Where AI Suggests One Step: Responsibility Boundaries in Scientific Writing

A recent solo paper in theoretical physics sparked discussion. The author is the well-known mathematical physicist Edward Witten, and the paper runs to about thirty-six pages. Unlike many conventional papers, the author did not leave the role of artificial intelligence outside the writing process. In the main text, he states that a key step came from a “possible variant” suggested by an AI assistant. In the acknowledgments, he further notes that AI tools assisted with several details in the paper.

What matters is not whether AI “wrote a paper,” and not a simple comparison of who is smarter, the human scholar or the model. The more concrete question is this: when AI proposes only a local idea inside a proof, how should the researcher disclose it, verify it, and define responsibility?

1. From “generating content” to “offering a local suggestion”

In research, AI does not play only one role. It may help organize references, polish language, check notational consistency, draft programs, or propose candidate approaches to a technical difficulty.

These uses do not carry the same academic meaning.

The following kinds of work can usually be treated as assistance:

1. Language-level editing
   This includes rewriting sentences, checking grammar, compressing abstracts, and improving section structure. Such work mainly improves expression. It does not directly determine the paper’s core conclusions.

2. Formal and mechanical checks
   These include checking equation numbers, finding symbol conflicts, confirming that definitions stay consistent, and helping generate simple verification scripts.

3. Exploration of candidate ideas
   AI may propose a possible lemma, transformation, or proof path based on existing conditions. Whether that suggestion holds still depends on the researcher’s independent judgment.

4. A critical argumentative contribution
   If a system proposes a previously unknown core theorem, construction, or decisive proof, and the researcher cannot explain its origin or correctness, it is no longer an ordinary writing tool. It has become an important knowledge-generating step in the research process.

The most useful detail in the original case is that the author did not say, in general terms, “this paper was completed by AI.” He limited the contribution to a specific place: one variant at a seam in the argument, plus assistance with several details.

That description is more informative than “AI was used.” It tells readers what AI did and did not do, which content the author proposed, which content was only suggested by a tool, and who made the final mathematical judgment.

2. Why “one step at the seam” is still worth disclosing

In a complex mathematical proof, the hardest part is not always the longest calculation. It may be the connection between two theoretical modules.

A proof may already have:

• a mature main algorithm;
• a set of known boundary conditions;
• a usable classification framework;
• several intermediate propositions that have already been checked.

What often blocks progress is how to join these parts. A transformation that looks small may decide whether the whole proof closes. It may occupy only a few lines, yet play a critical role in the logical structure.

“Contributing only one step” therefore does not mean “that step does not matter.” Academic contribution should not be measured only by word count or lines of code. It depends on where the step sits in the chain of reasoning.

There is still a difference between a suggestion and a contribution. If AI offers a candidate variant, that does not mean it has proved the variant. The researcher still needs to do at least four things:

• judge whether the suggestion fits the problem’s definitions;
• check whether it introduces hidden assumptions;
• verify that the derivation covers all boundary cases;
• place it back into the full proof and confirm that it does not break the surrounding logic.

In other words, AI can say “this might be changed in the following way.” It does not automatically earn the status of “therefore the argument holds.”

3. Why a solo paper still leaves full responsibility with the author

Authorship usually means more than participation in writing. It also means responsibility for the research results. A named author should be able to answer:

• how the core conclusions were reached;
• what justifies each step of the proof;
• which definitions, assumptions, and external results were used;
• under what conditions the results hold;
• how an error should be located and corrected.

An AI system does not have a stable, continuous, and accountable status as an academic agent. It may produce reasoning that looks plausible but is wrong, and it may give inconsistent answers to the same question at different times. Even if it offers an important suggestion, the author cannot transfer responsibility to the tool.

That is why “AI suggested a possible variant” must be kept strictly distinct from “AI proved the theorem.” The first is a disclosure of source. The second easily misleads readers about who is responsible and how thoroughly the result was verified.

A more careful statement should include three layers:

1. Scope of involvement: which section, which step, and in what way the tool participated;
2. Human review: whether the author re-derived and verified the suggestion;
3. Final responsibility: the named author remains responsible for the paper’s conclusions, errors, and explanations.

This kind of disclosure neither inflates the role of AI nor conceals its real effect on the research process.

4. How research papers should record AI involvement

If AI takes part in technical research, a single acknowledgment such as “thanks to AI for assistance” is often not enough. A more reproducible practice is to keep a concise usage record.

The record can include:

1. Purpose of use

State whether the tool was used for language editing, organizing materials, checking code, or exploring proof ideas. Different uses correspond to different levels of risk.

2. Location of involvement

Point to a specific section, proposition, or algorithm module, rather than saying only that “AI was used throughout the paper.”

3. A summary of inputs and outputs

It is not always necessary to publish the full conversation. At a minimum, summarize what question was asked, what kind of suggestion the tool returned, and which parts were actually adopted.

4. Method of human verification

Explain how the author checked the result. Examples include redoing a symbolic derivation, constructing counterexamples, running a formal check, or confirming the conclusion through an independent proof.

5. Version and time information

AI systems are updated continuously. The same prompt may produce different answers under different versions. If a tool materially affects a research conclusion, recording the time of use and the system version helps later review.

A simplified record template for a research project might look like this:

Research module: proof of Proposition 3
Assistance used for: exploring an intermediate transformation
AI suggestion: rewrite the boundary conditions in piecewise form
Actually adopted: only the variable rearrangement
Human verification: re-derived three classes of boundary cases and checked them with an independent program
Final conclusion: completed and owned by the author

The value of such a record is not formal complexity. It leaves an interpretable trace of the research process.

5. An acknowledgment is not sufficient disclosure

Acknowledgments are an important entry point, but they do not replace a methodological statement.

If AI was used only for grammar polishing, placing that fact in the acknowledgments is usually enough. If it took part in algorithm design, proof structure, experimental design, or data processing, a brief note in the acknowledgments still leaves readers unable to judge the scope of its influence.

A more reasonable practice is layered disclosure:

• Light use: state language and formatting assistance in the acknowledgments;
• Local technical suggestions: identify the specific location in a footnote or methods note in the relevant section;
• Important research contributions: add a separate “AI assistance statement” describing how the tool was used, how humans reviewed it, and what the limits were;
• Sensitive data or unpublished material: also state measures for data security, confidentiality, and access control.

The purpose of disclosure is not to create a headline. It is to help readers understand the paper’s sources of knowledge and the boundaries of its credibility.

6. Why AI-generated mathematical content must be verified independently

Large models are good at producing coherent text. That is not the same as being a rigorous prover. Typical problems include:

• writing a necessary condition as if it were sufficient;
• assuming that a function is continuous, invertible, or bounded;
• ignoring zero values, singularities, and boundary cases;
• using inconsistent notation across passages;
• citing an intermediate result that does not exist;
• treating an empirical pattern as a universal theorem.

Researchers should not include output in a paper merely because it sounds professional.

For theoretical work, a three-layer check is advisable:

Layer 1: Local derivation checks

Inspect formula transformations line by line and confirm that each step has a clear justification.

Layer 2: Counterexamples and boundary checks

Actively look for extreme conditions, low-dimensional cases, and degenerate cases, and test whether the conclusion still holds.

Layer 3: Independent-method checks

Try another proof path, a computational experiment, or a formal system as a cross-check. Confidence rises when different methods agree.

Where conditions allow, key propositions can also be converted into a machine-checkable form. AI then proposes candidate structure, the human makes the mathematical judgment, and a verification tool performs mechanical checking. The three roles complement one another.

7. What this means for developers and research teams

From an engineering point of view, this case is similar to code generation in software development. An intelligent tool may generate a critical piece of code, but whether that code can enter production depends on testing, review, dependency analysis, and security verification—not on how fluent the generator is.

Research needs a similar process:

1. Treat AI output as a candidate patch, not a final commit;
2. Keep prompts, outputs, and records of changes;
3. Build independent tests or proofs for critical conclusions;
4. Do not write unexplained results directly into formal work;
5. Make clear inside the team who reviews the work and who signs off.

Enterprise R&D teams should also add access control and data governance. Unpublished algorithms, customer data, internal code, or patent-related material should not be entered casually into an external system. Even if a tool improves efficiency, confidentiality, compliance, and intellectual-property requirements come first.

8. How authorship rules may change

As AI moves from language polishing into modeling, proof, experimental design, and code implementation, the traditional binary of “author versus tool” may need to be redefined.

Future papers and technical reports may more often include:

• human authors and the research modules they own;
• the scope of AI-tool use;
• automated verification tools and data-processing workflows;
• how key results were independently reviewed;
• reproducible experimental environments and version records.

This does not mean AI will naturally become a paper author. Authorship should remain tied to responsibility, the ability to explain the work, and academic commitment. A system that cannot defend the work, explain the method, or answer for errors is generally not suitable as an author in the traditional sense.

What is more likely to change is the granularity of contribution statements. In the past, people focused more on who proposed the topic and who ran the experiments. Going forward, it will also be necessary to say who designed the algorithm, who verified the proof, which steps an automated system proposed, and which content was reconstructed by humans.

Conclusion: Transparency matters more than myth

Witten’s paper drew attention not because AI suddenly replaced theoretical physicists, but because it showed a more mature research posture: do not exaggerate what the tool can do, and do not hide that it participated. Place a local suggestion in its accurate position, while the human author carries the full burden of verification and responsibility.

For researchers, the main principles can be summarized in three points:

• AI can propose ideas, but it cannot replace proof;
• AI can improve efficiency, but it does not automatically earn authorship or responsibility;
• AI can take part in research, but that participation must be transparent, explainable, and reviewable.

When these boundaries are clearly recorded, AI is no longer only a mysterious “black-box assistant.” It becomes a manageable, verifiable, and accountable technical step in the research workflow.