What must Cornell researchers now document when they use AI?
Cornell’s new interim policy makes principal investigators accountable for verification, disclosure, data control and human oversight. It is a binding institutional rule, not evidence that AI-assisted research has become safer or more reproducible.
The short version
- What happened
- The policy applies to everyone conducting research at Cornell on any campus and assigns ultimate accountability to the principal investigator or, for a solo student researcher, the faculty mentor.
- Who it affects
- Principal investigators and faculty mentors now have explicit responsibility for protocols, team training and records when AI materially affects research.
- What remains unproven
- The source is Cornell's own policy and announcement; no independent evaluation of implementation is available.
Keep in mind
The source is Cornell's own policy and announcement; no independent evaluation of implementation is available.
The policy applies within Cornell and does not establish a universal legal or disciplinary standard for other institutions or countries.
It gives principles and responsibilities rather than a single disclosure template or quantitative materiality threshold.
No baseline denominator for AI use, incidents, errors, productivity or disclosure is reported.
No post-implementation outcome has yet been measured because the policy was issued on 2 November 2026.
Specific sponsor, publisher, ethics, safety, patent and data rules can add stricter obligations.
Multi-source analysis · 2 linked sources. Primary analysed source date: 2 November 2026.
Keep following the evidence
Join The Impact Brief founding list for selected AI reports, their effect on people and the evidence limits. Read a sample first.
Read the full report · 10 min The evidence, methods, implications and limitations
The direct answer: keep enough evidence to reconstruct what the AI changed
Cornell's interim Policy 1.8, issued on 2 November, tells researchers to document AI use in enough detail for a reader to understand and, where applicable, reproduce the work. Principal investigators must ensure that use is recorded in project data or notes for as long as the underlying research records must be retained. Cornell urges disclosure across idea development, writing, code, presentations and figures, with the exact level shaped by sponsor, journal and disciplinary rules. Embedded AI in laboratory software should be identified with the software version when it affects methods.
The practical test is not whether a tool was called 'generative AI'. It is whether its output influenced the design, conduct, interpretation or communication of the research. A spelling correction may require less explanation than code that cleaned a dataset, a model that selected participants or an agent that changed a laboratory setting. The policy does not publish a universal disclosure template or a fixed prompt-log requirement. Teams therefore need a proportionate record showing the tool and version, the task, important inputs, human checks, changes accepted and any compliance approval.[1][2]
Who is accountable
The rule applies to everyone who conducts research at Cornell on any campus. Each team member remains responsible for verifying AI output and preventing bias or inaccuracy, but the principal investigator carries ultimate accountability for protocols and team understanding. When a student works alone or leads a project, the faculty mentor takes the PI role for this purpose. Cornell explicitly says responsibility for research conduct, participant protection, safety and regulatory compliance cannot be delegated to an AI system.
That allocation matters when an agent performs several linked tasks. A system can suggest an experimental design, generate code, control an instrument and draft a result without becoming the accountable investigator. The PI must decide what can be automated, who reviews it and what evidence is kept. Output connected to experimental design, hazard identification or regulatory compliance must also be reviewed with the relevant compliance or institutional safety office. A plausible answer from a model is not a substitute for approval by the body responsible for the risk.[1]
The distinction that matters
A practical record for material AI use
Cornell does not mandate one form, but its responsibilities imply a trace that lets another person understand the system's role and the human checks.
Tool and version
Record the service, model or embedded software version and approved environment
Identifies what system produced or influenced the work
Task and data
Describe the purpose, material inputs and any sensitive-data boundary
Shows what the system was allowed to see and do
Human verification
Keep checks, corrections, approvals and unresolved uncertainty
Demonstrates that accountability remained with the team
Disclosure and retention
Match sponsor, journal, participant and university rules and retain the record
Supports attribution, audit and reproducibility
Sensitive data face an approval boundary
Researchers must use institution-approved tools when available and examine a system's data-sharing practices before use. Cornell prohibits sharing high- or moderate-risk institutional data with an AI system unless that tool has passed appropriate internal review. The policy names private, export-controlled, sensitive and proprietary information, including identifiable participant data and student grades protected by US education privacy law, as material risks. When no approved tool exists, researchers must use the university's IT-governance process rather than choosing a public service informally.
The policy takes a deliberately cautious position on commercial promises that inputs will not be used for training. Researchers are told to begin from the assumption that information entered into a commercial tool could become training data and later be recoverable, unless the system runs in a Cornell-approved controlled environment. That is a risk-control rule, not a finding that every commercial service retains or leaks every prompt. Teams still need to inspect the actual contract, configuration, retention, region, access controls and deletion process for the tool they propose to use.[1][2]
Disclosure depends on materiality and context
Cornell requires disclosure when AI produces substantive portions of scholarly output or is an important part of the method. The purpose is attribution and reproducibility: readers should not mistake machine-generated material for human intellectual contribution, and they should be able to judge how the system affected the evidence. The policy recognises that expectations differ across fields, publishers and sponsors. Researchers must therefore follow the strictest applicable grant, journal, contract and professional requirements rather than assuming one university statement satisfies every destination.
Human-participant work has an additional layer. If AI affects participants, the institutional review board decides whether it must be described in consent. A study that hides AI use to test whether people can identify machine output is treated like other research involving deception and still requires ethical and regulatory review. Disclosure to a journal, disclosure to a participant and internal documentation serve different purposes. A team may need all three, with different detail, even when public release of a full prompt would expose confidential material or undermine a study design.[1]
Patent claims and creative work get specific warnings
The policy tells investigators not to use AI in a way that leaves a Cornell patent claim without an identifiable human inventor. An AI tool may help describe an invention, but the claimed inventive contribution must come from people and the project record must make that clear. This is not a new patent statute, and Cornell does not decide inventorship for every jurisdiction. It is an institutional control intended to preserve evidence before a patent filing or later dispute makes reconstruction difficult.
For scholarship that becomes a registered creative work, researchers must distinguish material made by people from more-than-minimal AI-generated content according to the norms of the field. The wording does not define a universal quantitative threshold. That leaves judgment with investigators, editors and rights specialists, but it also removes an easy excuse for silence. Teams using image, music, text or code generators should keep dated drafts, tool versions and human revisions so authorship and rights questions can be answered from records rather than memory.[1]
Agents and laboratory systems remain inside existing oversight
Cornell names autonomous, agentic, embodied and generative systems and says none is exempt from existing research oversight. Work involving biological agents, hazardous materials, radiation, physical hazards, people or animals may require review by the office already responsible for that domain. If an AI system can materially influence laboratory operations, safety decisions or participant welfare, researchers must preserve human oversight and accountability. The policy does not authorise unsupervised control merely because a vendor describes a system as an assistant or because a model can execute tools.
For laboratories, this means defining action boundaries before connecting an agent to instruments or workflows. A system that drafts a protocol poses a different risk from one that changes a temperature, selects a dose or schedules a participant. Useful controls include simulation, least privilege, approval gates, rollback, logs and an identified person able to stop the process. Cornell's policy lists responsibilities but does not certify a technical architecture. Each project still needs a hazard analysis matched to what the system can actually observe and change.[1]
The policy has enforcement, but no outcome study
Cornell's research-integrity, compliance and audit offices may investigate adherence. The policy says non-compliance can be handled through applicable procedures and progressive discipline, up to termination. It is also explicitly interim and will be reviewed annually while AI technology remains unstable. Those features make it more than optional advice. They do not tell researchers how often enforcement will occur, which documentation failures will be treated as minor or serious, or how appeals and remediation will work in a specific case.
No baseline accompanies the rule. Cornell does not report how many researchers use AI, what proportion already disclose it, how many sensitive-data incidents have occurred, or whether AI-assisted projects currently have more errors. There is no comparator university and no measured post-policy change in reproducibility, safety, time or publication quality. The evidence supports a governance conclusion: Cornell has assigned responsibilities and boundaries. It does not support a claim that the policy has already made research safer, faster or more trustworthy.[1]
What would show whether the policy works
A useful evaluation would begin with a transparent denominator: how many projects use AI, for which tasks and risk categories, and how many use approved environments. Cornell could then measure disclosure completeness, reproducibility problems, sensitive-data events, protocol deviations, participant concerns and corrections before and after implementation. Results should distinguish administrative burden from meaningful risk reduction and protect confidential research. Surveys alone would be insufficient if teams under-report prohibited or informal use.
The strongest evidence would compare similar projects or departments over time, document how reviews were applied and publish anonymised failure cases alongside successes. Researchers also need worked examples: minimum records for coding assistance, prompts affecting analysis, embedded instrument AI, participant-facing chatbots and autonomous laboratory agents. If annual revisions clarify proportionality while incidents and irreproducible outputs fall without blocking legitimate work, confidence would rise. If disclosure becomes a box-ticking exercise or pushes use into unrecorded channels, the policy would need redesign.[1][2]
Watch the background, then check the evidence
Illustrative photo · not a video frame
Google DeepMind · 1 min 52 sec · Background explainer
Protein folding explained
YouTube loads only when you choose; playback starts when you press play.
A short introduction to why a protein’s shape matters. It provides background for understanding what AI structure prediction contributes to biological research.
Evidence check: This is the developer’s explanation. Predicting a structure does not by itself establish a treatment’s safety or benefit for patients.
Published 2020-11-30. Source checked 14 October 2026. Check captions on the publisher’s video page.
Explore current researchContinue the story
Read next
Continue exploring ai disclosure: compare the finding and its evidence limits.
When should newsrooms disclose AI use?
A peer-reviewed case study based on 13 interviews with 12 Financial Times managers and 28 internal documents finds that AI disclosure is treated as a spectrum shaped by oversight, risk and context. It describes one newsroom's practice and does not test whether labels improve audience trust.
Read the next reportExplore the full source and evidence record
What this means for people
- Principal investigators and faculty mentors now have explicit responsibility for protocols, team training and records when AI materially affects research.
- Students and research staff need approved tools, clear data boundaries and a practical way to document corrections rather than relying on informal personal accounts.
- Participants and laboratory personnel retain human safeguards; an AI system cannot inherit responsibility for consent, welfare, hazards or regulatory compliance.
Global context
Research organisations are moving from broad AI principles toward operational rules for data, disclosure and accountability. Cornell's policy is notable for joining publication transparency to laboratory safety, participant protection and institutional discipline. Its reach is still local: other universities, funders, journals and jurisdictions use different definitions and thresholds. The transferable lesson is to preserve a verifiable human decision trail, not to assume Cornell's wording satisfies every research setting.
What the evidence does not yet show
- The source is Cornell's own policy and announcement; no independent evaluation of implementation is available.
- The policy applies within Cornell and does not establish a universal legal or disciplinary standard for other institutions or countries.
- It gives principles and responsibilities rather than a single disclosure template or quantitative materiality threshold.
- No baseline denominator for AI use, incidents, errors, productivity or disclosure is reported.
- No post-implementation outcome has yet been measured because the policy was issued on 2 November 2026.
- Specific sponsor, publisher, ethics, safety, patent and data rules can add stricter obligations.
What to watch next
- Cornell's worked examples or minimum documentation templates for common research tasks
- Which AI tools and configurations receive institutional approval for different data classes
- Anonymised counts of reviews, incidents, corrections and compliance findings
- How IRBs handle participant-facing AI and studies that conceal AI use
- Whether annual revisions define proportional records without encouraging box-ticking
- Measured effects on reproducibility, researcher workload, safety and legitimate AI adoption
Living evidence record
Impact record IAI-0P671E4
Evidence stage
Announced
Confidence
Supported
Reporting basis
Multi-source analysis
Independent or research support
Not yet
Record status
Monitoring
Last checked
3 November 2026
Source trail
2 direct sources across 2 source types.
People impact
Documented in this record.
Uncertainty
Limits and next checks are explicit.
Stages describe the evidence available—not whether a technology is good or bad. See the public method.
Evidence trail
Sources used for this report
Links checked 3 November 2026
Research question, topics and regions
How Cornell's interim AI research policy changes investigators' responsibilities and what outcomes remain untested
Topics: Research integrity · AI disclosure · Sensitive data · Human oversight · Reproducibility · University policy
Regions: United States
This report is labelled multi-source analysis. We summarise and analyse source material in our own words; company statements remain attributed claims until independently supported. Translated summaries preserve the meaning of the original source and link back to it. Read our editorial standards.
Reader commentary
What does this mean where you are?
Tell us what you have seen at work, in your community or in everyday life. A question is welcome too. No account is needed; we review notes before publishing them.
Explore commentary across the portal →Published reader notes
0No published reader notes yet. You can start the evidence-led discussion above.
Prefer a private correction or response? Contact the newsroom.