on an AI-detection tool, there is a major risk that double standards are applied when papers are assessed through the inaccurate premise that AI-written text is necessarily of low quality, while human-written text is not.

What about the possible counterargument that AI-written texts, as revealed by AI-detection tools, are of lower quality on average? If that is the case, and it might be, then we have to ask ourselves whether we want journals and other assessors of papers to engage in statistical discrimination or not. Assume, for example, that papers submitted by researchers at highly ranked universities are of higher quality on average than papers submitted by researchers at less prestigious universities. Should therefore reject a paper from a less prestigious university, or should they assess the individual paper on its merits? Should such a paper be flagged as “written by someone at a low-ranked institution”?

In my view, the same principle should apply to AI-written and human-written papers. There are bad and good papers in each category, and each paper should be evaluated on its scientific merits. Papers that an editor at a journal thinks uphold a certain, journal-specific quality level should go into the review process and be assessed fairly, no matter what an AI-detection tool shows.¹

AI is not only used for writing

Anyone familiar with how scientists work in 2026 knows that AI is used a lot. However, only one particular usage is captured by AI-detection tools: the writing of the paper. Everything else, including AI-proposed text that a human modifies, can be AI-produced, but that is not captured by the detection tools. A paper that is flagged as written 90–100% by AI can have had a human do everything but the actual writing, and that risks not being recognized as a proper scientific contribution by those favoring AI-detection tools.

The point is that we do not know the extent to which AI is involved in the scientific process, and it may be involved at each stage. AI can help formulate research questions, identify gaps in the literature, suggest hypotheses, develop theoretical arguments, derive proofs, write and debug code, clean data, propose empirical specifications, interpret results, suggest robustness tests, produce figures, and respond to comments from editors and referees.

None of this use will normally be visible to a tool that analyzes the final text. A paper written in an unmistakably human style could rely heavily on AI for its ideas, analysis, code, and interpretation. Another paper could contain text largely written by AI even though the human researchers developed the question, theory, design, evidence, and conclusions. The detection score could rank the second paper as involving more AI, although AI may have played a much more bigger role in the first.

If the only metric available deals with “who wrote the text”, we risk treating those whose relative strength may not be in writing (or whose time is better spent working on new scientific problems than writing) badly, while not noticing all the AI use going into the papers of people who choose to write the papers themselves.

Admittedly, writing is often part of thinking. The process of formulating an argument in words can reveal conceptual problems, force us to be more precise, and generate new insights. Delegating the writing can consequently involve delegating some intellectual work. Yet, an AI-detection tool cannot reveal how much thinking was delegated. It cannot distinguish between a researcher who accepts generated text without understanding it and a researcher who scrutinizes every sentence, rejects mistaken suggestions, restructures arguments, verifies claims, and takes responsibility for the final result.

In this way, AI-detection tools risk illustrating the joke referred to as “the drunkard’s search”. A person out walking at night comes across a drunk man scrabbling on the street under a lamppost. The man on the street says he lost his keys. When asked where he dropped them, he replies, “Oh, I dropped them over there, but the light’s better here”. Just because there is a “light” in the form of AI-detection tools does not mean we find what we need to find.

The light happens to shine on the written text, but the scientific contribution extends far beyond the writing of the text. Using the visible part as a measure of the whole process creates a distorted picture even when the visible part is measured perfectly.

AI contribution, transparency, and human responsibility

As I think is clear, I think the scientific community should accept AI-written papers in principle. They should be assessed on their merits, and if AI-written papers advance the frontiers of science, they should be warmly welcomed, as should human-written papers that do the same. AI-detection tools are not helpful here. They can be harmful from the point of view of scientific advancement if people rely on them simplistically.

There remains a problem which I think merits a serious discussion, viz., how the contribution of AI should be recognized when an AI-written paper has been produced in the interactive way described above. The human researchers put their names on the paper, although the paper may be the result of extensive interaction between humans and AI.

One can think of human co-authors working with each other. Such collaboration is usually intense and iterative, and also based on comparative advantage, just like collaboration with AI can be. An AI system may propose arguments, identify weaknesses, formulate text, suggest analyses, or improve explanations. If a human collaborator made the same contributions, these contributions could form part of the case for co-authorship.

Co-authorship, however, involves more than contributing to a text or an analysis. Human co-authors can approve the final paper, answer questions about it, correct errors, and take responsibility for the integrity of the work. AI cannot presently do these things in the same sense, and at present, journals do not accept AI as co-authors.

I think it is clear that human authors must accept full responsibility for a paper. This entails understanding and verifying what AI has contributed, as well as making active decisions about what to include. They should be held responsible for all content, including content proposed or written by AI.

The model that is being used at the moment is AI-use declarations, which could be required to specify what aspects of the scientific AI has assisted with and a declaration by the humans involved that they accept full responsibility for the entire work. This gives a much richer type of information about the role of AI throughout the scientific process, not just the writing.

However, this model creates what could be called a transparency trap. Studies show that people evaluate writing less favorably when they are told that AI was involved in producing it. For example, Li et al. find that disclosing AI assistance reduces average quality ratings, especially when AI generated new content. Raj et al. report a persistent AI-disclosure penalty across 16 preregistered experiments.

This gives authors reason to fear that complete honesty will cause editors, reviewers, and readers to evaluate their work less favorably. An author who discloses extensive interactive use may be penalized, while an author who conceals the use or uses AI extensively at less visible stages of the research process escapes the same penalty. A system of this kind rewards concealment and punishes transparency. Full disclosure becomes a kind of “honesty tax”.

If this means that declarations are not fully correct, are AI-detection tools then not a solution after all? I think not – in fact, the risk is that AI-detection tools reinforce the problem. Once a paper is classified as AI-written, the classification can affect how it is perceived. A high score that is formally presented as neutral information can become a quality judgment in practice. Researchers may then conceal their use, rewrite good text just to make it look more human (an inefficient use of scarce time), or avoid useful AI assistance altogether.

These incentives are especially unfortunate for researchers who are less comfortable writing in English. A scholar may have developed the research question, collected the data, conducted the analysis, and interpreted the results, but used AI to express the findings in clear English – only to be flagged by a detection tool. A fluent and linguistically talented writer who uses AI for code, analysis, and interpretation may leave no corresponding trace just by virtue of being able to write correct English.

So while declarations are probably the best way forward, journals need to be aware of the transparency trap. To ease the problem, my recommendation is to treat declarations as descriptive accounts of how the research was produced. They should not be used to downgrade a paper, infer low effort or quality, or question the abilities of the human authors; also, information about AI use should not be given to referees before they assess the scientific content.

Whether AI should eventually be recognized formally as a co-author can remain open for discussion. The answer depends partly on what we think co-authorship is for. It can represent intellectual contribution, but it also assigns credit, responsibility, and accountability. I recognize that AI cannot presently satisfy the latter components in the same way as humans; but when used in the interactive way described above, it can be seen as a functional (but not a formal) co-author.

So let us skip the AI-detection tools and recognize that AI can benefit scientific progress at all stages of the process. Scientific assessments should focus on the validity, accuracy, and originality of the work, and hold the named human co-authors responsible for the content. The central question should be whether the paper advances scientific knowledge and withstands scientific scrutiny, not whether a detector states that AI produced its sentences.²

See my previous posts on AI use in science:
”Freedom, Determinism, and AI”
”On the AI Policy of an Academic Journal”
”Should Academic Journals Use Detection Tools and Auto-Reject AI-Written Work?”
”Beyond Human Prose: Philosophical Routes to Defending Full AI Use in Academic Research”

Also see my paper on how people in academia can argue for the use of AI-detection tools and the regulation of AI-tool use on self-interested grounds.

1. Editors may respond that they cannot assess every paper in detail. Editorial time and referee time are scarce, and AI has made it easier to generate large numbers of submissions. I agree that this is a potential problem – but is the use of AI-detection tools a good solution? What I fear will happen is that a cheap signal placed in front of a time-constrained editor will become a substitute for the more costly assessment it is said merely to assist. The risk is that the AI-slop reflex becomes a consequence of the incentives facing editors. ↩︎
2. I consulted ChatGPT on the key arguments of this post, and it also proposed some (in my opinion) improved formulations, which I incorporated. ↩︎

A common argument in the debate about whether and how AI should be used in academic writing is the Human Authorship Thesis, which says that a necessary requirement for a genuine and legitimate scientific contribution is that the scientific text was written by the human researcher(s), rather than produced substantially or entirely with the help of AI.

The idea is that the human production of the prose has intrinsic value, independently of whether the claims are true, the methods sound, the evidence adequate, and the named authors answerable for the final product. The thesis assigns a special metaphysical status to human composition as such. The act of writing by a human hand is treated as a locus of scientific value.

I reject this thesis on deterministic grounds, as argued here. I asked ChatGPT to summarize the determinist argument and specify other philosophical views that imply the rejection of the thesis. After all, I realize that most people are not determinists like me. It came up with the following twelve positions (all direct quotes from the ChatGPT reply to my request).

1. Determinism and the rejection of ultimate authorship

One route begins with determinism, or more broadly with skepticism about ultimate self-authorship. No one is causa sui. A researcher did not create his own intelligence, temperament, discipline, upbringing, language, education, or cognitive style. If the value of authorship depends on being the ultimate source of a text, human beings do not possess that status either.

This point does not require crude fatalism. Nor does it imply that responsibility practices should disappear. We can still hold researchers responsible because doing so serves epistemic and institutional purposes: it encourages care, honesty, verification, and corrigibility. But such responsibility is forward-looking and practice-dependent. It does not rest on the mystical idea that the human author is an unconditioned origin of the scientific text.

On this view, AI does not introduce the first mediation between the scientist and the final article. Human authorship was already mediated by causes outside the author’s control. The demand that the text be “really written by the human researcher” therefore rests on a picture of agency that cannot be sustained.

2. Consequentialism: judge research by its effects on knowledge

A second route is consequentialist. On this view, the value of scientific practice lies in its consequences: more truth, better explanations, more reliable predictions, more useful theories, fewer errors, and more effective criticism. The question is not whether the prose was generated by a human unaided by AI. The question is whether the use of AI improves or worsens the epistemic consequences of research.

This route does not automatically vindicate every use of AI. AI can hallucinate, flatten arguments, obscure responsibility, reproduce errors, or generate plausible nonsense. But these are contingent objections. They show that AI use must be governed by verification, accountability, and transparency. They do not show that human composition has intrinsic value.

If AI helps researchers produce clearer exposition, better literature searches, more systematic robustness checks, more precise code, better translations, or more accessible explanations, then a consequentialist has reason to favor its use. If it worsens reliability, the consequentialist has reason to oppose that use. The relevant standard is epistemic performance, not human purity of production.

3. Reliabilism and process epistemology

A related but distinct route comes from reliabilist epistemology. What matters for warrant is whether a belief-forming or claim-producing process is reliable. “A human wrote this” is not a reliability condition. Humans are biased, inattentive, overconfident, status-seeking, tribal, and error-prone. “AI helped produce this” is not a defeater either. AI systems can be unreliable, but they can also be embedded in reliable workflows.

The relevant object of assessment is therefore not the metaphysical identity of the producer, but the whole epistemic process: prompting, source selection, data analysis, robustness testing, replication, human review, adversarial criticism, and final endorsement. A human-only process can be unreliable. A human-AI process can be reliable. The authorship thesis mistakes a feature of provenance for a feature of warrant.

4. The distinction between discovery and justification

A classic philosophy of science route distinguishes the context of discovery from the context of justification. Scientific ideas can arise from many sources: dreams, analogies, accidents, conversations, instruments, simulations, or automated searches. Their epistemic standing does not depend on their origin. It depends on whether they survive scrutiny.

AI-generated hypotheses, AI-assisted derivations, and AI-drafted text should therefore be judged by the standards of justification: evidence, argument, coherence, robustness, replicability, and openness to criticism. The fact that AI contributed to the formulation of a claim does not settle whether the claim is justified.

This distinction directly undermines one version of the authorship thesis. If what matters in science is justification, then the origin of the prose cannot carry decisive epistemic weight. A badly argued human-written paper is not redeemed by its human origin. A sound AI-assisted paper is not discredited by its nonhuman assistance.

5. Scientific realism and anti-psychologism

Scientific realism offers another route. If scientific claims aim at truths about the world, then their truth does not depend on the psychology or identity of the person who wrote them down. The law, mechanism, estimate, theorem, or causal relationship under investigation is not made true by human authorship.

A Fregean anti-psychologism points in the same direction. The content of a thought is not the same as the mental act by which someone grasps it. Likewise, the scientific content of a paper is not identical to the psychological process by which the prose was produced. To treat human composition as constitutive of scientific value is to confuse the origin of an expression with the validity of what is expressed.

This does not eliminate the need for responsible authors. Someone must be answerable for the claims. But answerability concerns endorsement, understanding, and defense. It does not require unaided human sentence production.

6. Pragmatism: ideas as instruments

A pragmatist route rejects the fetishization of origin. Ideas, models, concepts, and texts are instruments for inquiry. Their value lies in what they enable us to do: investigate, explain, predict, criticize, coordinate, and revise.

From this perspective, AI is one more instrument in the history of inquiry. The relevant question is practical and experimental: does it help the community inquire better under appropriate norms? A blanket insistence on human-written prose treats one historically contingent method of producing academic text as if it had intrinsic epistemic standing. The pragmatist asks instead whether the tool improves the conduct of inquiry.

7. Functionalism about authorship

The strongest defense of AI use should concede something important: authorship does matter in science. But it matters functionally, not metaphysically.

To be an author is not necessarily to have typed the sentences. It is to occupy a role. The author formulates or endorses the research question, takes responsibility for the methods, understands the evidence, checks the claims, resolves objections, approves the final text, and remains answerable after publication. These functions can survive extensive AI assistance.

This position is especially useful because it does not require determinism, skepticism about free will, or radical anti-humanism. It allows a robust conception of agency while denying that agency must be located in unaided prose composition. The morally and epistemically relevant question is not “Did the human write every sentence?” It is “Can the human author responsibly stand behind this work?”

8. Virtue epistemology: intellectual character, not manual production

Virtue epistemology provides another moderate route. On this view, good inquiry depends on intellectual virtues: honesty, rigor, curiosity, humility, courage, patience, and openness to criticism. But these virtues need not be exercised through unaided writing. They can be exercised in designing prompts, checking sources, testing claims, detecting weaknesses, revising arguments, disclosing assistance, and refusing to submit work one does not understand.

Indeed, AI may make some virtues more important rather than less important. The AI-assisted researcher must be more vigilant about error, more explicit about standards of verification, and more disciplined in distinguishing fluency from truth. The vice is not using AI. The vice is outsourcing judgment while pretending to possess it.

Virtue epistemology therefore rejects both extremes. It rejects the authorship thesis because human prose production is not intrinsically virtuous. But it also rejects an indiscriminate permissiveness in which the researcher becomes merely a name attached to machine output. The virtuous author may use AI fully, but not passively or irresponsibly.

9. Social epistemology: science as organized criticism

Science is not merely an individual achievement. It is a social institution organized around criticism, replication, peer review, methodological transparency, and reputational accountability. The epistemic standing of a paper depends partly on this social machinery.

From this perspective, the fixation on individual human composition looks misplaced. Published research is already the product of distributed cognition: coauthors, seminar participants, referees, editors, research assistants, statistical software, databases, laboratories, and disciplinary conventions. AI extends this distributed structure; it does not create it from nothing.

The important institutional question is therefore how AI affects the social organization of criticism. Does it make errors harder to detect, or easier? Does it enable more replication, or more noise? Does it increase access to scholarly production, or flood journals with low-quality submissions? These are serious questions. But they concern institutional design, not the intrinsic sanctity of human authorship.

10. The extended mind and the continuity of cognitive tools

Human cognition has always been technologically scaffolded. Writing, diagrams, libraries, microscopes, telescopes, calculators, statistical packages, search engines, reference managers, and collaborative documents all extend what researchers can think and do.

AI differs in degree, and perhaps in important qualitative ways, from older tools. But the mere fact that it contributes to cognition does not make it illegitimate. The boundary between “the human researcher” and “the external aid” has never been philosophically clean. Academic work is already hybrid.

The authorship thesis often assumes an ideal of pure human cognition that does not describe actual science. Once that ideal is abandoned, the question becomes comparative: which configurations of humans and tools produce better inquiry under defensible norms of responsibility?

11. Conventionalism about authorship norms

Authorship is also a convention. Different disciplines already allocate credit differently. In some fields, author order tracks contribution; in others, it follows alphabetical order; in still others, senior investigators or laboratory heads appear by convention. Some contributors receive authorship, others acknowledgments, and others no visible credit at all. Ghostwriting, editing, translation, statistical consulting, and research assistance have long complicated the idea that the named author literally produced the whole text.

This does not mean authorship norms are arbitrary. Conventions can be better or worse. They can promote accountability, fairness, and efficient coordination. But their justification is functional. We should ask what authorship rules are for, and then design them accordingly.

If authorship norms exist to assign responsibility, organize credit, and facilitate trust, then AI use should be governed by those purposes. The convention should not be reified into a metaphysical doctrine according to which human prose production has intrinsic value.

12. Liberal neutrality and pluralism

A final route comes from liberal pluralism. In a modern research community, scholars disagree about free will, personhood, virtue, authenticity, and the nature of mind. It would be inappropriate for academic institutions to build publication rules on one controversial metaphysical view of human authorship.

The better institutional standard should be one that researchers with different philosophical commitments can accept: published claims must be attributable to accountable human agents who understand, endorse, and can defend them. That standard does not require agreement about determinism, the metaphysics of agency, or the intrinsic value of human creativity. It requires only a shared commitment to responsibility and epistemic quality.

This gives defenders of full AI use an important argumentative advantage. They need not persuade everyone that determinism is true. They need only show that the anti-AI authorship thesis depends on a much more controversial premise than its defenders often acknowledge.

The concluding implication: accountable AI authorship

One can thus reject the Human Authorship Thesis on many philosophical grounds, while it is unclear to me how one can defend it except by invoking mystical and suspect metaphysical presuppositions. But if the Thesis is rejected, what should we aim for in its stead? How about this norm:

Researchers may use AI at any stage of academic work, including idea generation, literature review, coding, data analysis, outlining, drafting, writing, translation, response to referees, and stylistic refinement. But the human authors must be able to understand the work, verify its central claims, take responsibility for its methods, disclose AI use when relevant rules require it, and defend the final text under critical questioning.¹

This standard says that what matters is whether the final article is something for which the authors can be held epistemically accountable.

Under this norm, researchers who submit AI-generated text that they do not understand have violated a basic norm of authorship. The violation is that the researchers cannot responsibly stand behind the work (but the fact that it is AI-generated is not a problem as such). Conversely, researchers who use AI extensively but check the evidence, understand the reasoning, verify the citations, and endorse the final argument have not abandoned authorship. They have exercised authorship through direction, judgment, and accountability.²

An interesting thing is that the various philosophical perspectives come to a similar conclusion. Determinists reject ultimate self-authorship. Consequentialists care about epistemic results. Reliabilists care about truth-conducive processes. Philosophers of science distinguish discovery from justification. Realists distinguish truth from provenance. Pragmatists evaluate tools by their contribution to inquiry. Virtue epistemologists focus on intellectual character. Social epistemologists emphasize organized criticism. Extended-mind theorists reject a sharp boundary between mind and tool. Conventionalists treat authorship norms as revisable institutional devices. Liberal pluralists resist imposing a controversial metaphysics of human creativity on the whole academy. These views disagree about much else. But they converge on the view that human prose production is not the intrinsic locus of scientific value. What matters is truth, justification, reliability, criticism, understanding, and responsibility, irrespective of authorship.

A better norm for the AI age than “the human must have written the text” is that the human author must be able to answer for it. And that can be done with completely AI-written text.³

1. I favor relatively minimal AI disclosure rules. This is not because I regard deception about research practices as acceptable, but because disclosure requirements should be tied to epistemically relevant information. When AI use affects the production of data, coding, classification, statistical analysis, images, literature screening, translation of quoted material, or other elements that bear on verification and interpretation, disclosure can serve a genuine scientific purpose. By contrast, highly granular requirements concerning ordinary drafting or revision assistance, such as specifying every occasion, prompt, model version, or exact time of use, often seem to add little to the evaluation of the article. They may also invite irrelevant status judgments from readers who regard extensive AI use as intrinsically suspect because they accept some version of the Human Authorship Thesis. A better norm would require authors to state, where relevant, the kind of AI assistance used and to affirm that the listed human authors have checked, understood, endorsed, and taken responsibility for the final work. This preserves accountability without treating AI-assisted formulation as a scientific defect in itself. Disclosure: Most of this text was written interactively with ChatGPT. ↩︎
2. What about the objection that using AI for writing constitutes a form of plagiarism? By placing one’s name on a text, one allegedly represents oneself as the person who wrote it. This objection has force only if academic authorship is understood as a claim of sentence-level origination. I reject that understanding, both because of the deterministic considerations discussed in the text and because several other philosophical routes also detach authorship from unaided formulation. In scientific writing, authorship should instead be understood as responsible endorsement: the listed author understands the argument, has verified the central claims, accepts responsibility for the evidence and methods, and can defend the final text. Authorship so understood is compatible with extensive assistance in formulation, whether from coauthors, editors, translators, research assistants, software, or AI. This does not license plagiarism or deception. If AI is used to conceal uncited dependence on another scholar’s words, ideas, data, or distinctive formulation, the result is plagiarism; if the named author cannot understand or defend the submitted work, authorship itself is defective; and if AI use that is epistemically relevant to the evaluation of the work is deliberately concealed, there may be an integrity problem. But it does not follow that every failure to comply with highly granular or poorly justified AI disclosure rules is plagiarism or morally equivalent to plagiarism. Such failures may at most constitute breaches of local submission requirements, and whether they amount to genuine misconduct depends on whether the rule serves a defensible epistemic or institutional purpose. The central objections concern misappropriation, false representation, and lack of answerability, not the mere fact that the prose was formulated with AI assistance. ↩︎
3. This also implies that AI detection tools become wholly uninteresting in a research setting, even if such tools are 100% correct in their categorization of text. For more on proposals to use such tools to auto-reject academic work, see my earlier post. ↩︎

There are growing calls for journals to prohibit “extensive” AI use in manuscripts and referee reports, sometimes coupled with proposals to run AI-detection software and automatically reject texts flagged as AI-written. I worry this approach will lower scientific quality and productivity, and that it will do so for reasons that are only weakly connected to what journals are supposed to optimize. I propose a simpler principle: journals should evaluate manuscripts by scientific merit and enforce accountability for claims, rather than policing the provenance of prose.

Definition: what counts as “extensive” AI use?

By “extensive AI-tool usage”, I mean cases where most of the text of a manuscript or referee report has been drafted by a generative AI tool, as opposed to narrow uses such as copyediting, proofreading, or language polishing. The latter are functionally similar to human editing and rarely treated as ethically problematic.

Why do some people want categorical rejection?

The impulse behind categorical rejection appears to be an authorship-and-credit view: scholarly credit should attach only when the author personally produces the work, in a strong, almost artisanal sense. If an AI system writes 95 percent of the words, the argument goes, the author did not truly “do the work”, so the output should not count.

That position is not obviously correct, and it is not obviously aligned with the goals of science.

A simple thought experiment

Consider a simple case. A scientist has a novel idea with clear scientific value. Instead of spending weeks turning it into polished prose, the scientist iterates with an AI tool to produce a coherent paper. The scientist supplies the core idea, pushes the tool through multiple rounds of critique and revision, checks the logic, and ultimately submits an excellent manuscript. The journal runs an AI detector, it claims “95% AI-written”, and the paper is automatically rejected. The same happens elsewhere. The work is effectively blocked from the scientific record unless the scientist spends substantial time rewriting it purely to satisfy a production norm about who typed the sentences.

If the mission of a journal is to select and disseminate high-quality contributions, this is difficult to defend. It treats “who wrote the prose” as decisive even when the science is strong. That is a mismatch between means and ends.

Tool use is already central to modern research

A common reply is that writing is part of scholarship, so authors should do it themselves. I regard this as a category mistake. Many parts of the research process are already delegated to tools, including highly consequential parts. Empirical economists rely on statistical packages to estimate models, compute standard errors, and implement algorithms that almost nobody could reproduce by hand in practice. We do not reject papers because the author did not personally execute matrix algebra on paper. We instead require that the methods are appropriate, the results are correct, the code is available when needed, and the claims survive scrutiny.

If tool use is acceptable for inference, computation, and simulation, it is hard to see why tool use for exposition should trigger categorical rejection. The relevant question is whether the manuscript is correct, intelligible, and contributes to knowledge, not whether the prose was typed by a human.

Comparative advantage and scientific productivity

There is also an efficiency argument that matters for scientific progress. Researchers differ in strengths. Some have an unusual ability to generate fruitful ideas, identify sharp questions, and design credible empirical or theoretical strategies, while being comparatively weaker writers. If journals enforce a norm that valuable ideas must be expressed in wholly human-written prose to be admissible, they impose a tax on idea production and redirect effort toward mechanical tasks. The predictable result is lower output and slower progress, with little compensating gain in scientific reliability.

The epistemic objection: “Then we do not know who had the idea”

A related objection is epistemic and incentive-based: if extensive AI assistance makes it harder to infer who contributed the underlying ideas, then publication may become a noisier signal for credit allocation, and credit allocation matters because it shapes incentives for effort, originality, and integrity. This is a serious concern.

The institutional question is therefore which rule best aligns credit with contribution while preserving a high-quality scientific record. A blanket prohibition on AI-drafted prose is a blunt instrument because prose provenance is, at best, an imperfect proxy for intellectual contribution. It would encourage superficial compliance behavior (rewriting to satisfy a production norm) rather than the behavior journals ultimately need (truthful claims, accurate citations, and defensible inference).

A more coherent response is to tighten responsibility and signaling where it actually matters. Journals can require functional disclosure of AI use (especially where it could affect factual content), require explicit author contribution statements (for example, who formulated the core question, designed the strategy, conducted the analysis, and took responsibility for the argument), and enforce sanctions for misrepresentation. Under such a regime, even if AI contributed to ideation at the margin, credit remains tied to disclosed, accountable contributions rather than to who typed the sentences. AI-assisted ideation raises genuine questions about credit and incentives that deserve further thought, but those questions are best addressed through disclosure, contribution statements, and enforcement against misrepresentation, rather than through blanket bans or detector-based auto-rejection.

Why detection-and-auto-reject is a poor mechanism

Even if one endorses stricter norms around disclosure, AI-detection tools are a weak foundation for high-stakes editorial decisions. It seems to me they are opaque, error-prone, and unstable over time, particularly as writing models evolve and as humans adapt style in response. Rejecting manuscripts or referee reports on the basis of a probabilistic classifier raises fairness concerns because authors can be punished without transparent, contestable evidence.

Moreover, a detection-and-evasion arms race diverts resources away from research and toward compliance theater. It creates incentives to spend time “beating the detector” rather than improving arguments, verifying citations, or clarifying contributions. That is a poor use of high-skill labor, and it targets the wrong object. Journals should care about validity, novelty, and contribution, not stylometric guesses about authorship.

Suppose, for the sake of argument, that detection were perfect. Even then, an auto-rejection rule remains misguided. Perfect detection would establish only that an AI system produced much of the prose, which is comparatively superficial if the underlying contribution is sound. It would institutionalize a norm that elevates the mechanics of sentence production over the epistemic content of the work.

A more coherent policy stance

Journals should evaluate manuscripts by scientific merit and enforce accountability for claims. Extensive AI assistance should not, in itself, be a reason for rejection. If AI-assisted drafting produces weak argumentation, mischaracterizes prior work, fabricates citations, or introduces factual errors, then the manuscript should be rejected for the usual reason: it fails to meet scholarly standards.

However, manuscript policy and peer-review policy should probably not be identical. The right design depends on what is being protected.

1. Manuscripts: quality, transparency, and responsibility

For manuscripts, journals can protect the scientific record without detector-based gatekeeping.

First, I recommend maintaining and strengthening substantive standards. Require clear identification of the contribution, coherent argumentation, appropriate engagement with the literature, and claims that are testable and not overstated. Tool-assisted prose does not relax any of these requirements.

Second, I favor enforcing author responsibility. Authors should remain accountable for all claims, citations, data provenance, and ethical compliance, regardless of what tools were used to draft or edit the text. If an AI tool generated an argument, the author is still responsible for defending it. If an AI tool proposed a reference, the author is responsible for verifying it.

Third, I think disclosure could be used in a limited, functional way. If journals require disclosure of AI use, the purpose should be accountability and reproducibility rather than stigma. A sensible approach is to distinguish between editing and drafting, to note whether AI was used for translation or code assistance, and to clarify whether any tool could plausibly have affected factual content (for example, references, numerical claims, or descriptions of prior work). This is analogous to disclosing statistical software and computational tools: the point is transparency, not moral judgment.

Fourth, journals should focus on verifiability, not style. Where risks are salient, journals can require concrete checks: reference verification, access to data and code when relevant, and a review process that explicitly targets validity rather than rhetorical aspects. If AI use increases the frequency of certain failures, then journals should address those failures directly.

2. Referee reports: confidentiality and data handling

Referee reports raise an additional constraint: confidentiality. Journals are justified in imposing stricter rules here, because reviewers typically have access to unpublished ideas, data, and arguments. A journal can reasonably prohibit reviewers from uploading manuscript text, detailed descriptions of unpublished results, or other confidential materials into external AI systems. This is not about punishing AI use as such; it is about preventing disclosure.

Within that constraint, the relevant enforcement target should be confidentiality compliance and the substantive quality of the report, not stylometric detection. A journal could permit narrow uses that do not implicate confidentiality, such as local grammar correction, formatting assistance, or rewriting of a reviewer’s own notes, provided the reviewer remains intellectually accountable for the content. If an AI tool can credibly guarantee non-training and, crucially, stringent confidentiality safeguards (minimal retention, strong access controls, and auditable compliance), then the privacy concern is substantially mitigated, though not necessarily eliminated.

Conclusion

Science should be organized around producing and selecting true, important, and useful knowledge. A policy of rejecting work because an AI system wrote much of the prose substitutes an authorship norm for epistemic evaluation, encourages wasteful compliance behavior, and risks suppressing valuable contributions. Journals should not implement auto-rejection rules based on AI-detection. They should evaluate what matters: the quality of the science, transparency through functional disclosure where appropriate, and clear human accountability for the claims being advanced.

För några veckor sedan kom jag att tänka på en nära vän från förr. Jag lyssnade på Mireille Mathieu på YouTube och drömde mig tillbaka till mina tidiga tonår, då jag började läsa franska. Jag gjorde det tillsammans med min klasskamrat Pär, som var en riktig frankofil. Han älskade det franska språket! En dag frågade han mig om vad jag tyckte om ”Mireille”. Jag kände inte till henne, men han gav mig ett kasettband med sånger av henne, och jag fastnade direkt. Quelle voix!

Pär och jag förlorade kontakten efter gymnasiet, och när jag sökte efter honom på nätet för att se vad han sysslade med nuförtiden fann jag chockerande nog en minnessida. Pär dog 2021, 52 år gammal.

Jag har inte kunnat släppa denna upptäckt. Den har fått mig att minnas så mycket fint, men också en del fult, liksom den har fått mig att reflektera över livet (och döden).

Det fina. Pär och jag tågluffade i Frankrike sommaren 1985, efter 1:an i gymnasiet. Vilket äventyr och vilken livsglädje! Vi övade vår franska och njöt av stadslivet i Paris och Nice. Pär hjälpte mig se det rika i den franska kulturen, och även om jag aldrig riktigt har tagit den till mig som han, är jag glad över att någon förmedlade en passion till mig. Före resan hade en fin vänskap utvecklats under många år.

Det fula. Pär och jag hade något annat än franskan gemensamt: vi upptäckte att vi inte blev attraherade av flickor utan av pojkar. Detta talade vi dock inte om i vår ungdom, troligen beroende på att jag då var pingstvän. Jag kämpade emot och accepterade inte dessa känslor i mig själv. Men vi visste, vi förstod, vi kände att vi förenades av denna sak (i en värld där det inte ansågs ”normalt”). Det fula är att jag, när Pär i slutet av 80-talet berättade att han hade träffat en kille, var kall och fördömande mot honom. Detta har jag haft dåligt samvete för sedan dess. Det är en skamlig sak. Jag bad förvisso om förlåtelse i ett brev 1993, skickat från USA där jag tillbringade ett år som gästdoktorand, med en fras från operan Norma. (Är man homosexuell så är man.) Det är Adalgisa som efter ett avbrott i vänskapen med Norma sjunger:

Sì. Hai vinto. Abbracciami. / Ja. Du har segrat. Omfamna mig.
Trovo un’amica amor. / Jag finner en vän i kärlek.

Men vår vänskap återfann inte sin styrka. Jag hade saboterat den pga. mitt val att sätta det jag då uppfattade som kristen moral över kärlek och vänskap. Ett förfärligt misstag.

Reflektioner. Pär hade passion. Han älskade det franska, Mireille Mathieu, att skriva och att läsa. Men jag kan inte undgå att undra om han kom till sin rätt i livet. Vad jag vet fick han inte leva i enlighet med sin passion. Han mötte inte bara mitt fördömande utan även fördömande från andra i sin närhet. Han arbetade inte med det han brann för. Bröts han ned av den hårda verkligheten?

Jag minns Pär med kärlek och värme, med sorg och saknad, med ånger och skam, och jag har tårar i ögonen när jag skriver detta. Pär, je ne t’oublie pas!

Tillsammans med Therese Nilsson och Martin Ljunge har jag under flera år bedrivit forskning om tolerans. Med det menas i vårt fall en vidsynthet gentemot människor som är annorlunda. I synnerhet har vi försökt utröna vad som bestämmer graden av tolerans i ett samhälle, med fokus på institutioner. Några resultat:

• Rättssystemets kvalitet och penningvärdets stabilitet är positivt relaterade till tolerans mot homosexuella.
• Graden av marknadsekonomi inverkar positivt på tolerans, och ju mer människor litar på varandra i samhället, desto starkare är denna inverkan.
• Ju mer social och ekonomisk globalisering, desto större andel som anser det viktigt att lära barn tolerans och respekt.
• Lägre marginalskatter för dem med högst inkomst är positivt relaterade till mer tolerans mot ateister, kommunister och homosexuella (i USA).
• Invandrare med bakgrund i länder med låg andel muslimer, opartiska institutioner och en hög andel som anser det viktigt att lära barn tolerans och respekt är mer toleranta mot homosexuella än andra invandrare.
• Rättsstatens kvalitet är positivt, och graden av ekonomisk öppenhet mot andra länder är negativt, relaterad till antisemitism.

Man kan fråga sig varför vi anser det så viktigt att studera vad som kan tänkas bestämma tolerans i ett samhälle. Ett skäl är förstås att tolerans kan förväntas bidra till bättre livsvillkor och högre välmående för minoriteter av olika slag. Ett annat skäl ges i en ny studie, ”The Power of Tolerance vs. Unselfishness as a Cultural Determinant of Cooperation”, av Kimmo Eriksson, Brent Simpson och Irina Vartanova. De gör och finner, i kort sammanfattning, följande:

What makes some societies more cooperative than others? This question is important for a number of reasons, including the possibility that societal differences in cooperation may have played an important role in human cultural evolution (Henrich, 2004; Richerson et al., 2016). … While prior work leads straightforwardly to the hypothesis that higher societal valuations of unselfishness will lead to more cooperation, we explain why tolerance can be expected to be at least as important. We then report findings showing that a societal emphasis on tolerance plays a stronger role than an emphasis on unselfishness in predicting international differences in cooperation, cooperative norms, and cooperative metanorms.

Sambandet mellan en positiv värdering av tolerans i ett land och författarnas samarbetsmått illustreras i denna figur (där man kan notera att Sverige ligger högst i värdering av tolerans):

Det är alltså så att samhällen som kännetecknas av tolerans är mer samarbetsinriktade, vilket i sin tur är en central faktor för att samhällen ska fungera väl på en lång rad områden och lösa olika typer av kollektiva problem. Tolerans är till och med mer viktigt än osjälviskhet för att möjliggöra och upprätthålla samarbete människor emellan. Det är onekligen en viktig anledning att studera toleransens bestämningsfaktorer.

Robert Motherwell: Royal Dirge (1972), 182 x 137 cm:

Jag skrev för ett tag sedan om liberalismens problem med barn, där min poäng var att en politisk ideologi som sätter samtycke i främst rummet kan sägas hamna i en problematisk situation när man beaktar skaffandet av barn, eftersom barnets samtyckte till att bli till saknas.

Det slog mig, då jag lyssnade på filosofen Agnes Callard i samspråk med Russ Roberts, att problemet har ytterligare en dimension. Callard berättar om hur hennes äldste son reagerade på beskedet att han skulle få ett syskon – med starkt ogillande. Han hade uppfattat det som att han och föräldrarna hade ett implicit kontrakt där han var deras älskade barn och att inget annat barn skulle tillkomma som kunde konkurrera med dem om deras kärlek. Han blev upprörd över detta kontraktsbrott. Problemet ur liberal synvinkel här är att när väl ett eller flera barn finns, och om fler barn skaffas, görs det utan existerande barns samtycke. Detta trots att ankomsten av ett syskon är en revolutionerande och livsförändrande händelse av stora mått.

Någon kanske menar att liberalismen enbart är intresserad av samtycke från ”vuxna” individer och att man kan fatta beslut med konsekvenser för barn utan att bry sig om huruvida de samtycker eller ej. Så tycks förvisso rent faktiskt vara fallet – men att konstatera det är question-begging. Frågan är om detta är en rimlig ordning eller om det bör anses problematiskt att skaffa barn när barnets eget samtycke till att bli till eller till att få syskon saknas eller (i alla fall) saknar reellt värde.

Antag, i det fall vi diskuterar här, med ett nytt syskon, att föregående barn är 18 år och ”vuxen”. Bör inte då samtycke inför att få ett (ytterligare) syskon inhämtas? Eller är det så att när ett barn blir vuxet är samtycket irrelevant eftersom barnet, om det så önskar, kan sätta eget bo och distansera sig från sin kärnfamilj, om dess storlek och sammansättning inte befinns tilltalande? Eller att samtycket helt och hållet är irrelevant och föräldrarna anses ha oinskränkt rätt att skaffa barn (eller i alla fall en rätt som i sin omfattning inte påverkas av tidigare barns uppfattningar)?

Ett möjligt argument för den sista synen skulle kunna vara ett ”förväntan-argument”: om det kan förväntas vara så att ett existerande barn längre fram i livet accepterar ett nytt eller flera nya syskon är det i sin ordning att skaffa barn även om det existerande barnet motsätter sig det just nu. Detta argument kan ses som problematiskt eftersom det utgår från att föräldrarna vill och kan bedöma ett barns framtida samtycke, vilket förefaller tveksamt i många fall.

Jag inser att detta inlägg, liksom mitt förra i samma anda, är en intellektuell ”lek” i någon mening, och att nästan alla finner problematiseringen absurd. Men finns det ändå inte något som skaver här i det liberala sinnet?

Jag skrev för ett tag sedan följande:

För egen del är jag starkt påverkad av Schopenhauers förståelse av livets karaktär, men jag har svårt att se det som möjligt att uppnå ”frälsning” i form av viljelöshet.

Dvs. livet består av en cykel av lidande och tristess, och det är bara genom att sluta vilja som man kan komma ur den. En nationalekonom skulle kanske uttrycka det som att lösningen ligger i att sluta ha preferenser, att sluta ha mål att sträva mot, att sluta låta sig dras med i det tankemönster som säger att lycka kan uppnås genom preferenstillfredsställelse.

Men, som sagt, har jag inte upptäckt hur man kan sluta vilja, hur man kan sluta ha preferenser, hur man kan sluta ha mål att sträva mot, etc. Jag är fast i Schopenhauers cykel. Intressant nog läste jag just en mycket tankeväckande essä, ”The Wisdom of Surrender”, som tyder på att Samuel Beckett tänkte i liknande banor:

In his later years, Beckett occasionally seemed keen to distance himself from soteriological systems. In 1986, just three years before his death, he told an interviewer that, unlike Schopenhauer, unlike Leopardi and unlike certain ”oriental thinkers”, he didn’t propose a ”way out” or a ”hope of an answer” or a ”solution”. The only solution, Beckett said, grimly, was death.

Hopplöshetens filosofi?

Traditionellt har äktenskap, som rättslig företeelse, varit möjligt att ingå för personer av olika kön. På senare år har även personer av samma kön kunnat göra det. I Nederländerna, som var först ut, infördes denna möjlighet 2001, och i Sverige 2009. En reform av detta slag kan förstås motiveras på olika sätt. En liberal, principiell utgångspunkt om likabehandling är en utgångspunkt; en annan att en möjlighet att gifta sig har olika goda konsekvenser. En ny studie, ”Mental Health Effects of Same-Sex Marriage Legislation”, undersöker just konsekvenserna i Nederländerna på mental hälsa. Forskarna finner följande:

We study how this legislation affected mental health of sexual minorities in the Netherlands. Conducting a difference-in-differences analysis, we compare changes in mental health following the legalization between sexual minorities and heterosexuals. We find that SSML [same-sex marriage legislation] improved mental health of both married and nonmarried sexual minorities, which implies that marriage is not the only channel. Examinations of alternative mechanisms combined with literature suggest that the legislation may also take effect by improving societal tolerance as well as stabilizing partnerships and enriching the choice basket of partnership forms for sexual minorities.

Detta synes mig vara en viktig effekt av lagstiftningen: förbättrad mental hälsa hos en minoritet som har diskriminerats såväl på detta område som på andra genom historien och som erfar sämre mental hälsa än heterosexuella. En intressant del av resultaten är alltså att inte bara de homosexuella som gifter sig mår bättre, utan även andra homosexuella. Ibland förlöjligas reformer för minoriteter som ”symbollagstiftning”, men det tycks vara så att mer likabehandlande lagar, oavsett om de kommer till användning eller ej, medför goda effekter på tidigare diskriminerade grupper. De sänder en signal om inklusivitet.

Låt mig nämna två relaterade studier, som visar att likabehandling av homosexuella är relaterad till större livskvalitet hos homosexuella män och till större livstillfredsställelse hos befolkningen i stort. Ytterligare skäl att välkomna denna typ av lagar.

Många är bekymrade över det stöd som populistiska partier erhåller runtom i Europa. Det leder in på frågan vad som kan förklara detta stöd. En faktor som hittills inte har uppmärksammats är individens subjektiva välmående. En ny studie, ”Do Unhappy Citizens Vote for Populism?”, undersöker därför kopplingen mellan olika indikatorer på sådant välmående och sannolikheten för att rösta på ett populistiskt parti i Europa under de senaste två decennierna.

Här är en figur som antyder att det kan finnas ett samband:

Tesen undersöks ekonometriskt i ovan nämna studie, och forskaren finner följande: 

The results suggest that generalized unhappiness with one’s personal well-being – and not merely dissatisfaction with governments – could play a significant role in the rise of European populism. As such, low levels of subjective well-being are proposed as a valid predictor of shifts towards extreme movements.

Ett sätt att minska sannolikheten för att populistiska partier får stöd tycks alltså vara att försöka säkerställa att människor är lyckliga. Kanske genom bättre ekonomiskt, politiska och legala institutioner?

Se ett tidigare, relaterat inlägg: ”Har populismens framfart med globaliseringen att skaffa?”.

Fenomenet klaner har diskuterats mycket i svenska media på sistone, särskilt eftersom de i vissa fall kopplas till kriminell verksamhet bland invandrare. Det var därför intressant att stöta på ett blogginlägg av Jane Shaw Stroup, som utifrån två böcker presenterar tesen att en viktig förklaring till att västvärlden har utvecklats på det sätt den har, med rättsstat, demokrati och marknadsekonomi som kännetecken, har grund i beslut, inte minst av påven Gregorius I och andra inom Katolska kyrkan, som effektivt underminerade klaner som organisationsform i samhället.

Den ena boken är Deepak Lals Unintended Consequences: The Impact of Factor Endowments, Culture and Politics on Long-Run Economic Performance, och den andra är Joseph Henrichs The WEIRDEST People in the World: How the West Became Psychologically Peculiar and Particularly Prosperous. Jag har inte läst någon av dessa böcker, så jag förlitar mig på Stroups återgivande. Så här beskriver hon Henrichs resonemang:

Henrich gathers up hundreds of studies, including many personality tests, to argue that the destruction of traditional kinship led people to develop a wide range of unusual characteristics. They became “increasingly individualistic, independent, self-focused, nonconformist, and relationally mobile.” And that’s just for starters. [5]

Without the support of clans, Europeans in the Middle Ages created voluntary associations such as guilds, monasteries, and universities and many moved from agriculture to towns and cities. Because they learned to deal with people outside their kin, they were able to develop markets, vastly increasing their access to new products—and to new ideas. No longer did they have to deal only with family members because only family members could be trusted. WEIRD [Western, Educated, Industrialized, Rich, Democratic] people learned that outsiders could be trusted—and, to make sure of it,  they developed institutions like the “law merchant” and contracts. They developed other useful habits like attention to punctuality. … So, such characteristics led to markets, economic growth, and innovation. 

Jag ser detta som en tes, inte som ett etablerat faktum, men det är onekligen intressant att tänka sig att den institutionella ordning som kännetecknar västvärlden, och det välstånd den i hög grad har bidragit till, har kulturella rötter – kanske inte minst i hur Katolska kyrkan har sett på och utfärdat regler om äktenskap och familjebildning som har underminerat klaner. Deepak Lals tes, i den första boken, är att denna syn och dessa regler infördes av rent egoistiska skäl, för att kyrkan skulle kunna ta del av människors arv. I så fall är västvärldens utveckling en bieffekt av en egoistisk strävan av en mäktig organisation att bli ännu rikare!