Artificial intelligence is no longer a futuristic talking point in Michigan litigation. It is increasingly present in what parties produce during discovery, what lawyers rely on to prepare arguments, and what jurors may be asked to evaluate in the courtroom. That shift is not limited to the obvious headline problem of deepfakes. It also includes quieter, more routine questions that can become outcome-determinative: whether an AI-generated image is authentic, whether an AI “enhanced” video is a fair and accurate depiction, whether a chatbot’s output is hearsay, whether an algorithm’s score is an expert opinion, and whether the process that produced an AI output can be explained well enough to satisfy Michigan’s evidence rules. In practice, Michigan judges are being asked to apply familiar doctrines relevance, authentication, hearsay, best-evidence principles, and expert reliability to unfamiliar artifacts generated or transformed by models that may be opaque even to their designers.
Michigan’s Rules of Evidence provide a workable framework for these disputes, but they do not eliminate the need for careful lawyering and fact development. The central lesson emerging from early disputes around synthetic media and AI outputs is that admissibility fights are moving “upstream.” Instead of saving credibility attacks for cross-examination, parties increasingly litigate the integrity of digital evidence before trial, because the risk of unfair prejudice, confusion, and false confidence is materially higher when images, audio, and video can be fabricated or altered with consumer-grade tools. Michigan’s evidentiary system is built to handle contested proof, yet AI raises the stakes by collapsing the line between real and manufactured content in a way that can be persuasive to jurors even when it is wrong.
In Michigan, the threshold question is still whether a piece of proof is relevant and admissible under the ordinary structure of the Rules of Evidence, including the court’s authority to make preliminary determinations about admissibility. Michigan’s evidentiary rules are designed to promote fair proceedings and the development of evidence law toward truth-seeking, and that design matters when courts face technology-driven uncertainty.¹ The practical consequence is that AI evidence typically does not require courts to invent brand-new legal tests; rather, courts apply the same foundational requirements with heightened attention to the possibility of fabrication, alteration, and overstatement.
When AI touches evidence, the most common pressure points are authenticity, accuracy, and explanatory adequacy. Authenticity asks whether the item is what the proponent claims it is. Accuracy asks whether the item fairly depicts what it purports to depict, particularly when an image or recording is offered as a representation of a real-world event. Explanatory adequacy asks whether the proponent can explain the process that produced the item well enough for the court to evaluate reliability and for the opposing party to meaningfully test it. In an AI context, those three questions may collapse into one dispute, because a party challenging authenticity will often argue that the process is unknown or untrustworthy, while the proponent will argue that process evidence, metadata, or witness testimony can adequately support a finding of genuineness.
For many AI-related exhibits, the decisive step is authentication under Michigan Rule of Evidence 901. Michigan’s bench book guidance states the core proposition succinctly: evidence must be authenticated and identified before admission, and the proponent must produce evidence sufficient to support a finding that the item is what the proponent claims it is.² This is not a high bar in theory, because authentication is often described as a “sufficient to support a finding” standard rather than proof beyond doubt. In an AI era, however, the amount and type of foundation that judges expect can increase, not because the legal rule changes, but because the factual risk profile changes.
Deepfakes make authentication disputes less abstract. When a party offers a video or audio clip and the opponent claims it is synthetic or altered, the court has to decide what foundational showing is enough to permit the jury to see it. Traditional pathways witness recognition testimony, testimony about recording devices, chain of custody, metadata, and corroboration by surrounding circumstances still matter, but courts and litigants increasingly treat them as cumulative safeguards rather than as a single decisive factor. A witness saying “that looks like what I saw” may be less persuasive if the witness could be mistaken, if lighting or distance was poor, or if the exhibit is a copy with unknown provenance. The more consequential the exhibit is, the more likely courts will insist on a stronger foundation, not as a formal heightened standard, but as a practical response to the danger that jurors will give the exhibit undue weight.
This is where Michigan’s flexibility becomes important. Authentication can be built in multiple ways, and AI disputes encourage litigants to use multiple layers at once. A proponent who can present the original file from the capturing device, with intact metadata, with testimony from the person who captured it, and with evidence of preservation and handling, will be in a far stronger position than a proponent who offers a downloaded clip with no provenance. Conversely, the opponent who can show gaps in chain of custody, unexplained conversions, signs of compression artifacts inconsistent with the claimed device, or the absence of expected metadata may create enough doubt to support exclusion, limitation, or at least a pretrial hearing focused on authenticity.
Parties sometimes hope that self-authentication concepts will solve the AI evidence problem by allowing records to come in through certifications rather than live testimony. In the federal system, Rule 902 contains pathways for certifying records generated by electronic processes or systems and for certifying data copied from electronic devices, media, or files.³ Those provisions reduce the cost of proving that a record is what it purports to be, but they do not guarantee admissibility when the real dispute is whether the content has been manipulated or fabricated. A certification may establish that a file was produced from a system in a particular manner yet leave open the question of whether the content accurately reflects reality or whether it has been altered before entering that system.
Michigan practitioners should therefore distinguish between two categories of AI-related proof. One category involves routine electronic records that may be created, stored, or filtered through software that uses AI in the background, such as email platform features, document categorization tools, or automated transcription. Those can often be authenticated through conventional business-record and system testimony, along with the kinds of certifications that are increasingly common for electronic evidence. The other category involves synthetic or transformed media deepfakes, AI-generated images, or AI-enhanced recordings where the critical question is not merely whether the file came from a particular source, but whether the content is an accurate depiction. The second category tends to require more robust foundational proof and, in many cases, expert assistance.
Even when an item can be authenticated, Michigan courts still have to decide whether it should be admitted in the form offered. This is where familiar balancing principles matter. Visual and audio evidence can be highly persuasive, and synthetic media raises the risk that the persuasive force will exceed the exhibit’s actual reliability. Courts are likely to scrutinize whether an AI-enhanced or AI-generated exhibit invites the jury to treat an interpretation as though it were an objective recording.
Consider the difference between clarifying a video’s brightness so that the jury can see what the camera captured, and using AI to “fill in” missing pixels or to reconstruct details that were not actually recorded. The former is closer to a traditional enhancement that may be explained and tested; the latter may embed assumptions and model-driven guesses into the exhibit. When the technology is doing more than formatting or clarifying, the evidence begins to look less like a photograph and more like a demonstrative reconstruction. That does not necessarily mean it must be excluded, but it changes how courts and counsel should treat it. If a party is using AI to illustrate a theory rather than to present an original depiction, courts may require the exhibit to be clearly framed as demonstrative and may require limiting instructions or testimony explaining what the AI did and did not do.
A common mistake in early AI litigation is focusing so heavily on authenticity that lawyers ignore hearsay. A chatbot’s generated response may be offered for many different purposes, and admissibility can turn on whether the proponent is offering it for the truth of what it asserts. If an AI output is offered as proof that a fact is true, it is likely to trigger hearsay objections unless it falls within an exception or is not treated as a “statement” of a declarant under the relevant definitions. If, by contrast, the AI output is offered to show notice, effect on the listener, the reasonableness of an investigation, or a party’s state of mind, the hearsay problem may be less severe, but the proponent must be clear about the purpose.
The hearsay analysis can become complicated when AI is embedded in a business process. If a company uses an AI tool to summarize customer complaints, triage them, or generate risk scores, the resulting record may look like a business record, yet it may contain layers of assertions: the customer’s complaint, the system’s categorization, and the system’s inferred conclusion. A party seeking to introduce the record should be prepared to explain which layer is being offered for what purpose, and an opponent should be prepared to argue that AI-generated conclusions are not simply passive recordkeeping but constitute interpretive assertions requiring scrutiny.
Many AI disputes will turn on whether an expert is needed and, if so, whether the expert’s testimony is reliable and properly tied to the case. Michigan’s recent amendment to MRE 702 effective May 1, 2024, aligns Michigan’s expert rule with the federal approach and emphasizes the court’s gatekeeping function. ⁴ In practical terms, that alignment matters for AI evidence because litigants increasingly rely on experts to explain model behavior, digital forensic indicators, metadata interpretation, and the limits of detection tools. It also matters because courts are more likely to require a clear showing that the expert’s methods are reliable and that the expert has reliably applied them to the facts.
Deepfake disputes illustrate the point. A party alleging that a video is synthetic might retain a forensic analyst who uses artifact detection, compression analysis, lighting inconsistency evaluation, or provenance verification. The opposing party might counter with its own expert, challenge the detection method’s error rates, and argue that the analyst cannot reliably distinguish between deepfake artifacts and ordinary compression noise. Under modern gatekeeping, courts may hold hearings that look very much like reliability hearings in other technical domains, and the judge’s willingness to require rigor will shape outcomes. Michigan’s emphasis on a meaningful gatekeeping role, reflected in the post-amendment posture of Rule 702, increases the likelihood that courts will demand more than conclusory “trust me” assertions when the evidence rests on complex technology. ⁴
The same dynamic applies when a party offers an AI system’s output as substantive proof, such as an algorithmic risk score, an automated identification, or an AI-generated “match” between two datasets. When the output depends on a model’s internal parameters and training data, expert testimony often becomes the only way to explain the system and to address reliability. In these disputes, discovery about the system its training, testing, validation, update history, and known error rates can become as important as the evidence itself.
A recurring theme across jurisdictions is that courts are less interested in AI as a buzzword and more interested in whether the proponent can explain the process. If the evidence is generated by a system, judges typically want to know what the system does, how it does it, what inputs it used, and how the proponent can show that the system worked properly in this instance. That approach is consistent with how courts have treated other computer-generated evidence for decades, but AI complicates it because models may be probabilistic, updated frequently, and nontransparent.
For Michigan litigants, this means that “evidence about the evidence” will be more valuable than ever. When a party wants to introduce an AI-generated image, it may need to show the prompt, the model used, the date and time, the settings, the seeds or version information if available, and any post-generation editing steps. When a party wants to introduce an AI-enhanced recording, it may need to preserve the original recording, document every transformation, and be able to reproduce the enhancement process. When a party wants to rely on a chatbot transcript, it may need to preserve the full conversation context, including system prompts or hidden context that could affect the output, and not just a screenshot of the final response.
This emphasis on documentation changes litigation behavior. It rewards parties who adopt evidence-preservation habits early and punishes those who treat AI tools casually. If a client uses AI to generate content relevant to the dispute, the safest course is to preserve not only the final content but also the provenance data that can prove what happened. If a client uses AI to modify media that might later be evidence, the safest course is to preserve the original, preserve the modified version, and preserve a clear record of the tool, version, and settings used.
Deepfakes create a two-sided problem for courts. One side is the risk that fake evidence will be admitted and believed. The other is the risk that real evidence will be discounted because a party claims it could be fake. Commentary on synthetic media has highlighted how the mere possibility of deepfakes can be weaponized to create doubt, especially in high-stakes cases where a party benefits from undermining trust in recordings. ⁵ In the courtroom, that dynamic can distort ordinary evidentiary reasoning. If courts treat every recording as presumptively suspect, litigation becomes unmanageable; if courts treat recordings as presumptively trustworthy, deepfakes can exploit that trust.
Michigan’s evidentiary structure can manage this tension if courts keep two principles in focus. First, authenticity is not certainty; it is a threshold showing sufficient for a reasonable juror to find genuineness. Second, the court’s role includes preventing evidence from misleading the jury when its probative value is outweighed by the risks of unfair prejudice or confusion. Those principles allow courts to avoid both extremes by calibrating the foundation required for the circumstances of the case. A party raising a deepfake objection should be encouraged to articulate a concrete basis for the challenge provenance gaps, inconsistencies, or forensic indicators rather than relying on speculation. A party offering a recording should be encouraged to present provenance, metadata, and corroboration rather than relying solely on “it looks real.”
AI evidence disputes often start as discovery disputes. Parties will seek model logs, audit trails, metadata, version histories, and internal documentation. Opponents may argue that such discovery is burdensome, proprietary, or irrelevant. Courts then have to manage proportionality, confidentiality, and the practical need for an opponent to test reliability.
Michigan’s broader discovery framework is already oriented toward judicial management and proportionality, and Michigan benchbook materials emphasize the court’s ability to control scope and amount of discovery consistent with the Michigan Court Rules. That control becomes important when AI systems are involved because the “how it works” evidence can be extensive. The practical solution in many cases will be staged discovery that targets the most probative reliability information first, paired with protective orders that allow technical disclosure without unnecessary public dissemination. Courts may also prefer focused evidentiary hearings where parties present competing technical accounts and the judge can decide whether the evidence will be admitted, limited, or excluded.
When AI-generated evidence is offered, Michigan courts will typically want clarity about what the exhibit is. If the exhibit is purely synthetic an image, video, or audio clip that depicts an event that never happened its admissibility as substantive evidence should be rare, and its admissibility as demonstrative evidence should be tightly controlled and clearly labeled. If the exhibit is AI-generated text, such as a summary, timeline, or narrative, the court will likely ask whether it is being offered for its truth or merely as a tool to organize other admissible proof. If it is being offered as a substitute for underlying documents, courts may insist on the underlying sources or may treat the AI output as argument rather than evidence.
If the exhibit is AI-assisted rather than AI-invented, the admissibility question often turns on whether the assistance is more like formatting or more like inference. A transcription tool that converts speech to text is not neutral if it makes systematic errors, but its process can often be tested by comparing the transcript to the audio. An enhancement tool that denoises audio or stabilizes video may be acceptable if the original is preserved and the enhancement does not add content. A reconstruction tool that fills gaps is closer to an inference machine and is more likely to require expert testimony, more likely to trigger prejudice concerns, and more likely to be limited to demonstrative use.
In each of these scenarios, Michigan’s authentication requirement provides a common structure: the proponent must show what the item is, how it was produced, and why it can be treated as what it purports to be.² When the process is complex, MRE 702 becomes the gateway, with the court expecting a reliable explanation and a reliable application.⁴ And when the exhibit risks misleading the jury, traditional balancing principles provide the safety valve, allowing admission with limitations, redactions, or instructions, or exclusion when necessary to prevent unfairness.
Across the country, reporting and commentary on deepfakes and AI evidence points toward a consistent judicial instinct: courts are not eager to rewrite evidentiary rules, but they are increasingly willing to demand stronger foundations and clearer explanations when synthetic media is at issue. ⁵ That instinct will likely shape Michigan practice as well. Michigan judges already have doctrinal tools to manage these issues, and the recent sharpening of expert gatekeeping underscores that Michigan’s judiciary expects rigor where technical proof is offered. ⁴
For litigants, the most important adaptation is strategic rather than doctrinal. Evidence preservation should be approached as provenance preservation. Exhibits should be framed with precision so that the judge and jury understand whether they are seeing an original depiction, a clarified depiction, or an AI-assisted interpretation. Experts should be selected not merely for credentials, but for the ability to explain methods in a way that a judge can evaluate and that an opponent can test. And when deepfakes are alleged, parties should expect courts to prefer concrete, technically grounded challenges over generalized skepticism.
Michigan courts have handled technological shifts before, from digital photos to social media and body camera footage. AI is different not because it makes the rules obsolete, but because it changes how easily reality can be simulated and how confidently jurors may believe what they see and hear. The law’s response, at least in the near term, will be less about creating AI-specific doctrines and more about insisting that parties do the foundational work that the existing rules already require.
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).
Footnoted Sources:
- Michigan Supreme Court, Michigan Rules of Evidence (official PDF published by Michigan Courts, current as posted online; includes rules governing relevance, authentication, and general evidentiary principles). www.courts.michigan.gov/490404/siteassets/rules-instructions-administrative-orders/rules-of-evidence/michigan-rules-of-evidence.pdf
- Michigan Courts, Michigan Judicial Institute Criminal Proceedings Benchbook, “Admission of Physical Evidence” (discussion of MRE 901 authentication standard and foundational requirements). https://www.courts.michigan.gov/4a4b91/siteassets/publications/benchbooks/csbb/csbbresponsivehtml5.zip/CSBB/Ch_9_Evidence/Admission_of_Physical_Evidence.htm?rhtocid=_0_0_13_1
- Legal Information Institute, Cornell Law School, Federal Rule of Evidence 902: Evidence That Is Self-Authenticating (includes provisions on certified records generated by electronic processes or systems). https://www.law.cornell.edu/rules/fre/rule_902
- Michigan Supreme Court Order, ADM File No. 2022-30, “Amendments of Rules 702 and 804 of the Michigan Rules of Evidence,” adopted March 27, 2024, effective May 1, 2024 (amending MRE 702 and reinforcing gatekeeping requirements). www.courts.michigan.gov/48d1c1/siteassets/case-documents/briefs/msc/2022-2023/163120/163120_69_ac_brf-lcj.pdf
- Thomson Reuters, “Deepfakes on trial: How judges are navigating AI evidence” (analysis of authentication challenges, judicial skepticism, and courtroom treatment of deepfake-related claims). https://www.thomsonreuters.com/en-us/posts/ai-in-courts/deepfakes-evidence-authentication/#:~:text=The%20current%20legal%20framework%20for,familiar%20with%20the%20speaker’s%20voice.
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.