03 · What You Need to Know
How Responsibility Works When AI Enters the Research Chain
Start by Separating Cause From Responsibility
Suppose an AI assistant generates a nonexistent citation and a researcher includes it in a published article.
At one level, the model caused the false citation to appear in the interaction. That is a causal description.
But scholarly responsibility asks different questions. Who chose to use the system for reference generation? Who accepted the output? Who was expected to verify the reference? Who approved the manuscript? Who had authority to correct the problem?
Causal contribution
What event, system, person, or process contributed to the error occurring?
Accountability
Which human or organizational actors had duties, authority, or responsibility for preventing, detecting, correcting, explaining, or responding to the error?
Keeping those questions separate prevents the technologically interesting part of the story from obscuring the scholarly one.
AI Is Not an Accountable Research Author
Generative AI can produce text, code, analysis suggestions, summaries, images, and other substantial outputs. It still does not assume the human responsibilities associated with scholarly authorship or research accountability.
The European Commission's current living guidelines emphasize accountability, transparency, responsibility, and research integrity in generative AI use and place responsibility for scientific outputs on human researchers using these systems. The third edition was completed in May 2026 and updates the guidance as generative technologies and their research uses evolve.
This creates an important asymmetry:
AI can contribute content to research without being able to take responsibility for that content.
It cannot answer an editor's integrity inquiry as the accountable author. It cannot approve a correction. It cannot accept responsibility toward a participant whose confidential information was mishandled. It cannot defend a methodological choice as a scholarly actor.
The Researcher Who Uses AI Does Not Automatically Bear Every Possible Responsibility Alone
Human accountability does not mean that every AI-related failure should automatically be assigned entirely to the individual researcher who typed the prompt.
Research is often conducted through distributed systems.
A principal investigator may choose a tool. A research assistant may operate it. A statistician may validate the analysis. An institution may approve a platform. A service provider may make technical representations about the system. A collaborator may upload shared data without informing the rest of the team.
Responsibility may therefore be distributed.
The appropriate allocation depends on what each actor was expected and able to do.
A Useful Responsibility Analysis Asks Who Controlled What
When an AI-related error occurs, reconstruct the workflow.
| Question |
Why it matters |
| Who selected the AI system? |
Tool selection may carry responsibility for suitability and known limitations. |
| Who supplied the data or prompt? |
Input choices can create privacy, bias, scope, or interpretation problems. |
| Who relied on the output? |
Reliance determines where generated material affected the research. |
| Who was supposed to verify it? |
Defined review responsibilities help locate failures of oversight. |
| Who had the expertise to recognize the problem? |
Responsibility can depend partly on assigned competence and role. |
| Who approved the final research output? |
Final approval normally carries responsibility for what enters the scholarly record. |
| Were institutional safeguards adequate? |
Organizations may bear responsibility for governance, training, approved systems, or infrastructure. |
| Did the system behave contrary to reasonable representations? |
Some failures may implicate vendors, developers, or service providers in addition to researchers. |
This is more useful than searching for one universal person to blame.
Foreseeability Matters
Some AI failure modes are well known.
Generative systems can produce false factual information and fabricated citations. NIST identifies confabulation as a core generative-AI risk and also warns about automation bias, where users may defer excessively to apparently reliable AI outputs.
If a researcher uses an unverified generative model to create references and publishes them without checking whether the sources exist, “the AI hallucinated” is not a strong defense. The failure mode was sufficiently foreseeable that verification should have been part of the workflow.
Responsibility becomes more complicated when a system fails in an unexpected way that a reasonable researcher could not readily anticipate or detect.
That does not eliminate the need to correct the research record, but it can affect judgments about fault or negligence.
Errors and Misconduct Are Not Automatically the Same Thing
An AI-related error should not automatically be labeled research misconduct.
A researcher may make an honest error despite reasonable precautions. A system may fail unexpectedly. A collaborator may misunderstand a workflow. A software defect may produce an incorrect result that was not reasonably detectable at the time.
Conversely, knowingly using unreliable generated material, fabricating evidence through AI, concealing substantial AI use where disclosure is required, or ignoring obvious verification duties may raise more serious integrity concerns depending on the facts and applicable definitions.
Responsibility analysis should therefore distinguish the existence of an error from the separate question of whether someone's conduct meets a formal threshold for misconduct, negligence, questionable research practice, or another category.
The More Consequential the AI Output, the Stronger the Researcher's Verification Duty Usually Becomes
If AI suggests three possible manuscript titles and one is poor, little is at stake.
If AI generates the code that determines the study's primary result, the consequence is much greater.
This is why generative AI safeguards should be proportional to the research risk.
| AI contribution |
Possible consequence of error |
Expected researcher response |
| Title suggestion |
Minor misrepresentation |
Check fit and accuracy |
| Language revision |
Changed meaning or claim strength |
Compare against intended meaning and evidence |
| Generated citation |
False scholarly evidence |
Verify existence, metadata, relevance, and support |
| Generated analysis code |
Incorrect research results |
Inspect, test, validate, and understand the implemented method |
| Automated data classification |
Systematic distortion of evidence |
Validate performance, errors, bias, and methodological suitability |
| Generated interpretation |
Unsupported conclusion |
Evaluate against data, method, theory, uncertainty, and literature |
The researcher cannot reasonably apply the same level of checking to every interaction. Responsibility involves matching oversight to consequence.
“I Checked It” Is Not Enough If the Check Could Not Detect the Error
Suppose AI generates complex analysis code and the researcher cannot understand the programming language. The researcher runs it, obtains a graph, and asks the AI whether the code is correct.
The researcher has technically performed a check.
But the check lacks independence and may lack the expertise required to detect the problem.
Meaningful verification should be capable of discovering the relevant class of error. This is one reason researchers should not completely delegate tasks they cannot competently evaluate.
Co-Authorship Creates Shared Responsibilities
AI-related errors can also expose ordinary authorship problems.
Suppose one author uses AI to generate a literature section containing fabricated references. The other authors never check that section before approving the manuscript.
The person who generated and inserted the references may bear a particularly direct responsibility. But co-authors may also have responsibilities under the authorship and publication standards governing the work.
The precise allocation depends on roles, disciplinary norms, journal requirements, and the circumstances. It should not be assumed that “Author 2 used the AI” automatically absolves everyone else who approved the final article.
Principal Investigators and Supervisors Have Governance Responsibilities Too
Research assistants and students increasingly encounter AI within research teams. A principal investigator who simply says “use AI responsibly” without defining permitted tools, data restrictions, verification expectations, or documentation practices may create avoidable ambiguity.
Senior researchers do not need to inspect every prompt personally. But where AI use is consequential, research leadership may need to establish a defensible workflow.
This could include:
- which systems are approved for which information;
- what uses require additional permission;
- how outputs affecting data or analysis are validated;
- who reviews generated code or classifications;
- how AI use is documented;
- who checks current journal, funder, or institutional requirements.
Responsibility becomes easier to allocate when responsibilities were actually allocated before something went wrong.
Institutions Can Share Responsibility for the Environment They Create
Research organizations increasingly determine which AI systems researchers can access, what data may be processed, what training is available, and what governance applies.
The European Commission's living guidelines are explicitly addressed not only to researchers but also to research organizations and funding organizations, reflecting the fact that responsible AI use is not solely an individual matter.
An institution may therefore have responsibilities involving policy, infrastructure, approved services, training, risk assessment, data governance, and mechanisms for responding to problems.
That does not erase researcher responsibility. It means accountability can operate at several levels simultaneously.
Vendors and Developers Can Also Contribute to Failures
Researchers should not assume that all responsibility for an AI failure necessarily ends with the user.
A service may behave contrary to documented functionality. A software defect may produce erroneous processing. Security or privacy controls may fail. A model update may unexpectedly alter behavior. A provider may make misleading claims about capabilities or data handling.
Questions of legal liability depend on jurisdiction, contracts, facts, and applicable law and should not be collapsed into general research-integrity advice.
For scholarly practice, the important point is narrower: recognizing provider responsibility does not permit researchers to ignore risks that were within their own reasonable control, just as researcher responsibility does not mean providers can never be responsible for defective systems.
Responsibility Also Includes What Happens After the Error Is Discovered
Accountability is not only about preventing mistakes.
Suppose a researcher discovers after publication that AI-generated code mishandled missing data and changed the reported results.
The responsible response is not to ask whether the AI can be blamed convincingly enough to leave the paper untouched.
The researchers need to establish the extent of the error, rerun the analysis correctly, determine which findings or conclusions are affected, notify the appropriate co-authors or institutions, and work with the journal or other responsible body to correct the research record as required.
The original cause matters for understanding what happened and preventing recurrence. The obligation to respond to the known error remains.
Transparency Does Not Transfer Responsibility Either
A disclosure stating “Generative AI was used in the analysis” can be valuable.
It does not mean the researcher is no longer responsible for the analysis.
Transparency helps readers understand the process. It does not operate as a waiver of accountability.
The same principle applies throughout AI-assisted research: disclosure, verification, documentation, and human oversight complement responsibility rather than replace it.
A Practical Responsibility Test Is: Who Had the Last Reasonable Chance to Catch It?
No single test can allocate responsibility in every case, but one useful question is:
Who had the authority, information, expertise, and reasonable opportunity to prevent or detect this error before it affected the research record?
Sometimes that will point strongly toward one researcher. Sometimes it will reveal a team failure. Sometimes the institution failed to provide necessary governance. Sometimes the technical system behaved unexpectedly despite reasonable safeguards.
The purpose of asking is not merely to assign blame. It helps identify which safeguard failed and where the workflow should be repaired.
