What Should Researchers Do When AI Rules Change During an Ongoing Research Project?

What if the Rules Change After Your Research Has Already Started?

You began a multiyear project when your university permitted a particular AI service. Six months later, the policy changes. Or your funder revises its generative AI requirements, an ethics committee issues new guidance, or the journal you intend to submit to introduces an AI disclosure policy after your manuscript is already drafted.

Do you have to stop? Redo previous work? Apply the new policy retroactively? Can you simply continue under the rules that existed when the project began?

There is no universal answer. The critical questions are what changed, which authority issued the new requirement, when it takes effect, what activities it governs, whether it applies to existing projects, and whether your previous or planned AI use is affected.

Do Not Assume Either Automatic Retroactivity or Automatic Exemption

Policy Changes Need to Be Analysed by Timing, Scope, and Authority

First Confirm That the Policy Actually Changed

Do not rely on an email subject line, a colleague's warning, or a social-media post saying that "AI is now banned." Retrieve the authoritative policy and compare it with the version that previously applied.

Look for a version number, publication date, effective date, amendment history, transition provision, or explanation of what changed.

The University of Oxford's current Policy for using Generative AI in Research provides a useful example of explicit version control. Its September 2026 version is identified as version 1.1, lists amendments made on 7 September 2026, identifies the office responsible for maintaining the policy, and states that the policy will be reviewed on a termly basis.

UK Research and Innovation similarly maintains an update history for its generative AI application and assessment policy. Its official page records an update on 6 July 2026 and identifies earlier amendments.

This kind of version information allows researchers to distinguish a genuinely new requirement from a policy they simply had not noticed before.

Separate Past, Current, and Future Activities

A mid-project policy change should not be treated as one undifferentiated problem.

This temporal separation prevents two opposite mistakes: assuming that every new policy automatically invalidates earlier work, or assuming that a project started under older rules can ignore all later requirements until completion.

Check the Effective Date, Not Just the Publication Date

A policy may be published on one date and become effective on another. It may also contain transitional arrangements for applications, studies, awards, or submissions already in progress.

Read the operative wording carefully. Terms such as "effective immediately," "for applications submitted on or after," "from the next funding round," or "for new studies" materially change the answer.

For example, NIH's 2025 policy concerning substantially AI-developed research applications did not simply state a new principle without timing. NIH specified that the policy applied to applications submitted to the September 25, 2025 receipt date and beyond. That date matters when determining which applications fall within the requirement.

Do Not Assume New Policies Are Retroactive

Researchers sometimes react to a policy change by assuming that every earlier AI-assisted activity must now be judged under the new rule. That conclusion requires support from the policy or applicable authority.

A new disclosure requirement at a journal, for example, may apply to manuscripts submitted after its effective date even if the AI assistance occurred earlier. In that case, the earlier use may need to be disclosed at submission without having been prohibited when it occurred.

By contrast, a new institutional prohibition on uploading certain data to a particular AI service does not automatically establish that every earlier upload was a policy violation. The earlier conduct should be assessed against the requirements applicable when it occurred, unless the new rule or another governing obligation says otherwise.

Do Not Assume Existing Projects Are Grandfathered Either

The opposite assumption is equally risky. Starting a project under one policy does not automatically freeze every applicable requirement for the project's entire lifespan.

A university can revise information-security rules governing future data processing. A funder can update conditions where its authority and award terms permit. An ethics body can require researchers to respond to newly identified participant risks. A journal can apply the submission requirements in force when a manuscript is submitted.

Whether an existing project is exempt, subject to transition arrangements, or immediately affected must be determined from the relevant authority rather than inferred from the project's start date.

A Change in Tool Approval Can Require Immediate Attention

Some policy changes concern not abstract principles but the technology itself. An institution may withdraw approval for an AI service because its terms, security posture, data handling, ownership, or technical configuration changed.

If you are actively processing sensitive or confidential research information through a service that has just become prohibited, continuing to upload new material while waiting for philosophical clarity would be unwise. Pause the affected activity and contact the responsible office.

Stopping new processing does not necessarily mean deleting data, terminating the entire study, or taking other irreversible action. Those decisions should follow authoritative guidance and applicable data-retention requirements.

Changes Affecting Human-Participant Research May Require Ethics Review

A new institutional AI requirement may force you to change a procedure described in an approved ethics protocol. For example, you may need to replace an AI transcription service, alter where participant data are processed, introduce additional human review, or stop a planned automated analysis.

That creates two questions rather than one. First, what does the new institutional rule require? Second, does changing the research procedure require ethics review or an amendment?

Researchers should therefore revisit whether the ethics body overseeing the study needs to approve the revised AI workflow before implementation.

Funder Policy Changes Need Their Own Applicability Check

Funders regularly revise policies, but not every update applies identically to existing applications and awards.

UKRI's generative AI application and assessment policy, for example, was updated in July 2026, and UKRI announced further changes to its grant-assessment approach in September 2026 in response partly to generative AI. Researchers should use the funder's official policy history and the terms of the relevant opportunity or award to determine what applies.

This is why funder-specific AI requirements should be tracked separately from university policy.

Publication Policies May Change Long After the Research Was Conducted

Publication creates a particularly common timing problem. Research may begin years before the resulting article is submitted.

Elsevier states that it will continue to monitor developments and update its generative AI journal policies as practices, technologies, and standards evolve. Its journal policies were updated in 2026. Springer Nature likewise maintains evolving editorial policies based on how AI is used, while ICMJE expanded its recommendations on AI in publishing in January 2026.

Researchers should normally check the publication requirements that apply when preparing and submitting the manuscript, not assume that the journal's policy from the project's start date remains controlling.

If the new journal rule conflicts with an institutional requirement, use the process for resolving university and journal AI policy conflicts rather than ignoring either policy.

Changing Rules Can Affect Documentation Even When They Do Not Invalidate Past Work

Suppose you used AI for manuscript drafting when your target journal had no explicit disclosure requirement. Before submission, the publisher introduces one requiring authors to identify the tool and purpose.

The new requirement may not mean your earlier drafting was improper. It may mean you now need sufficient records to make an accurate declaration.

This is one reason documentation should begin when AI is used rather than when a journal finally asks about it.

Policy Changes May Reveal a Problem That Already Existed

Sometimes a new AI policy does not create a new underlying obligation. It clarifies how an existing rule applies.

For example, a university may publish new guidance explicitly stating that identifiable participant data must not be entered into public generative AI systems. That guidance might implement existing privacy, confidentiality, or security obligations rather than creating those obligations from scratch.

Researchers should therefore ask whether the amendment changes the rule, clarifies an existing rule, or changes only the recommended procedure. The answer can affect how earlier activity should be assessed.

An AI Service Becomes Restricted Halfway Through a Study

Common Mistakes When AI Policies Change Mid-Project

Treat a Policy Change as a Controlled Research Change

The goal is neither to panic nor to ignore the update. Treat the policy change as you would another consequential change in the research environment: verify it, determine its scope, assess its effect, obtain necessary approvals, and document the response.

For long projects, a formal process for tracking changing AI policies is more reliable than hoping someone notices an update before it matters.

When You Discover That an AI Policy Has Changed

Frequently Asked Questions About Changing AI Rules

A New AI Rule Does Not Erase the Project's Timeline

How to Cite This Guide

This guide is intended to be read, shared, and used in research, teaching, and academic work. If you draw on its ideas, explanations, or other content, please acknowledge the source by citing the guide. Doing so gives appropriate credit and helps your readers locate the original resource.