The Two Extremes—and What Happens to Those Who Used AI Before the Rules Caught Up
Artificial intelligence has entered scientific research much faster than scientific institutions have been able to develop rules for it.
Researchers are already using AI to write and edit manuscripts, generate code, analyse data, search literature, summarize papers, formulate hypotheses, prepare figures, interpret results and respond to reviewers.
The policies governing these activities, however, are still evolving.
That creates an uncomfortable question:
What happens to researchers who have already used AI extensively, at a time when the rules were unclear or nonexistent?
There are two extreme possibilities.
At one extreme, AI becomes an accepted and largely invisible part of scientific research. At the other, the scientific community develops a retrospective AI "audit" culture in which previously published work is scrutinized for undisclosed AI use.
The future is likely to fall somewhere between these extremes—but understanding both is important.
Extreme 1: AI becomes as ordinary as spell-checking
The optimistic scenario is that AI eventually becomes completely normalized.
Thirty years ago, researchers worried about whether using a computer to perform calculations was somehow different from doing calculations manually. Today, nobody declares:
"Microsoft Excel was used to calculate the standard deviation."
Similarly, nobody normally declares that a word processor was used to correct spelling.
AI could eventually follow the same trajectory.
A researcher might routinely use an AI assistant to:
- correct grammar;
- translate text;
- suggest clearer wording;
- generate routine code;
- debug scripts;
- search literature;
- summarize papers;
- organize references;
- generate figures from verified data;
- automate repetitive analyses.
Under this model, journals would increasingly focus not on whether AI was used, but on whether the scientific work is reliable.
This direction is already visible.
Elsevier's current policy, for example, distinguishes basic spelling, grammar and punctuation assistance from more substantive AI assistance.
Springer Nature has gone further by explicitly introducing a risk-based framework: relatively low-risk language assistance is treated differently from AI that substantially shapes scholarly content, while AI that replaces authorship or generates unsupported scholarly content is prohibited.
Under this future, an AI declaration might eventually become almost routine:
"AI-assisted tools were used for language editing and code development. All outputs were reviewed and validated by the authors."
Nobody would regard the declaration itself as suspicious.
Indeed, Springer Nature explicitly says that an AI declaration should not itself be considered a cause for concern.
Extreme 2: AI becomes the new research-integrity fault line
The pessimistic scenario is very different.
Imagine that AI becomes sufficiently powerful that it is possible to reconstruct, with reasonable confidence, whether a paper was produced with substantial AI assistance.
Suppose a paper published in 2025 contains:
- AI-generated paragraphs;
- AI-generated code;
- AI-assisted statistical analysis;
- AI-generated interpretations;
- fabricated or incorrect references;
- undisclosed AI-generated figures.
Five years later, the journal discovers evidence that the authors never disclosed the AI assistance.
What happens?
The issue may no longer be:
"Did you use AI?"
It becomes:
"Did you accurately disclose what AI did?"
That is a much more serious question.
The declaration itself may become evidence
This is an important shift.
Historically, authors generally made declarations about:
- conflicts of interest;
- funding;
- data availability;
- ethics approval;
- competing interests;
- authorship.
AI declarations are now joining this ecosystem.
If an author writes:
"AI-assisted language editing was used for paraphrasing portions of the text."
that creates a representation about what AI did.
If subsequent evidence demonstrates that AI was actually used to:
- generate substantial portions of the Discussion;
- write the analysis code;
- interpret results;
- generate hypotheses;
- identify literature;
- construct figures;
then the problem may not simply be AI use.
It may be inaccurate disclosure.
That distinction could become extremely important.
ICMJE's current guidance already says that nondisclosure of AI use may require corrective action and may, in some circumstances, be construed as misconduct.
COPE likewise places responsibility on authors for AI-assisted content and expects transparent disclosure of AI use in writing, image production and data collection/analysis.
The middle ground is much more likely
The two extremes are unlikely to become reality in their pure forms.
A much more plausible future is a risk-based AI governance system.
Think of AI use as a spectrum.
Green: communication assistance
Examples:
- grammar;
- spelling;
- readability;
- translation;
- formatting;
- minor paraphrasing.
These uses are likely to become increasingly routine.
Amber: substantive assistance
Examples:
- drafting sections;
- substantial rewriting;
- literature synthesis;
- code generation;
- statistical assistance;
- reference discovery;
- generating outlines;
- restructuring arguments.
These will probably require disclosure and increasing documentation.
Red: replacement of scientific judgment
Examples:
- AI independently generating scientific interpretations;
- AI generating results;
- AI fabricating data;
- AI generating unsupported conclusions;
- AI creating research images that appear to represent experimental evidence;
- using AI to replace the researcher's intellectual contribution.
These uses are likely to remain prohibited or subject to extremely strict controls.
Springer Nature's current policy already resembles this model.
The really important transition: from disclosure to documentation
This may be the biggest change coming.
Initially, journals asked:
"Did you use AI?"
Then:
"How did you use AI?"
The next question may become:
"Can you demonstrate how AI was used?"
This is already beginning to appear in discussions of research integrity.
A recent framework published in Research Integrity and Peer Review argues that disclosure alone is not documentation and advocates recording what AI was used for, what inputs were provided, what outputs were incorporated and how those outputs were verified.
This makes considerable sense.
Consider two researchers.
Researcher A
"ChatGPT was used."
That tells us almost nothing.
Researcher B
"ChatGPT was used on 14 June to generate an initial Python implementation of the sequence-filtering algorithm. The generated code was independently tested against simulated datasets and subsequently rewritten by the authors. No AI-generated code was used without validation."
That is reproducible.
The second approach is much more likely to become the norm.
What happens to researchers who used AI before the rules existed?
This is probably the most difficult question.
There are millions of research outputs produced during the period when AI policies were:
- absent;
- vague;
- inconsistent between journals;
- changing rapidly;
- unknown to many researchers.
It would be unrealistic to treat all historical AI use as misconduct.
A researcher who used ChatGPT in 2023 to correct grammar cannot reasonably be treated in the same way as someone who used AI to fabricate experimental results.
Therefore, a sensible future system will probably distinguish between use, misuse and concealment.
Category 1: AI was used appropriately and transparently
For example:
AI was used for grammar and language editing.
This is unlikely to become a major problem.
In fact, such uses are already accepted by many publishers.
Category 2: AI was used substantively but the work remains scientifically sound
For example:
AI generated some code, but the researchers independently validated the code and the results.
Or:
AI substantially rewrote sections of the manuscript, but all scientific claims came from the authors and were independently checked.
This situation is more complicated.
The likely future response will be:
"Disclose it."
A journal might request an updated declaration rather than retracting the paper.
Category 3: AI was used substantively but described inaccurately
This is where the risk becomes greater.
Imagine a declaration saying:
"AI-assisted language editing was used for paraphrasing portions of the text."
But records later show that AI was used to:
- generate large sections of the manuscript;
- formulate interpretations;
- write the analysis code;
- suggest conclusions;
- generate references.
The problem is now not merely AI use.
It is potentially misleading disclosure.
The eventual response could include:
- correction of the AI declaration;
- editorial investigation;
- requests for documentation;
- examination of the underlying analysis;
- institutional review in serious cases.
The severity will probably depend on whether the undisclosed use affected the scientific substance.
Category 4: AI was used to fabricate or manipulate research
This is fundamentally different.
If AI was used to:
- invent experimental results;
- fabricate datasets;
- manufacture citations;
- generate false images;
- create nonexistent experiments;
- conceal manipulation of scientific evidence;
then the AI issue becomes secondary.
The primary issue is research misconduct.
No future AI policy is likely to make fabricated science acceptable simply because an AI system produced it.
Recent publishing guidance is already explicit on this point. Nature Computational Science, for example, describes fabrication of data, images or citations using generative AI as a serious breach of research integrity.
Will journals retrospectively investigate old papers?
Probably—but selectively.
It would be impossible to investigate every paper published before comprehensive AI policies existed.
Instead, investigations are more likely to be triggered by:
- credible complaints;
- obvious discrepancies;
- fabricated references;
- unexplained analytical errors;
- suspicious figures;
- inconsistencies in methodology;
- admissions by authors;
- evidence of substantial undisclosed AI generation;
- other research-integrity concerns.
In other words:
AI detection is unlikely to become a universal fishing expedition.
It is much more likely to become another component of targeted research-integrity investigations.
AI detectors are unlikely to solve the problem
There is another reason retrospective enforcement will be difficult.
AI detectors are not equivalent to forensic proof.
A detector saying:
"This paragraph is probably AI-generated"
does not establish:
"This paragraph was generated by ChatGPT on a particular date."
Recent research-integrity work has emphasized that retrospective detection is uncertain and can produce both false positives and false negatives. Prospective documentation is therefore potentially much more useful than attempting to reconstruct AI use after publication.
This is important for students and researchers.
A future investigation is much more likely to value:
- version histories;
- Git commits;
- laboratory records;
- prompts;
- AI-use logs;
- drafts;
- code histories;
- analysis notebooks;
than an AI detector's probability score.
The Git repository may become part of the research record
This is particularly relevant for computational biology.
Imagine that a paper says:
"The analysis was performed using Python."
In the future, reviewers may reasonably ask:
"Who wrote the code?"
The answer could be:
"The initial implementation was generated using an AI coding assistant."
That is not necessarily problematic.
What matters is whether:
- the code is available;
- the code was tested;
- the analysis is reproducible;
- the authors understand what it does;
- the results can be independently reproduced.
This could actually make computational research more transparent, not less.
AI may eventually become analogous to a laboratory instrument
There is an interesting possibility here.
We routinely report:
"Sequencing was performed using an Illumina NovaSeq."
We report:
"Images were acquired using a Leica SP8 confocal microscope."
We report:
"Phylogenetic analyses were performed using IQ-TREE."
In the future, we may routinely report:
"Code development was assisted by GitHub Copilot."
or:
"Manuscript language was edited using an AI-assisted language model."
The AI system becomes part of the methodological provenance of the research.
But there is a critical difference:
A microscope does not invent an interpretation.
An AI system potentially can.
That is why AI governance will probably remain more complicated than ordinary software disclosure.
What will journals increasingly ask authors?
I suspect future submission systems will move toward structured questions such as:
AI use during research
Was AI used in the research itself?
- No
- Yes — literature analysis
- Yes — research design
- Yes — coding
- Yes — data analysis
- Yes — interpretation
- Yes — image processing
- Yes — other
AI use during manuscript preparation
Was generative AI used in preparing the manuscript?
- No
- Language editing
- Translation
- Paraphrasing
- Drafting
- Summarization
- Substantive rewriting
- Other
Accountability
Did the authors independently verify all AI-assisted content?
Data confidentiality
Was unpublished or confidential material provided to an external AI system?
Documentation
Can the authors provide additional information about AI use if requested?
That is a much more useful system than a single checkbox saying:
"AI used: Yes/No."
What should researchers do today?
The safest strategy is surprisingly simple.
Don't try to predict the final AI policy.
Instead:
1. Record what you do
Maintain an AI-use log.
2. Be more specific than the minimum declaration
If AI paraphrased text, say so.
If AI generated code, say so.
If AI helped interpret results, say so.
If AI generated a figure, say so.
3. Preserve evidence
Keep:
- prompts;
- important outputs;
- code history;
- drafts;
- relevant version histories.
You don't necessarily need to save every trivial interaction forever. But significant AI contributions should be reconstructable.
4. Separate language assistance from scientific assistance
This distinction will probably become increasingly important.
5. Never allow AI to become an undocumented scientific collaborator
If AI contributed materially to scientific reasoning, analysis or interpretation, document that contribution.
6. Check the policy immediately before submission
Policies are changing rapidly.
For example, Elsevier updated its journal AI policies in August 2026, while Springer Nature has recently introduced a substantially more explicit risk-based framework.
A policy that was correct when a project started may not be the policy that applies when the paper is submitted.
The most important principle
There is a fundamental difference between:
"I used AI."
and
"I concealed how I used AI."
The first is rapidly becoming normal.
The second is likely to become increasingly problematic.
The direction of travel in scientific publishing is therefore not really toward AI prohibition.
It is toward:
AI + transparency + human accountability + provenance.
ICMJE, COPE, Springer Nature, Elsevier and other organizations are already converging on the principle that humans remain responsible for AI-assisted work and that meaningful AI use should be disclosed.
So what will happen to today's researchers?
Probably not a mass retrospective purge.
A more realistic future looks like this:
| Historical AI use | Likely future treatment |
|---|---|
| Grammar/spelling assistance | Essentially normalized |
| Minor paraphrasing | Usually low concern |
| Substantial paraphrasing/restructuring | Disclosure increasingly expected |
| AI-generated manuscript sections | Disclosure and scrutiny |
| AI-generated code | Increasingly expected to be documented |
| AI-assisted data analysis | Methods-level documentation likely |
| AI-assisted interpretation | Increasingly scrutinized |
| AI-generated figures | Strict provenance requirements |
| Undisclosed substantive AI use | Possible correction/investigation |
| AI-assisted fabrication | Research-integrity investigation |
| False AI declaration | Potentially more serious than simple non-disclosure |
The crucial question for researchers who have already used AI is therefore not:
"Have I ever used AI?"
Almost everyone eventually will.
The better questions are:
What did AI actually do?
Did it influence the scientific content?
Can I reconstruct what happened?
Did I accurately disclose it under the rules that applied at the time?
And, perhaps most importantly:
Can I demonstrate that the final scientific work is genuinely mine and that I remain accountable for it?
The future may actually reward transparency
There is an ironic possibility.
Researchers currently worry that declaring AI use will make their work look suspicious.
But if AI becomes ubiquitous, the opposite may happen.
A paper with a detailed declaration saying:
"AI was used to generate an initial coding implementation; all code was independently tested; AI was used to paraphrase approximately 8% of the manuscript; no AI was used for data interpretation or scientific conclusions; all references were independently verified."
may ultimately look more trustworthy than a paper that simply says:
"AI was used for language editing."
The first tells the reader what happened.
The second leaves the reader guessing.
And that may be the fundamental lesson of the AI transition:
The scientific community may ultimately care less about whether you used AI than whether you can honestly account for what AI did.
That is why researchers working with AI today should think beyond the declaration required by the journal.
Treat AI use as part of the provenance of the research.
Document it.
Verify it.
Disclose it.
And preserve enough information that, five years from now, you can still explain exactly how the work was produced.
No comments:
Post a Comment