Artificial intelligence is rapidly becoming part of the evidentiary record in civil litigation. Businesses use AI systems to evaluate insurance claims, screen employment applicants, detect fraud, forecast losses, analyze medical images, monitor equipment, recommend prices, summarize communications, and make operational decisions. Parties and experts are also using generative AI to create reports, reconstructions, demonstrative exhibits, transcripts, translations, timelines, and summaries. At the same time, AI tools can fabricate or alter photographs, audio recordings, videos, documents, and electronic communications with a level of realism that may be difficult to detect.
These developments do not mean that courts must abandon traditional evidentiary principles. An AI-generated exhibit is not automatically admissible merely because a computer produced it, and it is not automatically inadmissible merely because artificial intelligence played some role in its creation. The central questions remain familiar: Is the evidence relevant? Is it authentic? Does it contain hearsay? Is it the product of a reliable process? Does it require expert testimony? Does the evidence fairly represent the underlying information? Would its probative value be substantially outweighed by the risk of unfair prejudice, confusion, or misleading the jury? ¹
What has changed is the complexity of answering those questions. AI systems may involve proprietary models, undisclosed training data, continuously updated software, probabilistic outputs, human review, and multiple layers of data processing. A polished AI result may conceal uncertainty, assumptions, or mistakes that would have been obvious in a conventional calculation. Lawyers seeking to admit or exclude AI evidence therefore need to begin building the evidentiary record during discovery, rather than waiting until the eve of trial to determine how the technology works.
The phrase “AI evidence” can describe several materially different forms of proof. In one case, the evidence may be an output generated by an AI system during the ordinary course of business, such as a fraud score, employment recommendation, insurance valuation, medical classification, or equipment-failure alert. In another case, a retained expert may use machine learning as one component of an opinion. A party may also offer an AI-enhanced photograph, computer-generated reconstruction, translated recording, voice identification, or predictive model. Finally, a party may dispute whether apparently conventional evidence is authentic by alleging that a photograph, video, voice recording, email, or document is an AI-generated deepfake.
These distinctions matter because each category presents a different evidentiary problem. An AI-generated business output may require proof that the system operated reliably and that the underlying data were accurate. An expert’s reliance on AI may be evaluated through the standards governing expert testimony. An AI-generated animation may function only as an illustrative aid rather than substantive evidence. A deepfake objection, by contrast, ordinarily concerns authenticity and the allocation of responsibility between the court and the jury.
Counsel should therefore avoid addressing “AI” as though it were a single technology. The foundation for a generative-language-model summary will not necessarily resemble the foundation for a computer-vision classification, an automated damages calculation, or a synthetic video. The proponent must identify what the system actually did, what human beings contributed, what information the system received, and what proposition the resulting evidence is offered to prove.
The Federal Rules of Evidence do not presently contain a special, effective rule governing all artificial-intelligence evidence. The ordinary rules of relevance, authentication, hearsay, expert testimony, original writings, and unfair prejudice continue to apply. The absence of a separate AI rule does not create an evidentiary vacuum. Instead, it requires courts and litigants to apply established principles to unfamiliar technology.
A useful starting point remains the analytical framework courts have long applied to electronically stored information. In Lorraine v. Markel American Insurance Co., the court explained that electronic evidence must be examined for relevance, authenticity, hearsay, compliance with the original-writing rules, and the possibility of exclusion under Rule 403.⁶ Although Lorraine predates modern generative AI, its central lesson is especially important today: electronic evidence is not admissible simply because it appears in a party’s files or is attached to a motion. The proponent must address each applicable evidentiary requirement.
AI evidence can fail at any point in that analysis. A document may be authentic but contain inadmissible hearsay. An AI output may satisfy a business-record exception but still lack a reliable foundation. An expert may be qualified, but the model used by the expert may not reliably apply to the facts of the case. A visually compelling reconstruction may accurately reflect an expert’s assumptions but remain unfairly prejudicial because it appears to portray an actual event rather than a disputed theory.
Federal Rule of Evidence 901 requires the proponent to produce evidence sufficient to support a finding that an item is what the proponent claims it is. ¹ The rule provides several possible methods, including testimony from a witness with knowledge, distinctive characteristics, comparison with authenticated material, voice identification, and evidence describing a process or system and showing that it produces an accurate result.
For conventional electronic evidence, authentication may be established through a participant in the communication, a custodian, a forensic examiner, metadata, account information, internal characteristics, surrounding circumstances, or a combination of those sources. The Sixth Circuit’s decision in United States v. Farrad illustrates that social-media photographs and similar digital evidence cannot be authenticated solely by assuming that content appearing on an account must have been created or posted by the account holder.⁴ The proponent must connect the item to the person, event, or system it is offered to represent.
AI makes that connection more important. A witness may recognize the scene depicted in a photograph, but recognition alone may not establish that the photograph was not materially altered. A manager may testify that a report came from the company’s software, but that testimony may not establish what version of the model produced it, whether the inputs were complete, or whether a human edited the output. A party may authenticate an AI-generated report as a document maintained in its records without proving that the report’s conclusions are accurate.
The foundation should therefore match the proponent’s claim. When a party offers an AI output merely to prove that the output was received and influenced a decision, it may be enough to establish that the decision-maker received and relied on it. When the output is offered to prove that its substantive conclusion was correct, the proponent will ordinarily need a more substantial foundation concerning the system, the data, and the reliability of the result.
Rule 901 specifically permits authentication through evidence describing a process or system and showing that it produces an accurate result. ¹ That provision is particularly relevant when no human being directly observed the fact reflected in the evidence. Examples may include automated sensor readings, algorithmic classifications, computer-generated calculations, fraud alerts, or machine-produced comparisons.
A competent foundation will normally explain the purpose of the system, how information enters it, what processing occurs, what controls exist, how outputs are stored, whether the software was operating properly, and whether the result can be reproduced or independently verified. The witness does not always need to be the original programmer. A sufficiently knowledgeable administrator, engineer, analyst, custodian, or expert may be able to establish the relevant facts. The witness must nevertheless understand enough about the system to provide more than the conclusory assurance that “the computer generated it.”
The Sixth Circuit’s decision in United States v. Ganier is instructive even though it arose in a criminal case and did not involve generative AI. The court concluded that testimony interpreting information produced through forensic computer software involved specialized knowledge and should have been treated as expert testimony.⁵ The decision demonstrates an important boundary: a witness may testify as a fact witness about ordinary observations, but interpreting technical computer-generated results may require qualification and disclosure as an expert.
That boundary becomes especially significant when an employee is offered a sponsoring witness. An employee who regularly reviews AI reports may be able to explain how the company uses them. The employee may not be qualified to testify that the model is scientifically valid, that its error rate is acceptable, or that it reliably applied to a particular individual or event. Those issues may require testimony from a person with technical expertise.
Federal Rules of Evidence 902(13) and 902(14) allow certain electronically generated records and copied electronic data to be authenticated through a certification rather than live testimony. ¹ These provisions can be valuable when the disputed issue concerns whether data were generated by a particular electronic process or whether a forensic copy matches information obtained from a device, storage medium, or file. Hash values and other digital-identification methods may help establish that an electronic file has not changed.
Self-authentication does not resolve every admissibility issue. A certification may establish that the exhibit is an accurate copy of a file retrieved from a server, but it does not necessarily establish who created the file, whether the information in it is true, whether the file contains hearsay, or whether an AI system reliably generated its conclusions. Authentication establishes identity and integrity; it does not automatically establish substantive accuracy.
For AI evidence, counsel should decide whether the contested question concerns the integrity of the electronic file, the identity of its source, the reliability of the process that created it, or all three. A Rule 902 certification may eliminate the need to call a forensic technician merely to prove that two files have matching hash values. It will not ordinarily eliminate the need for a witness or expert who can explain what the AI system did and why its output is relevant.
Generative AI makes it possible to create fabricated photographs, videos, voices, and documents that appear genuine. That capability may lead litigants to challenge authentic evidence by asserting that it “could be” a deepfake. Courts must take genuine manipulation seriously, but they are unlikely to permit a speculative accusation to defeat otherwise adequate authentication.
As of July 2026, the federal Advisory Committee on Evidence Rules has continued to study whether a special rule should govern deepfake challenges. The Committee has recognized that deepfakes can be easy to create and difficult to detect, while also observing that the traditional Rule 901 threshold is relatively low. Its working approach has contemplated requiring the opponent to make an evidentiary showing that an item may be a deepfake before imposing a heightened burden on the proponent. The Committee had not adopted that approach as an effective Federal Rule of Evidence as of its May 2026 report.⁸
Under the existing rules, a meaningful deepfake challenge should identify concrete reasons to question the evidence. Those reasons might include inconsistent metadata, unexplained gaps in custody, visual or audio anomalies, conflicting original files, evidence of editing software, missing source recordings, inconsistent timestamps, or expert analysis suggesting synthetic generation. A general statement that artificial intelligence makes fabrication possible should carry little weight because the same assertion could be made about virtually any digital exhibit.
The proponent should be prepared to preserve and produce the highest-quality source file, associated metadata, device information, transmission history, earlier versions, and testimony from persons who created, received, or maintained the evidence. When the authenticity dispute is technically substantial, the court may need a pretrial evidentiary hearing and expert testimony rather than attempting to resolve the issue through objections made in front of the jury.
Hearsay analysis requires careful attention to the distinction between human assertions and purely machine-generated information. Under the federal definition, a declarant is a person who made a statement. A calculation or measurement generated automatically by a machine may not itself be a human statement. The underlying data, however, may contain human assertions, and the system’s output may repeat or summarize them.
For example, an AI-generated summary of employee complaints may include several layers of potential hearsay. The complaints are human statements. Labels or descriptions added by investigators may be additional statements. The AI-generated summary may combine, paraphrase, or mischaracterize those statements. Even when the final computer output is not independently treated as hearsay, the underlying assertions still require an exclusion from hearsay or an applicable exception if offered for their truth.
The purpose for which the evidence is offered remains critical. An employer defending a discrimination claim may offer an AI recommendation not to prove that the recommendation was objectively correct, but to explain what information the decision-maker received and why the decision was made. In that circumstance, the evidence may be offered for its effect on the recipient rather than for the truth of the model’s conclusion. The court may nevertheless consider whether a limiting instruction is needed and whether the jury is likely to misuse the evidence as proof that the AI determination was accurate.
A business-record foundation also requires precision. The fact that an AI output was routinely stored in a company’s records may support admission under the business-record exception, but routine retention does not validate an unreliable model. The opponent may challenge the method or circumstances of preparation as indicating a lack of trustworthiness. Moreover, human statements incorporated into the AI record may create hearsay within hearsay, requiring a separate basis for each layer.
Many disputes concerning AI evidence will ultimately be governed by Federal Rule of Evidence 702. The current rule requires the proponent to demonstrate to the court that it is more likely than not that the expert’s specialized knowledge will help the factfinder, that the testimony rests on sufficient facts or data, that it results from reliable principles and methods, and that the expert reliably applied those principles and methods to the case. ²
When an expert relies on AI, counsel should not treat the software as an invisible assistant. The expert should be able to explain the role the system played, what inputs were provided, what output was generated, whether the expert independently evaluated the output, and whether the opinion would change if the AI component were removed. The court may also need information about validation, testing, known limitations, error rates, data quality, model drift, version changes, and the similarity between the model’s development environment and the facts of the litigation.
Proprietary status does not excuse the proponent from establishing reliability. A vendor may resist disclosing source code or training data, and the court may use protective orders to address legitimate confidentiality concerns. But a party cannot ordinarily rely on a black-box conclusion while preventing meaningful examination of the basis for that conclusion. If neither the expert nor the vendor can explain how the result was generated or tested, the court may conclude that the opinion has not been reliably supported.
The expert’s own analysis remains indispensable. An expert who merely submits documents to a generative AI system and repeats the resulting answer may have difficulty demonstrating a reliable methodology. A stronger presentation will show that the expert selected appropriate data, tested the output against independent evidence, examined contrary information, assessed uncertainty, and exercised professional judgment rather than deferring to the machine.
The federal Advisory Committee has considered a proposed Rule 707 addressing AI-generated or machine-generated evidence offered without an accompanying expert witness. The proposal was designed to apply reliability concepts resembling Rule 702 when a party seeks to introduce a technically derived conclusion without presenting a human expert who can be examined about it.
The proposal generated substantial public comment, including disagreement about its scope, terminology, interaction with Rule 702, treatment of ordinary electronic processes, and whether an expert should be required. In May 2026, the Advisory Committee reported that it was not recommending action on proposed Rule 707 at that time. Instead, the Committee planned further study and possible coordination with its work on deepfake evidence.⁸ Proposed Rule 707 should therefore not be cited as though it were currently effective federal law.
The proposal is nevertheless useful as an indication of where evidentiary disputes are heading. It reflects concern that parties may attempt to place AI-generated conclusions before juries without satisfying the safeguards that would apply if the same conclusion were offered through an expert. Until a specific rule is adopted, courts can address that problem through Rules 104, 403, 611, 702, 901, and the court’s authority to manage the presentation of evidence.
The original-writing rules apply to electronically stored information, including AI inputs and outputs. ¹ In the electronic context, an “original” can include a printout or other readable output that accurately reflects electronically stored information. A duplicate may be admissible unless a genuine question is raised about authenticity or it would be unfair to admit the duplicate.
Although a screenshot or PDF export may qualify for some purposes, the native material is often strategically important. A screenshot may omit metadata, prompts, system instructions, revisions, attachments, hidden content, or information showing when and how the output was created. A PDF report may not reveal that a user regenerated the answer several times before selecting the most favorable version.
When AI evidence may become significant, preservation should include the input data, prompts, system settings, model and software versions, output files, audit logs, user edits, validation records, and information concerning human review. If the system is continuously updated, preserving only the final output may make it impossible to reproduce the result later. Counsel should work with the client and appropriate technical personnel before ordinary retention practices overwrite relevant logs or model information.
Failure to preserve relevant AI records may implicate the same discovery and spoliation principles that apply to other electronically stored information. Federal Rules of Civil Procedure 26, 34, and 37 provide the primary framework for identifying, requesting, producing, and addressing the loss of ESI. ³
The admissibility record should begin with the Rule 26(f) conference. When a party’s claims or defenses involve automated decision-making, predictive analytics, generative AI, or machine learning, the parties should discuss preservation, custodians, relevant systems, production formats, metadata, proprietary information, and whether technical discovery should proceed in
stages. ³
Requests for production should be tailored to the disputed use of AI. In an employment case involving an automated screening recommendation, relevant discovery may include the criteria used by the system, the applicant data supplied, the resulting score or recommendation, validation materials, override procedures, and communications showing how human decision-makers used the recommendation. In a commercial case involving an AI forecast, relevant information may include the source data, assumptions, model version, confidence intervals, alternative outputs, and records showing whether the forecast was revised.
Discovery should also identify human involvement. Many systems described as “automated” include substantial human judgment. Employees may select the inputs, alter thresholds, choose among outputs, correct classifications, or make the ultimate decision. Conversely, a company may characterize a process as human-reviewed when reviewers routinely approve the machine’s recommendation without meaningful scrutiny. That distinction may affect relevance, causation, reliability, and the identity of the witnesses required at trial.
Proportionality remains important. A demand for an entire source-code repository or all training data may be unduly burdensome when the dispute can be resolved through validation reports, targeted testing, documentation, and testimony. In other cases, source code or detailed technical information may be necessary because the output cannot otherwise be meaningfully examined. Courts can use phased discovery, sampling, neutral experts, confidentiality orders, and restricted review environments to balance evidentiary need against burden and trade-secret concerns.
AI can create sophisticated timelines, animations, accident reconstructions, medical illustrations, and visual summaries. These materials may help a jury understand complex evidence, but their realism can also give disputed assumptions the appearance of established fact.
Federal Rule of Evidence 107, effective December 1, 2024, governs illustrative aids used to help the factfinder understand evidence or argument. The court may allow such an aid when its utility is not substantially outweighed by the danger of unfair prejudice, confusion, misleading the jury, delay, or wasted time. An illustrative aid is not itself evidence and ordinarily should not be provided to the jury during deliberations unless the parties consent or the court finds good cause. The rule also contemplates preserving the aid in the record when practicable.⁷
Counsel should clearly identify whether an AI-generated visual is offered as substantive evidence, a Rule 1006 summary, or a Rule 107 illustrative aid. The categories are not interchangeable. A Rule 1006 summary may be admitted to prove the content of voluminous admissible materials. A Rule 107 aid merely assists the factfinder and does not become evidence. A computer reconstruction offered as substantive proof may require authentication, expert testimony, and a showing that it accurately incorporates the relevant data.
The proponent should disclose the assumptions embedded in the visual and provide the opposing party with sufficient time to inspect it. Labels, camera angles, timing, colors, perspective, facial expressions, and reconstructed movements can subtly influence the jury. AI may also add details not supported by the record. A visual that appears realistic should be examined frame by frame to ensure that it does not portray speculation as fact.
AI evidence can carry a special aura of objectivity. Jurors may assume that a numerical score, computer classification, or algorithmic recommendation is more neutral than human judgment. This tendency, sometimes described as automation bias, can magnify the prejudicial effect of an output that is based on incomplete or unreliable information.
Rule 403 permits exclusion when probative value is substantially outweighed by dangers including unfair prejudice, confusion, and misleading the jury. ¹ Courts may use the rule to limit AI evidence that cannot be adequately explained, that creates a false impression of certainty, or that would require a disproportionate trial within a trial concerning the underlying technology.
The proponent can reduce those concerns by using plain-language explanations, disclosing uncertainty, identifying the system’s limited role, and avoiding unsupported claims that the AI was “objective,” “neutral,” or “scientifically proven.” The opponent should identify the specific way the evidence may mislead rather than relying on generalized distrust of artificial intelligence.
Limiting instructions may help where an AI output is admitted for a restricted purpose. For example, the jury may be instructed that an automated recommendation is admitted to show what a decision-maker received, not to establish that the recommendation was correct. In other circumstances, redaction, explanatory testimony, or exclusion of unnecessary technical details may provide a better solution than complete exclusion.
A motion in limine concerning AI evidence should explain the technology, the purpose for which the evidence is offered, the proposed foundation, and the particular evidentiary rules implicated. Broad arguments that AI is inherently unreliable or inherently trustworthy are unlikely to assist the court.
The proponent should identify each sponsoring witness and the part of the foundation that witness will supply. A custodian may establish recordkeeping practices. A user may explain the inputs and ordinary operation of the system. A forensic witness may establish file integrity. A technical expert may address validation and reliability. A decision-maker may explain how the output affected conduct. Attempting to obtain the entire foundation from a witness who has only superficial knowledge creates avoidable risk.
Where admissibility depends on technical questions, the parties should request a Rule 104 hearing before the evidence is displayed to the jury. A pretrial hearing allows the court to examine model documentation, expert testimony, source files, and competing forensic analyses without exposing jurors to potentially inadmissible material. It also permits the court to determine whether the evidence may be admitted for one purpose but not another.
The opponent should preserve specific objections. An objection that evidence is “AI-generated” does not identify whether the alleged defect concerns authentication, hearsay, expert reliability, original writings, unfair prejudice, discovery violations, or lack of personal knowledge. Separating those grounds allows the court to make a reasoned ruling and protects the record for appeal.
Michigan litigators will encounter many of the same issues under the Michigan Rules of Evidence. Michigan Rule 901 requires evidence sufficient to support a finding that an item is what its proponent claims and recognizes authentication through testimony, distinctive characteristics, voice identification, and proof that a process or system produces an accurate result. Michigan Rule 702 requires the proponent to demonstrate that it is more likely than not that the expert’s knowledge will assist the factfinder, the testimony is based on sufficient facts or data, the principles and methods are reliable, and the expert reliably applied them.⁹
Michigan practitioners should not assume, however, that every federal evidence provision has an identical Michigan counterpart. The current Michigan Rules should be consulted directly, particularly with respect to self-authentication, expert disclosures, learned treatises, summaries, and state procedural requirements. The meaning of “Rule 707” also differs: proposed Federal Rule 707 concerns AI-generated evidence, while Michigan Rule 707 addresses the use of learned treatises for impeachment.
The Michigan Judicial Council’s Generative AI and the Courts Workgroup has recognized that generative AI presents evolving issues for judges, lawyers, court personnel, and litigants. ¹⁰ That institutional attention reinforces the practical need for transparency and competent preparation, but it does not replace the governing Michigan Rules of Evidence or controlling precedent.
Michigan counsel should therefore develop the same foundational record that would be prudent in federal court: preservation of source information, identification of human and automated components, targeted discovery, appropriate expert testimony, and a clear explanation of the purpose for which the evidence is offered.
The admissibility of AI evidence will rarely turn on the label “artificial intelligence” alone. Courts will examine what the system did, what information it used, how the output was preserved, what human beings contributed, whether the process was reliable, and what proposition the evidence is offered to prove.
For proponents, the most persuasive approach is transparency. Preserve the inputs and outputs, identify the model and version, disclose material assumptions, retain qualified witnesses, validate the result, and distinguish substantive evidence from an illustrative aid. For opponents, the strongest challenges will be specific and evidence-based. Examine provenance, metadata, system operation, source data, testing, error rates, human intervention, alternative outputs, and the fit between the technology and the disputed facts.
Artificial intelligence may be new, but the essential task remains familiar. Evidence must be connected to the case, supported by a competent foundation, tested through the adversarial process, and presented in a manner that helps rather than misleads the factfinder. Lawyers who address those requirements during preservation and discovery will be far better positioned than those who first confront the technology in a motion in limine or at the courthouse door.
Contact Tishkoff
Tishkoff PLC specializes in business law and litigation. For inquiries, contact us at www.tish.law/contact/. & check out Tishkoff PLC’s Website (www.Tish.Law/), eBooks (www.Tish.Law/e-books), Blogs (www.Tish.Law/blog) and References (www.Tish.Law/resources).
Sources:
1- Federal Rules of Evidence, Rules 104, 401–403, 801–807, 901–902, and 1001–1008, as amended through December 1, 2024. https://www.uscourts.gov/sites/default/files/2025-02/federal-rules-of-evidence-dec-1-2024_0.pdf
2- Federal Rules of Evidence, Rules 701–705, as amended through December 1, 2024. https://uscode.house.gov/view.xhtml?path=/prelim@title28/title28a/node232&edition=prelim
3- Federal Rules of Civil Procedure, Rules 26, 34, and 37, as amended through December 1, 2025. https://www.uscourts.gov/sites/default/files/document/federal-rules-of-civil-procedure.pdf
4- United States v. Farrad, 895 F.3d 859 (6th Cir. 2018). https://caselaw.findlaw.com/court/us-6th-circuit/1943586.html
5- United States v. Ganier, 468 F.3d 920 (6th Cir. 2006). https://law.justia.com/cases/federal/appellate-courts/F3/468/920/524366/
6- Lorraine v. Markel American Insurance Co., 241 F.R.D. 534 (D. Md. 2007). https://app.minerva26.com/case_law/17344-lorraine-v-markel-am-ins-co
7- Federal Rule of Evidence 107, Illustrative Aids, effective December 1, 2024. https://www.uscourts.gov/sites/default/files/2025-02/federal-rules-of-evidence-dec-1-2024_0.pdf
8- Advisory Committee on Evidence Rules, Report to the Standing Committee, May 17, 2026. https://www.uscourts.gov/forms-rules/records-rules-committees/committee-reports/advisory-committee-evidence-rules-may-2026
9- Michigan Rules of Evidence, Rules 104, 401–403, 702, 801–807, 901–902, and 1001–1008, updated through Michigan Supreme Court orders effective January 28, 2026. https://www.courts.michigan.gov/492ca5/siteassets/rules-instructions-administrative-orders/rules-of-evidence/michigan-rules-of-evidence.pdf
10- Michigan Judicial Council, Generative Artificial Intelligence and the Courts Workgroup Report, 2024. https://www.courts.michigan.gov/administration/special-initiatives/mjc/
This publication is for general informational purposes and does not constitute legal advice. Reading it does not create an attorney-client relationship. You should consult counsel for advice on your specific circumstances