Everyone Uses AI in Research. Nobody Writes It Down.

If today's papers were re-read against the old standards, many acknowledgments pages would have to be rewritten.

The old standards were generous. Early in a project, someone explains the field's open problems and results. You thank them. When the work stalls and a friend takes a walk with you, listens to your confusion, and says one sentence that opens things up. You thank them. At a conference, you corner an expert and ask a few key questions. You thank them. You read an important paper, email the author about a derivation, and get a reply explaining the details. You thank them too. Go further. Someone verifies a derivation, checks code, helps run data, or offers important input on the writing plan or the research plan. By the old ruler, that kind of help may belong in the author list.

That standard was never applied evenly. Authorship and acknowledgments are shaped by power relations, familiarity, the visibility of a contribution, and the evaluation system. But it represented an ideal: help should be seen, and the chain of contribution should be traceable.

AI has magnified the problem.

To survey an unfamiliar field, the routine move is now to hand Kimi or Qwen the task and get a preliminary report. To understand an unfamiliar paper, you feed the material to a model. A derivation that once took a week to grasp takes a few hours to understand. When intuition says a pattern might be universal, you ask AI to search for counterexamples, explanations, another route. Sometimes it finds a path and gives a clue to why the pattern holds. It does more than polish language. It has entered literature review, conceptual understanding, derivation checking, code assistance, and idea generation. Applied strictly, the old standards would require a large share of papers to thank AI. By the old measure of "important input on the research plan," some work would even have to list AI as an author.

The dispute is not about technology. It is about norms. The research norms of the AI era have not been built.

1. Acknowledgments Acknowledge Help; Authorship Carries Responsibility

Authorship includes several things: substantive contribution to the research; consent to publication; responsibility for the content; being reachable and able to respond to criticism. AI can generate text, propose ideas, check code. But it cannot consent, cannot be held accountable, cannot answer for error, and cannot stand up when the paper is challenged and say, "This is where I judged wrong." So AI cannot be an author. Accountability has to land on someone who can answer for the work.

AI cannot sign, but that does not excuse silence. "No AI was used" is not an adequate disclosure. Many journals have established one rule: AI cannot be an author. They have not established the other: how AI contributions should be disclosed in specific terms. Everyone knows they are using it. Everyone knows everyone else is using it. In the paper, it is either omitted entirely or covered with one vague phrase: "language polishing." The rules offered no safe way to say it, so users chose silence.

2. The Machinery of Silence

Many people do not mention AI. They know it was useful. They fear that affirming AI's contribution will make others think "the work was done by AI." Research evaluation has long stressed originality, individual contribution, and ownership of ideas. AI use is easily equated with laziness, ghostwriting, padding. A loop forms: the less people disclose, the more those who do disclose look guilty; the more only the established and successful dare to speak, the more it reinforces "only big names can afford transparency."

Witten can talk about AI's specific contributions because no one thinks he is coasting on AI. A young researcher, a postdoc, someone in an up-or-out position may not dare carry the cost of being misread. In the corner of the whiteboard, formulas from the previous meeting were still there. He minimized the Kimi window and switched back to Excel; the spreadsheet held next week's deadline progress.

The problem is institutional. The system has not de-moralized disclosure. A reasonable AI statement should work like "used this software," "used this database," "used this statistical method": part of methodological transparency. It does not judge how much you used it; it says where it was used, how it was verified, who is responsible. Disclosure is part of reproducibility.

3. No Authorship, but Layered Disclosure

Keep AI out of the author list. Make its contributions clear in layers.

Language polishing, translation, formatting: state which tool was used.

Literature review, source summary, background mapping: state how it shaped the path by which you entered the problem.

Code assistance, data checks, derivation verification: in the methods section, state which steps were AI-assisted and how humans reviewed them.

Idea generation, alternative explanations, counterexample search: state what kind of suggestions AI offered and how the human authors verified, corrected, and took responsibility for them.

Writing assistance: if AI drafted passages or restructured arguments, disclose that honestly; do not fold it into "polishing."

These statements do not make AI a collaborator, let alone an author. They express: what AI took part in, how humans verified it, who carries responsibility. The granularity can be debated. "Say nothing at all" should no longer be the default.

In the submission system, the "AI-use disclosure" field is one line. Many people fill in "none." After that line, you scroll down, and there is the "author contributions" section.

4. Evaluation Culture

Beyond the rules, there is culture. If the evaluation system treats AI use as an original sin, even detailed rules will be avoided. Journals should provide a common template; institutions should state in research-integrity training that "disclosure does not mean a penalty"; reviewers should not assume lower quality because AI use is disclosed; mentors and senior researchers should lead by describing AI contributions specifically.

Statistical software, code libraries, public databases, and informal email discussions with peers have all changed the boundaries of authorship and acknowledgment. AI enters understanding, generation, verification, and writing at the same time; it cuts through the whole pipeline. The old norms look inadequate because they were never as tight as we imagined. Rather than saying 99% of researchers should thank AI, it is better to say that AI has made a long-standing problem impossible to avoid: contribution disclosure depends on goodwill and lacks fine-grained rules.

5. Three Lines on the Acknowledgments Page

AI cannot be an author, but it must be disclosed more specifically. Acknowledgments are part of the accountability chain. Disclosure states how the human authors verified and took responsibility.

The worst state is the one we have now: AI used by everyone, acknowledged by no one. The norms should give AI an accurate position: the tool can be strong, the collaboration can be deep, but responsibility stays on the human side.

Before the next submission, delete that phrase "language polishing" from the acknowledgments page and replace it with three lines: what tool was used, in which steps, and who reviewed which outputs. When the three lines are done, switch back to the document; the cursor sits in the author-contributions field. Outside the window, a parcel-locker light flickered on.