Loading
Expert Witness Insights by Area of Expertise
8 min read
Areas of Expertise :
Artificial Intelligence, Digital Forensics, Data Security, Computer Science, Software Engineering
Published Date:
Last Modified Date:
Artificial intelligence in litigation can create expert witness issues involving reliability, transparency, authentication, validation, and the use of machine-generated evidence. This article explains how Rules 702 and 901 apply to AI-assisted expert work, the risks of hallucinations and deepfakes, proposed Rule 707, and practical steps for creating verifiable and defensible AI-supported opinions.
Artificial intelligence is becoming harder to ignore in litigation. In many matters, it shows up in expert workflows, in exhibits, and in the way teams process large volumes of digital material. Used well, AI can help an expert organize information, test scenarios, and explain complex material more clearly. Used poorly, it can produce outputs that look persuasive but are difficult to explain, difficult to reproduce, or harder to defend under the rules that govern expert testimony.
This guide focuses on what matters most when AI touches expert work. For litigators, the question is not only whether AI use is permitted under court orders, ethics duties, and confidentiality obligations. The bigger question is whether an AI-assisted opinion can survive admissibility review, and whether the record shows exactly how the opinion was formed, checked, and applied to the dispute. As a practical matter, if AI plays a meaningful role in the expert’s work, it is often explored in deposition.
Federal Rule of Evidence 702 sets the core admissibility requirements for expert testimony in federal court. To be admitted, the opinion has to rest on an adequate factual foundation, use a reliable method, and apply that method reliably to the facts at issue. The opinion also needs to help the trier of fact on an issue in the case.
Rule 702 was amended in 2023, and those changes matter in real litigation. Admissibility is decided by the judge under Rule 104(a). And the party offering the expert must show it is more likely than not that Rule 702’s requirements are satisfied.
In AI-related disputes, that usually means more than a confident conclusion and a polished report. The expert should be able to walk through the workflow step by step, including what went in, what came out, what checks were performed, and where independent expert judgment began and ended.
Courts generally evaluate AI-assisted opinions through the usual Rule 702 lens. The question is whether the expert opinion is reliable, explainable, and tied to the case in a way the court can assess.
In practice, these are common reliability and application issues that show up in depositions and Rule 702 challenges. They are not a universal checklist. What matters most depends on the method, the data, and the case question.
A useful pressure test is this: could a neutral reviewer rerun the process, when feasible, and reach materially similar results? If the answer is no, the reliability fight often gets harder. If exact reruns are not possible because of closed systems or nondeterministic tools, the expert should document why, and then document the stability checks used instead.
Digital Evidence
Electronic Authentication
Evidence Analysis
Computer Data
Data Analysis
Data Mining Algorithm
Machine Bias
Software Audit
Litigation Support
Artificial intelligence is becoming harder to ignore in litigation. In many matters, it shows up in expert workflows, in exhibits, and in the way teams process large volumes of digital material. Used well, AI can help an expert organize information, test scenarios, and explain complex material more clearly. Used poorly, it can produce outputs that look persuasive but are difficult to explain, difficult to reproduce, or harder to defend under the rules that govern expert testimony.
This guide focuses on what matters most when AI touches expert work. For litigators, the question is not only whether AI use is permitted under court orders, ethics duties, and confidentiality obligations. The bigger question is whether an AI-assisted opinion can survive admissibility review, and whether the record shows exactly how the opinion was formed, checked, and applied to the dispute. As a practical matter, if AI plays a meaningful role in the expert’s work, it is often explored in deposition.
Federal Rule of Evidence 702 sets the core admissibility requirements for expert testimony in federal court. To be admitted, the opinion has to rest on an adequate factual foundation, use a reliable method, and apply that method reliably to the facts at issue. The opinion also needs to help the trier of fact on an issue in the case.
Rule 702 was amended in 2023, and those changes matter in real litigation. Admissibility is decided by the judge under Rule 104(a). And the party offering the expert must show it is more likely than not that Rule 702’s requirements are satisfied.
In AI-related disputes, that usually means more than a confident conclusion and a polished report. The expert should be able to walk through the workflow step by step, including what went in, what came out, what checks were performed, and where independent expert judgment began and ended.
Courts generally evaluate AI-assisted opinions through the usual Rule 702 lens. The question is whether the expert opinion is reliable, explainable, and tied to the case in a way the court can assess.
In practice, these are common reliability and application issues that show up in depositions and Rule 702 challenges. They are not a universal checklist. What matters most depends on the method, the data, and the case question.
A useful pressure test is this: could a neutral reviewer rerun the process, when feasible, and reach materially similar results? If the answer is no, the reliability fight often gets harder. If exact reruns are not possible because of closed systems or nondeterministic tools, the expert should document why, and then document the stability checks used instead.
Even when the underlying analysis is strong, credibility can collapse if a filing includes unverified citations or unsupported factual claims. At least one court has sanctioned lawyers for filing materials that included fabricated case citations generated by an AI tool. One well-known example is Mata v. Avianca, Inc.
AI can assist with research and drafting, but it cannot replace verification. In Mata, the court described the submission of non-existent opinions and fake citations as a serious breakdown in accuracy obligations. When citations fall apart, it invites scrutiny of everything else that was submitted.
For expert work, the takeaway is similar. If AI is involved at any step, someone has to verify the authorities and quotes being relied on, and someone has to be able to explain how the work was checked before it was put in front of the court.
Reliability is one fight. Authenticity is another. AI tools can make digital items look real when they are not, including altered images, synthetic documents, and deepfake audio or video.
Rule 901 sets a threshold for authentication. The proponent has to offer enough evidence so a factfinder could conclude the exhibit is what it is claimed to be. It is a foundation the proponent must clear, not a guarantee that authenticity cannot be challenged.
In practice, common ways parties try to meet Rule 901’s foundation include custody evidence, metadata, and process-based proof. In disputes involving suspected manipulation, parties often focus on foundations like:
If the proponent cannot explain provenance and preservation, and cannot offer a foundation that supports authenticity, the evidence becomes easier to challenge.
Questions also arise when a party tries to introduce a machine-generated output without putting an expert behind it. Proposed Federal Rule of Evidence 707 has been discussed as a way to address machine-generated evidence offered without a human expert witness, particularly when the output functions like an opinion about a disputed fact.
This is a proposal published for public comment, not a rule currently in effect.
Rule 707 is aimed at situations where machine output is offered without an expert, but the output would be treated like expert testimony if a human witness offered the same inference.
The practical concern is straightforward. If an AI output comes in through a lay witness, the opposing party may have less ability to test reliability through a qualified witness who can explain validation, limits, and failure modes. There may be no expert to defend the methodology through focused cross-examination.
The proposal is out for public comment, and the committee has emphasized that publication is not a presumption the rule will be enacted.
As a practical matter, when software is used to draw inferences or make predictions from data, the reliability questions start to look like expert testimony questions. If that output is offered to prove a disputed fact, parties should expect Rule 702-style challenges even if the evidence is framed as a simple “system result.”
Outside the U.S., one example of a court-issued approach is the Supreme Court of New South Wales Practice Note SC Gen 23. It was issued on January 28, 2025 and it commenced on February 3, 2025. It is not U.S. authority, but it is a useful reference point for how a court can approach disclosure and risk control when generative AI is used in connection with court documents, including evidence-related materials.
When Gen AI is used to prepare written submissions or argument summaries, the author must verify that citations and authorities exist, are accurate, and are relevant. The Practice Note also makes clear that verification cannot be done using Gen AI alone.
The Practice Note also addresses expert reports. It states that Gen AI must not be used to draft expert report content without prior leave. If leave is granted, it requires disclosure about what relied on Gen AI and record-keeping steps, including keeping a record of prompts and settings used, unless the court dispenses with that requirement.
Even if U.S. courts do not adopt the same approach, the takeaway is still practical. When AI is used to draft, summarize, or support evidence-related material, undisclosed use and unverifiable sources can create credibility problems fast.
If your goal is admissible, defensible expert work, “transparent and reproducible” is a strong north star. These practices tend to help when the record is contested:
A simple way to keep the work court-ready is to write a short workflow memo that links each step to the opinion you ultimately offer. At minimum, consider preserving: (1) the tool and version, (2) the exact inputs or prompts used, (3) the outputs you relied on, (4) the checks performed, and (5) a clear explanation of how independent expert judgment was applied. If exact reruns are not possible, document why and show the stability checks you used instead.
Can AI be an expert witness in court?
AI tools do not appear in court as witnesses. In practice, testimony is given by human witnesses who can be examined, and any expert opinion must satisfy Rule 702. An expert may use AI tools as part of their work, but a human expert must sponsor and defend the opinion.
What is proposed Rule 707 trying to address?
It is aimed at machine-generated outputs introduced without a human expert, where reliability scrutiny could be sidestepped. The focus is less on the label “AI” and more on preventing opinion-like outputs from being treated as self-proving facts without meaningful testing. The proposal is still under consideration and is not currently in effect.
How do deepfakes change authentication disputes?
They raise the stakes for provenance, metadata preservation, and reliable foundations for authenticity. In practical terms, parties often lean more heavily on chain of custody and reproducible forensic methods where appropriate.
What is the biggest risk when AI is used in expert work?
Lack of verifiability. Missing inputs, unclear workflow steps, weak validation, and unverified citations can undermine reliability and credibility quickly. Once the court doubts the foundation, everything else becomes harder to defend.
Even when the underlying analysis is strong, credibility can collapse if a filing includes unverified citations or unsupported factual claims. At least one court has sanctioned lawyers for filing materials that included fabricated case citations generated by an AI tool. One well-known example is Mata v. Avianca, Inc.
AI can assist with research and drafting, but it cannot replace verification. In Mata, the court described the submission of non-existent opinions and fake citations as a serious breakdown in accuracy obligations. When citations fall apart, it invites scrutiny of everything else that was submitted.
For expert work, the takeaway is similar. If AI is involved at any step, someone has to verify the authorities and quotes being relied on, and someone has to be able to explain how the work was checked before it was put in front of the court.
Reliability is one fight. Authenticity is another. AI tools can make digital items look real when they are not, including altered images, synthetic documents, and deepfake audio or video.
Rule 901 sets a threshold for authentication. The proponent has to offer enough evidence so a factfinder could conclude the exhibit is what it is claimed to be. It is a foundation the proponent must clear, not a guarantee that authenticity cannot be challenged.
In practice, common ways parties try to meet Rule 901’s foundation include custody evidence, metadata, and process-based proof. In disputes involving suspected manipulation, parties often focus on foundations like:
If the proponent cannot explain provenance and preservation, and cannot offer a foundation that supports authenticity, the evidence becomes easier to challenge.
Questions also arise when a party tries to introduce a machine-generated output without putting an expert behind it. Proposed Federal Rule of Evidence 707 has been discussed as a way to address machine-generated evidence offered without a human expert witness, particularly when the output functions like an opinion about a disputed fact.
This is a proposal published for public comment, not a rule currently in effect.
Rule 707 is aimed at situations where machine output is offered without an expert, but the output would be treated like expert testimony if a human witness offered the same inference.
The practical concern is straightforward. If an AI output comes in through a lay witness, the opposing party may have less ability to test reliability through a qualified witness who can explain validation, limits, and failure modes. There may be no expert to defend the methodology through focused cross-examination.
The proposal is out for public comment, and the committee has emphasized that publication is not a presumption the rule will be enacted.
As a practical matter, when software is used to draw inferences or make predictions from data, the reliability questions start to look like expert testimony questions. If that output is offered to prove a disputed fact, parties should expect Rule 702-style challenges even if the evidence is framed as a simple “system result.”
Outside the U.S., one example of a court-issued approach is the Supreme Court of New South Wales Practice Note SC Gen 23. It was issued on January 28, 2025 and it commenced on February 3, 2025. It is not U.S. authority, but it is a useful reference point for how a court can approach disclosure and risk control when generative AI is used in connection with court documents, including evidence-related materials.
When Gen AI is used to prepare written submissions or argument summaries, the author must verify that citations and authorities exist, are accurate, and are relevant. The Practice Note also makes clear that verification cannot be done using Gen AI alone.
The Practice Note also addresses expert reports. It states that Gen AI must not be used to draft expert report content without prior leave. If leave is granted, it requires disclosure about what relied on Gen AI and record-keeping steps, including keeping a record of prompts and settings used, unless the court dispenses with that requirement.
Even if U.S. courts do not adopt the same approach, the takeaway is still practical. When AI is used to draft, summarize, or support evidence-related material, undisclosed use and unverifiable sources can create credibility problems fast.
If your goal is admissible, defensible expert work, “transparent and reproducible” is a strong north star. These practices tend to help when the record is contested:
A simple way to keep the work court-ready is to write a short workflow memo that links each step to the opinion you ultimately offer. At minimum, consider preserving: (1) the tool and version, (2) the exact inputs or prompts used, (3) the outputs you relied on, (4) the checks performed, and (5) a clear explanation of how independent expert judgment was applied. If exact reruns are not possible, document why and show the stability checks you used instead.
Can AI be an expert witness in court?
AI tools do not appear in court as witnesses. In practice, testimony is given by human witnesses who can be examined, and any expert opinion must satisfy Rule 702. An expert may use AI tools as part of their work, but a human expert must sponsor and defend the opinion.
What is proposed Rule 707 trying to address?
It is aimed at machine-generated outputs introduced without a human expert, where reliability scrutiny could be sidestepped. The focus is less on the label “AI” and more on preventing opinion-like outputs from being treated as self-proving facts without meaningful testing. The proposal is still under consideration and is not currently in effect.
How do deepfakes change authentication disputes?
They raise the stakes for provenance, metadata preservation, and reliable foundations for authenticity. In practical terms, parties often lean more heavily on chain of custody and reproducible forensic methods where appropriate.
What is the biggest risk when AI is used in expert work?
Lack of verifiability. Missing inputs, unclear workflow steps, weak validation, and unverified citations can undermine reliability and credibility quickly. Once the court doubts the foundation, everything else becomes harder to defend.
Digital Evidence
Electronic Authentication
Evidence Analysis
Computer Data
Data Analysis
Data Mining Algorithm
Machine Bias
Software Audit
Litigation Support
Expert Witness Insights by Area of Expertise
Expert Witness Insights by Area of Expertise