AI-related evidence is appearing in federal litigation in several forms. It may include AI-generated text, synthetic images or audio, facial-recognition results, predictive scores, automated transcripts, or expert analysis supported by machine-learning or other automated tools.
These different forms of evidence raise different issues. A synthetic recording offered to prove what someone said may raise an authentication issue and, depending on the purpose for which it is offered, a hearsay issue. An expert opinion that materially relies on an AI model may require closer review of the methodology, validation, and case-specific application.
For now, these disputes remain governed by the existing Federal Rules of Evidence. Proposed Rule 707 remains under consideration, but it has not been adopted. In May 2026, the Advisory Committee revised the published proposal and deferred further action while it continued studying the issue.
This guide explains how Rules 901, 702, 703, 801, and 403 may apply to AI-related evidence. It also highlights preservation gaps, weak validation, and expert-vetting failures that may support discovery relief, an evidentiary objection, limits on testimony, or a motion to exclude, depending on the circumstances.
What Counts as AI-Related Evidence?
AI-related evidence may include:
- AI-generated text, images, audio, or video
- Facial-recognition or image-classification results
- Machine-learning predictions, rankings, or risk scores
- Automated translations, transcripts, or summaries
- Expert opinions developed with AI-assisted tools
- Reports or conclusions produced by AI or other automated systems
The type of evidence, the claim made about it, and the purpose for which it is offered shape the required foundation.
An AI-generated image may raise a Rule 901 authentication issue. An automated transcript may contain human statements that require a hearsay analysis. A predictive model used by an expert may require closer review under Rule 702.
A useful starting point is to ask: What is the evidence, how was it created, which people and systems contributed to it, and what is it being offered to prove?
Authentication Is Not Reliability
A file can be authentic and still contain inaccurate information or support an unreliable conclusion.
Rule 901 requires enough evidence to support a finding that an item is what the proponent claims it is. Depending on the evidence and the nature of the authentication challenge, relevant foundation may include:
- The system, model, and version used
- The operator, inputs, prompts, and settings
- The original output
- Metadata, logs, timestamps, or audit records
- Any later editing or manipulation
- The chain of custody and preservation history
Not every item will be necessary in every case. The required showing depends on what the proponent claims the evidence is, the nature of the authenticity challenge, and the available methods of proof.
Rule 901(b)(9) provides one possible method of authentication by allowing the proponent to describe the process or system and show that it produces an accurate result for the purpose at issue.
Even when the exhibit is authenticated, separate questions may remain about the method’s validity, the accuracy of the particular output, and the reliability of any conclusion drawn from it.
When an expert offers an opinion based on that result, the testimony must also satisfy Rule 702.
Testing AI-Assisted Expert Methodology
An expert should not treat an AI result as reliable simply because software produced it. The expert must understand and evaluate the method well enough to explain the basis for relying on it, its material limitations, and its application to the facts of the case.
Under Rule 702, the proponent must show, more likely than not, that each of the rule’s requirements is satisfied:
- The expert’s specialized knowledge will help the trier of fact understand the evidence or decide a disputed fact
- The testimony is based on sufficient facts or data
- The testimony is the product of reliable principles and methods
- The expert’s opinion reflects a reliable application of those principles and methods to the facts of the case
Naming the software is not enough. Nor is it enough to say that the platform is widely used, commercially available, or considered advanced.
Depending on the method and the disputed issue, relevant questions may include:
- Whether the system was validated for the specific task
- Whether the tested and deployed versions were the same or materially different
- Whether the data were sufficient and reasonably representative for the task
- Whether error rates, thresholds, and limitations are known or meaningful for the method
- Whether the expert independently tested, checked, or otherwise evaluated the result
- Whether the available information permits a qualified person to meaningfully review or evaluate the process
These are context-dependent questions, not mandatory elements in every case. Their importance depends on the type of system, the opinion offered, and the disputed reliability issue.
A model may perform well overall but still produce an inaccurate or unreliable result under the conditions present in a particular case.
For example, a facial-recognition system may perform well in one evaluation but show different error rates when the image quality, capture conditions, or evaluated population differs.
One key question is whether the specific version, data, settings, and conditions used in the case support a reliably derived result, not simply whether the technology performs well in general.
A Common Exclusion Problem
A polished report can hide a weak methodology.
Consider an expert who relies on a proprietary classification model but cannot adequately explain or document:
- The model version and settings
- The inputs and preprocessing
- The validation population and error rates
- The decision threshold
- Whether the expert independently checked, tested, or otherwise evaluated the result
No single missing item necessarily requires exclusion. The court may consider the significance of the missing information, whether the expert can otherwise support the opinion, and whether narrower measures can address the problem.
Preprocessing refers to changes made to the data before the system analyzed it, such as cleaning, normalization, resizing, transformation, selection, or feature extraction. A decision threshold is the cutoff the system used to turn a score into a match, classification, or conclusion.
The report itself may be authentic, and it may even be possible to prove that the software produced it on a specific date. But that does not show that the method was properly tested or reliably applied.
The problem is not simply that the expert used AI. The problem is that the available information may leave the proponent unable to establish, and the court unable to evaluate, whether the conclusion was reliably reached.
This is why early expert vetting matters. Counsel should determine whether the expert can critically evaluate and explain the underlying method or is simply repeating the system’s output.
What Other Evidence Rules Apply?
Rule 703
Rule 703 allows an expert to rely on otherwise inadmissible facts or data when experts in the particular field would reasonably rely on that kind of information when forming an opinion on the subject.
Rule 703 addresses the kinds of facts or data an expert may rely on. Rule 702 separately governs whether the opinion has a sufficient basis and is reliably derived.
That does not make the underlying material automatically admissible. If the material is otherwise inadmissible, the proponent may disclose it to the jury to explain the opinion only when its value in helping the jury evaluate the opinion substantially outweighs its prejudicial effect.
In an AI-related case, counsel should ask whether experts reasonably rely on this kind of output from this kind of system for the purpose at issue. Using an AI tool as one input in a broader analysis may raise different questions from relying on its output as the principal basis for the conclusion, but either approach must be evaluated under Rules 702 and 703.
Rule 801
Machine-generated output does not always fit neatly into traditional hearsay analysis.
Rule 801 defines a statement as a person’s oral or written assertion, or nonverbal conduct intended as an assertion. Information generated and recorded automatically by a machine may fall outside the hearsay rule when it does not contain or communicate an assertion made by a person.
This does not mean that all computer-produced evidence is nonhearsay. The analysis depends on whether the challenged content is an assertion attributable to a person and whether it is offered for its truth.
The analysis may change when the output repeats or incorporates an assertion supplied, labeled, or recorded by a person. Human involvement alone does not make the entire output hearsay, so counsel should identify the particular assertion and determine whether it is offered for its truth.
Relevant questions may arise when the system:
- Reproduces a human statement
- Summarizes or translates human communication
- Incorporates human-entered statements, or labels or classifications intended as assertions
For example, a chatbot log may be generated and stored automatically, but a statement typed by a user and reproduced in that log remains a human assertion. Counsel must separately determine whether that statement is offered for its truth and, if so, whether a hearsay exclusion or exception applies.
Counsel should identify the source of each material assertion rather than treating the entire output as either human or machine generated.
Spoken words reproduced in an automated transcript must be analyzed separately to determine whether they are offered for their truth and whether the hearsay rules permit their admission. The transcript’s accuracy and the admissibility of the speaker’s words are separate questions.
Counsel should also determine whether the transcript itself is being offered as evidence or used only as an aid for reviewing an authenticated recording.
Rule 403
AI-related evidence can appear more certain than it really is.
Rule 403 permits exclusion when the evidence’s probative value is substantially outweighed by dangers such as unfair prejudice, confusing the issues, or misleading the jury.
Rule 403 does not permit exclusion merely because evidence is technical, damaging, or difficult to challenge. The identified danger must substantially outweigh the evidence’s probative value.
Depending on how it is presented, a numerical score, technical visualization, synthetic recording, or computer-generated conclusion may appear more objective or certain than the underlying method supports.
Counsel should consider whether explanatory testimony, limiting instructions, redactions, or a clearer presentation can reduce that risk. When those safeguards are not enough, exclusion may be appropriate.
Preserving and Vetting the Technical Foundation
A final PDF or screenshot may not contain the information needed to examine how a result was produced.
Important details may appear in prompts, logs, system settings, earlier file versions, or records held by a third-party provider.
Records held by a third-party provider may require prompt preservation requests, contractual review, consent, or third-party discovery. Their availability will depend on the governing discovery rules and the party’s legal right or practical ability to obtain them.
Depending on the dispute, counsel should consider preserving:
- Original files and native outputs
- Prompts, inputs, and preprocessing
- Model names, versions, and settings
- Logs, metadata, hashes, and timestamps
- Validation and human-review records
Preservation decisions should remain tied to the claims, defenses, anticipated disputes, and information reasonably within the party’s possession, custody, or control.
Early preservation can make later authentication and reliability review easier. Systems, access settings, and available versions can change over time, which may make the original process harder to examine later.
The expert should also match the actual dispute. Depending on the case, that may require experience in machine learning, digital forensics, biometrics, statistical validation, software engineering, or the field in which the model was used.
During vetting, counsel should ask whether the expert can:
- Explain the system clearly
- Evaluate the data and the adequacy of the validation
- Review versions, settings, and preprocessing
- Reproduce, test, or otherwise evaluate the result when the available information allows it
- Identify missing information, uncertainty, and limitations
A strong expert should explain what can be tested, what cannot, and how those limits affect the opinion.
NIST materials can help counsel ask better questions about validation, documentation, system limits, and risk management. They do not create a federal evidence standard or establish admissibility on their own.
What Is the Status of Proposed Rule 707?
Proposed Federal Rule of Evidence 707 is not currently in effect.
Proposed Rule 707 was published for public comment. The published proposal would have applied Rule 702-style reliability requirements to certain AI-generated evidence offered without supporting expert testimony. The comment period ran from August 15, 2025, through February 16, 2026.
As of the May 17, 2026 Advisory Committee report, the Committee had revised Rule 707 but had not recommended action on it. It planned further study at its Fall 2026 meeting, together with a separate possible Rule 901(c) addressing deepfake authentication.
The published proposal should not be treated as final.
Until any new rule takes effect, courts must analyze these disputes under the existing Federal Rules of Evidence, governing precedent, and any other applicable law.
AI Evidence Checklist for Attorneys
Before offering or challenging AI-related evidence, counsel should ask:
- What is the evidence offered to prove?
- Is authenticity disputed, reliability disputed, or both?
- Which system, model, version, operator, and settings can be identified?
- Are the original inputs, outputs, metadata, and logs preserved?
- What human review or intervention occurred?
- Was the system used within the purpose and conditions for which it was validated?
- Are error rates, thresholds, and limitations known or meaningful for the method?
- Does the output contain or depend on human assertions?
- Is the output itself being offered, or only an opinion based on it?
- Who possesses the relevant technical records, and can they be preserved or obtained?
- Could the presentation mislead the jury?
These questions are most useful before or early in expert discovery, and before deposition or motion practice. They help separate a meaningful technical foundation from a conclusory claim that the evidence is reliable simply because a computer produced it.
