1. Introduction

Algorithm version changes mid-trial refers to a situation in which an artificial intelligence (AI) or machine-learning system used in a clinical trial, medical research study, or regulatory investigation is modified after the trial has begun but before it is completed.

Examples include:

A diagnostic AI model is updated halfway through a clinical trial.

A genomic prediction algorithm changes its risk-scoring method during patient recruitment.

An AI system used to determine clinical-trial eligibility is retrained using new patient data.

A software update changes the algorithm's sensitivity, specificity, outputs, or treatment recommendations.

A sponsor changes the model but continues reporting results as though the same system had been used throughout the study.

The central legal question is:

When must researchers, sponsors, software developers, or trial investigators disclose an algorithm change, to whom must it be disclosed, and what happens if they fail to do so?

There is no single universal statute specifically titled Algorithm Version Changes Mid-Trial Disclosure Law. Instead, the applicable obligations arise from clinical-trial regulations, research ethics, informed-consent requirements, data-integrity rules, medical-device regulation, contractual duties, and litigation disclosure principles.

The precise obligations depend on the jurisdiction, the type of trial, the significance of the change, and whether the algorithm affects participant safety or the validity of the study's results.

2. What constitutes an algorithm version change?

Not every software modification has the same legal significance.

Level 1 — Technical maintenance

Examples: correcting a display error, improving loading speed, or fixing a non-substantive interface issue.

Disclosure may be required by change-control policies, but the modification might not require a new clinical protocol or participant notification if it has no effect on trial conduct, data, or safety.

Level 2 — Performance or configuration change

Examples: changing a classification threshold, updating a reference dataset, or adjusting a risk-score calibration.

May require documented validation, a protocol amendment, regulatory notification or approval, and an assessment of whether results from different versions remain comparable.

Level 3 — Material model change

Examples: retraining the model, changing its architecture, adding new clinical inputs, or materially changing diagnostic predictions.

Requires especially careful assessment of safety, performance, trial validity, regulatory status, consent implications, and the need to pause or modify the trial.

A change's legal significance depends on its actual effect, not merely its software version number. Even a small code change can be material if it changes patient classification or treatment decisions.

3. Principal legal duties arising from a mid-trial algorithm change

A. Duty to disclose the change to regulators

The sponsor must assess whether the new version changes the approved or submitted clinical-trial protocol, the investigational device, or the scientific validity of the trial.

In the United States, 21 CFR § 312.30 requires an investigational new drug (IND) protocol amendment for specified changes that significantly affect safety, the scope of the investigation, or scientific quality. Such changes generally must be submitted before implementation, subject to the regulation's exception for changes intended to eliminate an apparent immediate hazard.

LII / Legal Information Institute

+1

A mid-trial algorithm update may trigger this requirement when, for example, it:

Changes the primary endpoint or the way it is measured.

Alters participant eligibility or treatment allocation.

Materially changes diagnostic performance.

Changes the safety-monitoring process.

Undermines the scientific assumptions underlying the trial.

A software update does not automatically require an IND amendment. The legal question is whether the change falls within the applicable regulatory requirements.

For medical devices, separate investigational-device rules may apply. The FDA's guidance on modifications during a clinical investigation explains how sponsors assess whether a change can be implemented without prior agency approval or requires additional regulatory review.

FDA

B. Duty to inform the ethics committee or institutional review board

The relevant ethics committee or institutional review board (IRB) must receive information about changes that affect the ethical acceptability, risk profile, consent materials, or approved conduct of the trial.

The sponsor and investigator should assess:

Whether the change falls outside the currently approved protocol.

Whether additional validation or a revised risk assessment is required.

Whether participants face a new or increased risk.

Whether the change requires prior review and approval.

Whether any urgent safety exception applies.

A change should not be treated as ethically insignificant simply because it was deployed through an automated software update.

C. Duty to inform trial participants

Disclosure to regulators and ethics committees is distinct from disclosure to participants.

Not every technical change requires fresh informed consent. However, participant notification or renewed consent may be necessary if the change materially affects:

The foreseeable risks or benefits of participation.

The procedures performed on participants.

The collection or use of genetic or other personal data.

The treatment or monitoring participants receive.

Information that a reasonable participant would consider important when deciding whether to continue participating.

If the revised model changes a patient's clinical management, the investigator must also assess any immediate duty to explain the resulting medical decision.

D. Duty to preserve records and disclose the change in study results

A defensible clinical-trial record should identify:

The exact algorithm version used for each participant and assessment.

The date and reason for each update.

The model's training data, configuration, and validation status, as relevant.

The performance of the old and new versions.

Whether earlier data were reprocessed using the new version.

Any effect on endpoints, statistical analysis, missing data, and safety outcomes.

ICH Good Clinical Practice guidance emphasizes controlled computerized systems, traceable changes, audit trails, data integrity, and risk-based review.

European Medicines Agency (EMA)

+1

The key principle: the trial record must make it possible to reconstruct which system generated each result and how that result was handled in the final analysis.

4. Important case laws and their application

There is an important qualification: the following are established decisions concerning informed consent, expert evidence, scientific reliability, or research ethics. They are not direct precedents holding that a particular AI algorithm update during a clinical trial is unlawful. Their relevance is by analogy to the legal duties implicated by such an update.

Case 1: Daubert v. Merrell Dow Pharmaceuticals, Inc. (1993)

United States Supreme Court · 509 U.S. 579 (1993)

Legal principle: Scientific evidence offered in federal court must satisfy standards of relevance and reliability.

In Daubert, the Supreme Court addressed whether expert scientific evidence concerning an alleged connection between a drug and birth defects was admissible. The Court assigned trial judges a gatekeeping role in evaluating scientific expert testimony under the Federal Rules of Evidence. Relevant considerations include testability, peer review, potential error rates, and standards governing the technique.

LII / Legal Information Institute

+1

Application to algorithm changes mid-trial

Suppose a sponsor uses AI Version 1 for the first 500 patients and AI Version 2 for the remaining 500. Version 2 has a lower reported error rate, but the sponsor has not validated the difference between the versions.

If the sponsor later presents the combined dataset as evidence that the diagnostic method is reliable, the court may examine:

Whether each version was independently validated.

Whether the results are comparable across versions.

Whether the reported error rate accurately represents the system used.

Whether the expert disclosed the update and its implications.

Daubert does not itself impose a clinical-trial notification obligation. Its importance is that an undisclosed change can undermine the scientific foundation of expert testimony.

Case 2: Canterbury v. Spence (1972)

United States Court of Appeals for the District of Columbia Circuit · 464 F.2d 772

Legal principle: Information material to a patient's treatment decision may need to be disclosed so that the patient can make an informed choice.

The court emphasized the patient's right to make an informed decision and the importance of disclosing material risks. The case concerned medical treatment, not AI or clinical-trial software.

biotech.law.lsu.edu

+1

Application

Imagine that participants consent to a trial in which an AI system identifies candidates for a particular treatment. Midway through the trial, researchers introduce a new model that substantially changes who qualifies for treatment or creates a previously unrecognized risk.

If that change materially affects the risks or nature of participation, the investigator should assess whether participants need updated information or renewed consent.

The legal analysis should distinguish between:

A change that is merely technical and has no material effect on participation.

A change that alters the risks, procedures, or treatment participants may receive.

A change that affects the continued validity of the participant's original consent.

The case supports a materiality-based disclosure analysis; it does not establish that every algorithm update requires re-consent.

Case 3: Samira Kohli v. Dr. Prabha Manchanda (2008)

Supreme Court of India · Civil Appeal No. 1949 of 2004 · decided 16 January 2008

Legal principle: Valid medical consent must be voluntary and based on adequate information concerning the nature and purpose of the proposed procedure.

The Supreme Court of India explained that consent must enable a patient to make a balanced judgment about whether to undergo the proposed treatment. The case concerned consent for medical procedures, rather than clinical-trial algorithms.

Indian Kanoon

Application in India

Consider a clinical trial using an AI-based genomic diagnostic system. Participants are told that a specified model will be used, but the sponsor subsequently introduces a materially different model that changes the diagnostic process or the risks associated with participation.

The principles in Samira Kohli support examining whether the original consent remains adequate in light of the new circumstances.

However, the case should not be interpreted as requiring consent for every internal software modification. The question is whether the change materially affects what participants are agreeing to and whether applicable clinical-trial rules require an updated consent process.

Case 4: Moore v. Regents of the University of California (1990)

Supreme Court of California · 793 P.2d 479 (Cal. 1990)

Legal principle: Physicians may have disclosure obligations concerning material personal interests that could affect their professional judgment.

In Moore, the court considered allegations concerning a physician's undisclosed economic and research interests relating to biological materials obtained from a patient.

Application to algorithm changes

Suppose a trial sponsor adopts a new AI model developed by an affiliated company and has a financial interest in demonstrating that the updated model performs better than competing systems.

The case may provide an analogy for examining whether a material conflict of interest should have been disclosed. It does not establish a general duty to disclose every commercial relationship involving AI software, nor does it directly govern version changes. Any actual disclosure duty depends on the applicable law and facts.

5. Regulatory framework: United States, European Union, and India

United States

Key rules: 21 CFR § 312.30; applicable IRB rules; investigational-device requirements where relevant.

The sponsor must evaluate whether the change requires a protocol amendment or other regulatory action. Material changes affecting safety, study scope, or scientific quality may require an amendment before implementation. A change intended to eliminate an apparent immediate hazard can fall within a specific exception.

LII / Legal Information Institute

+1

Relevant parties include the FDA, the IRB, the sponsor, and the principal investigator.

European Union

Key framework: Clinical Trials Regulation (EU) No 536/2014, applicable medical-device rules, and Good Clinical Practice requirements.

A significant change to a clinical trial may qualify as a substantial modification requiring regulatory and ethics review under the applicable framework. The sponsor must assess the nature and consequences of the change rather than assuming that software updates are automatically exempt.

For AI used as a medical device, separate device-change and conformity-assessment considerations may also apply. The exact procedure depends on the product and trial.

India

Key framework: New Drugs and Clinical Trials Rules, 2019; applicable ethics committee requirements; informed-consent requirements; and other relevant drug, medical-device, and data-protection laws.

An algorithm update must be assessed to determine whether it changes the approved clinical-trial protocol, participant risk, study procedures, or the basis on which consent was obtained.

The sponsor and investigator should follow the applicable amendment, ethics-review, safety-reporting, and consent procedures. The exact obligation depends on whether the AI is part of an investigational drug trial, a medical-device investigation, or another research setting.

An important international reference is the SPIRIT-AI extension, which recommends specifying the AI algorithm version in the trial protocol and, where applicable, describing differences from versions used in earlier studies and explaining why those changes were made. This is a reporting guideline, not an independent statute imposing a universal disclosure duty.

PMC

6. Who must receive disclosure?

The required recipient depends on the nature of the change.

These are not identical duties. For example, a sponsor may have to inform the regulator without every participant necessarily requiring a new consent form. Conversely, an urgent clinical risk may require prompt participant protection even before every administrative process has been completed.

7. What if the sponsor fails to disclose the change?

Depending on the applicable law and the consequences, nondisclosure may lead to several different forms of liability.

Regulatory noncompliance: An unreported or improperly implemented change may result in inspection findings, corrective actions, restrictions on trial conduct, or other regulatory consequences.

Research ethics violations: Failure to obtain necessary review or provide material participant information may breach research-ethics requirements.

Scientific-integrity concerns: Combining results from materially different model versions without appropriate disclosure or analysis may render the reported findings misleading or unreliable.

Negligence or product-liability claims: If an inadequately validated update contributes to a harmful clinical decision, the sponsor, manufacturer, investigator, or other responsible party may face claims, depending on causation and applicable law.

Evidentiary consequences: In litigation, an undisclosed version change may undermine an expert's credibility or the admissibility or weight of the evidence, depending on the governing procedural rules.

Nondisclosure alone does not automatically establish civil liability. A claimant generally must satisfy the elements of the relevant legal claim, including any required proof of harm, causation, and breach.

8. Practical example: a genomic AI trial

Assume a clinical trial evaluates an AI system that predicts whether patients have a genetic condition.

Month 1 — Trial begins

Version 1 is validated, specified in the protocol, and used to classify participants.

Month 4 — Version 2 is deployed

The sponsor retrains the model. Its sensitivity improves, but it also changes the number of patients classified as high risk.

Required assessment

The sponsor assesses validation, participant safety, protocol compliance, regulatory reporting, ethics approval, and whether new consent information is needed.

Final report

The sponsor identifies the version used for each assessment, explains any reprocessing, documents the change, and evaluates whether the combined results remain scientifically valid.

If Version 2 is materially different, the sponsor should not simply merge both versions' outputs and report a single performance figure without evaluating comparability. Depending on the protocol and regulatory requirements, the appropriate response might be an amendment, separate analyses, additional validation, a pause in deployment, or a revised trial design.

9. Recommended compliance checklist

Before introducing a new algorithm version during a trial, the responsible team should document the following:

The exact changes from the previous version.

The rationale for the update and its expected clinical impact.

Validation results, including performance differences and limitations.

Whether the approved protocol or regulatory submission must be amended.

Whether prior ethics committee or IRB approval is required.

Whether participants need updated information or renewed consent.

The implementation date and the participants assessed with each version.

Whether previously collected data will be reprocessed.

How the statistical analysis will address version differences.

Audit trails, access records, training, and retained validation documents.

This checklist is a practical compliance framework, not a substitute for jurisdiction-specific legal advice.

10. Conclusion

The governing legal principle is that a clinically material algorithm change during a trial must be assessed, documented, and disclosed to the appropriate recipients whenever the applicable regulatory, ethical, consent, or procedural rules require it.

Three propositions are particularly important:

A software update is not automatically a reportable protocol amendment; its effects and the governing rules determine the obligation.

A change that affects participant safety, treatment, consent, endpoints, or scientific validity deserves substantially greater scrutiny than routine technical maintenance.

Version histories, validation records, audit trails, and transparent reporting are essential to demonstrating that the trial's conclusions remain trustworthy.

The most relevant authorities discussed here are Daubert v. Merrell Dow Pharmaceuticals for scientific reliability, Canterbury v. Spence for material-risk disclosure, and Samira Kohli v. Dr. Prabha Manchanda for informed consent in India. They support the underlying principles, but none should be mischaracterized as a direct judicial ruling on mid-trial AI algorithm version changes.