Can AI Change the Balance of Risks and Benefits in an Approved Research Protocol?

Can AI Make an Approved Study More Efficient but Less Ethically Justifiable?

You have received ethics approval for a qualitative study examining university instructors' experiences with institutional expectations concerning generative AI. The approved protocol describes semi-structured interviews, secure data storage, and researcher-led thematic analysis.

After collecting the interviews, you discover an AI system that can summarize transcripts and generate preliminary analytical categories. Using it could substantially reduce the time needed for initial data organization.

However, the proposed system operates through an external provider. The transcripts contain identifiable professional experiences, and the AI-generated categories may introduce interpretations that were not anticipated in the original analytical plan.

The new workflow offers an apparent benefit, but it may also introduce risks that the ethics committee never evaluated.

Does faster analysis improve the study's risk-benefit balance, or could the additional risks outweigh the practical advantages?

Yes, AI Can Change Both the Risks and the Benefits of Research

How AI Can Alter the Risk-Benefit Balance of Research

What Is a Research Risk-Benefit Assessment?

A research risk-benefit assessment examines foreseeable harms or burdens associated with a study in relation to its anticipated benefits and scientific or social value.

Depending on the applicable framework, researchers and ethics committees may consider the likelihood and seriousness of potential harms, whether risks have been minimized, and whether remaining risks are reasonable in relation to anticipated benefits and the knowledge expected from the study.

The Belmont Report identifies beneficence as a fundamental ethical principle and emphasizes the importance of minimizing possible harms while maximizing possible benefits.

Under the United States Common Rule, 45 CFR 46.111(a)(2) requires that risks to subjects be reasonable in relation to anticipated benefits, if any, to subjects and the importance of the knowledge reasonably expected to result.

These frameworks do not establish that every benefit can compensate for every risk. Ethical acceptability also depends on other requirements, including informed consent, equitable participant selection, and appropriate protections.

Distinguish Research Benefits From Researcher Convenience

Generative AI may help researchers complete tasks faster, organize large datasets, or reduce repetitive work.

These advantages may improve the feasibility of a study. However, saving the researcher's time is not necessarily equivalent to producing a direct benefit for participants or increasing the scientific value of the research.

For example, reducing preliminary coding time from several weeks to a few days may be operationally useful. It does not establish that the resulting interpretations are more accurate or that participants benefit from the change.

Conversely, an appropriately designed AI-assisted accessibility feature might reduce a participant's burden or improve their ability to participate.

Researchers should identify who receives the proposed benefit and whether it is supported by evidence.

An operational advantage may contribute to research feasibility, but it should not automatically be counted as a participant benefit.

What Changes Might Affect the Risk-Benefit Balance?

Researchers should examine the specific differences between the approved protocol and the proposed AI-assisted procedure.

These are possible effects rather than established outcomes. Their significance depends on the actual system, research population, data, and safeguards.

Why the Original Ethics Approval Does Not Settle the New Balance

An approved protocol is evaluated under particular procedures and anticipated conditions.

If those conditions change, the original assessment may no longer address all relevant risks or benefits.

For example, a study approved for researcher-led analysis on institutional computers may later introduce an external AI service that processes identifiable transcripts.

The research questions remain unchanged, but the confidentiality arrangements differ.

Researchers should not assume that approval of the original methodology constitutes approval of every alternative workflow.

The appropriate response depends on the institution's post-approval modification procedures and whether the proposed change is material.

Researchers should review the requirements for introducing AI after ethics approval before implementing consequential modifications.

How Should Likelihood and Severity Be Considered?

Risk assessment should distinguish the possibility of harm from its expected likelihood and consequences.

For example, an AI-generated summary may contain an inaccurate interpretation. If it is used only as a preliminary suggestion and checked against the original transcript, the consequences may be limited.

The same error could be more consequential if the classification were used to evaluate an identifiable participant's professional competence.

Similarly, disclosure of a nonconfidential research outline differs from exposure of transcripts containing sensitive workplace allegations.

Researchers should consider the seriousness of potential harms, the likelihood of their occurrence, and the people or groups who may be affected.

Where reliable probability estimates are unavailable, qualitative risk assessments may be appropriate, but they should not be presented as precise quantitative predictions.

Can AI Reduce the Risks of a Study?

Potentially. Some AI applications may reduce participant burden or support protective procedures.

For example, an appropriately authorized system may assist researchers in identifying direct identifiers before preparing transcripts for further analysis.

AI-assisted accessibility tools may also support participation where their suitability has been established for the intended population.

However, these benefits require evaluation. Automated de-identification may overlook contextual identifiers, and AI-generated translations may distort participant meaning.

Researchers should compare the AI-assisted procedure with feasible alternatives and consider whether the proposed improvement is supported by appropriate evidence.

AI should not be described as reducing risk merely because the system automates a task previously performed by a human.

Risk Minimization Is Not the Same as Risk Elimination

Safeguards can reduce certain risks without eliminating them completely.

For example, removing direct identifiers from transcripts may reduce identification risks, but distinctive institutional details may still reveal participants' identities.

Similarly, using an authorized enterprise AI service may provide stronger contractual and technical protections without eliminating every possibility of unauthorized access or inappropriate interpretation.

Researchers should assess residual risk, meaning the risk that remains after safeguards have been implemented.

They should avoid claiming that a procedure is risk-free simply because a provider offers security controls or because a researcher intends to verify the output.

The question is whether the remaining risks are acceptable under the relevant ethical framework and institutional determination.

What Is the Difference Between Risk Reduction and Risk Transfer?

AI may reduce workload for researchers while transferring certain risks to participants or other organizations.

For example, outsourcing transcription to an AI provider may reduce the research team's administrative burden but introduce additional data processing arrangements.

The researcher benefits from efficiency, while participants may face a new confidentiality exposure.

That does not automatically make the arrangement unacceptable. It does mean that the benefits and burdens are not distributed identically.

Researchers should examine who benefits from the change and who bears the residual risks.

Fairness in risk distribution is particularly important when participants have limited influence over how their information is processed.

Why Data Minimization Matters to the Risk-Benefit Balance

Researchers may be able to reduce certain AI-related risks by limiting the information submitted for processing.

For example, an AI system used to organize teaching-related excerpts may not require participant contact information, signed consent forms, or unrelated personal disclosures.

Removing unnecessary information can reduce exposure without necessarily diminishing the analytical purpose.

However, excessive removal of context may undermine the validity of qualitative interpretation.

Researchers should preserve information necessary for defensible analysis while avoiding unnecessary personal data processing.

The principle of data minimization in AI-assisted analysis can therefore inform the design of proportionate safeguards.

Can AI-Generated Inferences Change the Balance?

Yes. AI may produce classifications or inferred characteristics beyond those collected for the original research purpose.

For example, a study examining instructors' AI-related teaching practices may generate AI-based predictions about participants' psychological states or professional competence.

These inferences may be unnecessary, unreliable, or sensitive.

If linked to identifiable participants or used in consequential decisions, they may introduce additional risks without a corresponding research benefit.

Researchers should examine whether the inferred information is necessary and justified before allowing it to become part of the analytical workflow.

The concern is not simply whether AI can generate sensitive inferences, but whether their creation and use improve the study sufficiently to justify the associated risks.

Should Researchers Use a Numerical Risk-Benefit Score?

A numerical scoring system may help organize a preliminary assessment, but it should not be mistaken for a universally valid ethical calculation.

For example, assigning a risk score of three and a benefit score of five does not establish that the proposed procedure is ethically acceptable.

Risks and benefits may differ in kind, affect different people, and involve uncertainties that cannot be meaningfully combined into a single number.

Some ethical requirements, such as appropriate consent and lawful data processing, cannot simply be traded away because a researcher assigns a high score to expected efficiency.

The NIST AI Risk Management Framework (2023) supports context-sensitive identification, measurement, and management of AI risks rather than treating risk evaluation as a universal arithmetic exercise.

Researchers may use structured matrices to clarify decisions, but the final ethical judgment requires substantive reasoning and applicable institutional review.

How Should You Prompt AI to Compare Risks and Benefits?

AI can assist with organizing a comparison between approved and proposed procedures. It should not be asked to determine ethical acceptability on behalf of the researcher or review body.

This prompt may help identify matters requiring further examination. The researcher should verify factual claims and obtain the applicable institutional determination before implementing a modification requiring prior approval.

When Does the New Balance Require Ethics Review?

A proposed AI activity may require institutional notification, amendment, or additional approval when it materially changes the research procedures or participant risks.

For example, introducing an AI interviewer may change how participants interact with the study and how unexpected responses are managed.

Using an external provider for identifiable transcripts may change confidentiality and data protection arrangements.

However, not every modification necessarily requires a full review. Institutions may have different procedures for minor and consequential changes.

Researchers should seek an authorized determination when the proposed AI use may alter the conditions under which the original protocol was approved.

Submitting a modification request does not necessarily authorize implementation while the decision is pending.

When Should the Researcher Reject the AI-Assisted Procedure?

Sometimes the proposed AI workflow offers limited benefit while introducing risks that cannot be adequately justified or managed.

For example, submitting identifiable transcripts to an external service may be unnecessary when the intended analytical task can be performed through an approved local workflow.

If the proposed provider lacks suitable safeguards or the processing is inconsistent with consent and institutional authorization, the researcher should not proceed merely because the system is convenient.

The appropriate decision may be to select a safer service, redesign the workflow, or retain the original procedure.

Researchers should consider when not using AI is the more appropriate ethical decision.

Reassess the Balance When the AI Workflow Changes Again

An initial risk-benefit assessment does not necessarily cover every future modification.

For example, a researcher may first use AI to organize de-identified excerpts and later expand the task to complete transcripts or participant profiling.

The expanded workflow may involve different information, purposes, and consequences.

Changes in provider arrangements or institutional requirements may also affect the assessment.

Researchers should therefore revisit the risk-benefit balance when consequential changes occur and document the resulting decisions.

Ethical evaluation should remain connected to the actual research procedures rather than become a one-time statement made at protocol submission.

Comparing Manual Analysis With External AI-Assisted Analysis

The decision cannot be reduced to whether AI is faster than manual analysis. The researcher must consider whether the proposed arrangement changes the risks, whether the expected benefits are credible, and whether the residual risks are acceptable under the applicable ethical requirements.

Common Misconceptions About AI and Research Risk-Benefit Assessment

Compare the Ethical Consequences, Not Just the Efficiency of the Tools

Before introducing AI into an approved study, compare the existing procedure with the proposed alternative. Identify what benefits are expected, who receives them, and what new or altered risks may arise.

Then evaluate whether feasible safeguards reduce those risks sufficiently and whether the proposed modification requires institutional review.

Document the comparison, including unresolved uncertainties and alternatives considered.

The purpose of reassessment is not to defend a preferred technology. It is to determine whether the proposed AI-assisted procedure remains ethically justifiable in light of its actual benefits and risks.

Before Changing an Approved Research Protocol to Include AI

Frequently Asked Questions About AI and Research Risk-Benefit Assessment

AI May Improve the Workflow Without Improving the Ethical Balance

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.