Artificial intelligence is rapidly entering the practice of law, and for good reason. Attorneys now routinely use generative AI tools to summarize records, analyze depositions, review scientific literature, and draft legal arguments, and these tools deliver real value. As these technologies become more capable, a natural question arises for expert witnesses: how should experts use artificial intelligence thoughtfully and ethically in forming their opinions?
The most widely discussed risks, such as fabricated case citations, hallucinated authorities, and AI-generated filings submitted without verification, are very real, but they are also the most visible. Experts and courts can identify and correct for obvious errors. Subtler problems are potentially more consequential.
The more important question is not whether expert witnesses should or will use artificial intelligence. Used well, AI is a genuinely powerful tool, and experts who engage with it thoughtfully will have real advantages. However, we must consider whether AI shapes an expert's reasoning process in ways the expert does not fully recognize, and whether experts have the methodological discipline to manage that risk.
This is not an entirely new problem. Expert witnesses have always operated in environments that can subtly shape their reasoning before independent analysis begins. Colleagues offer informal opinions. Templates from prior reports impose familiar structures. The diagnostic categories of the DSM organize clinical thinking in ways that are easy to take for granted. AI does not introduce a categorically different risk so much as it introduces a more powerful and less visible version of a familiar one.
Medicine and other technical fields have long studied how computerized systems influence professional decision-making. Studies of clinical decision-support systems have shown that physicians may accept incorrect recommendations generated by automated systems, even when the available clinical evidence points elsewhere, a phenomenon commonly described as automation bias.[1]
Notably, providing clinicians with AI-generated explanations designed to counteract bias did not reliably improve decision-making.[2] The anchoring effect persisted even when clinicians were made aware of it.
Automation bias has since been documented across several domains including aviation, military command systems, and healthcare decision-making.[3] The common feature is that automated systems influence how humans reason about problems, even when the human decision-maker remains formally responsible for the final judgment.
Generative AI introduces a similar dynamic. Modern AI systems produce confident, fluent, and persuasive explanations. When those explanations appear early in an analytical process, they may shape how an expert frames the problem, what alternative explanations are considered, and how conflicting evidence is weighed. The resulting opinion may appear entirely the expert's own. The reasoning process that produced it may not be.
Courts are beginning to look beyond expert conclusions to the process by which those conclusions were formed. Legal commentators have surveyed several cases in which courts examined the role of artificial intelligence in expert analysis, finding that scrutiny increasingly extends to methodology, not just results.[4]
This reflects longstanding principles governing expert testimony. Under Federal Rule of Evidence 702, opinions must reflect a reliable application of professional knowledge and methodology. Courts evaluate not only what an expert concludes, but how that conclusion was reached.
An expert who reproduces AI-generated analysis without independent evaluation risks failing that methodological standard. But the concern extends further. Even an expert who carefully reviews and verifies AI-generated content may find that the system has already shaped the analytical framework, determining which questions were asked, which interpretations were foregrounded, and how the evidence was organized before independent judgment was applied.
Generative AI tools can be valuable when used appropriately. In complex cases involving extensive records, these systems can help experts organize large volumes of documents, summarize deposition transcripts, identify relevant scientific literature, and generate alternative interpretations for independent evaluation. These uses resemble traditional research and organizational tools. The expert remains responsible for evaluating the information and exercising professional judgment.
Problems arise when AI begins to shape the analytical framework itself rather than assist with information management. An expert who asks an AI system to propose diagnostic interpretations, causation theories, or litigation narratives may unknowingly anchor their reasoning around the structure of the generated response. Even if the expert ultimately modifies or rejects the AI's suggestion, the initial framing may determine how the problem is analyzed and what possibilities are never considered at all.
The influence of generative AI may extend into the drafting process in ways that are particularly difficult to detect. When experts use AI tools to assist with writing, the suggested language does more than fill in words. It proposes a structure, an ordering of findings, an emphasis among interpretations, and a narrative logic that organizes the clinical picture. Even careful editing may leave that underlying architecture intact.
This risk operates both within a single report and across an evaluator's practice over time. Within a report, the structure an AI imposes on one section can quietly constrain how the evaluator thinks through the next. A causation narrative that takes shape in section two may determine what evidence feels relevant in section three, before the evaluator has consciously made that judgment. Over time, repeated reliance on AI-assisted drafting may gradually reshape how experts approach forensic questions, not through any single report, but through the slow accumulation of AI-structured thinking.
This is the dimension of AI influence that existing legal and ethical frameworks may be least equipped to address. It is invisible to opposing counsel, difficult to probe on cross-examination, and may not be apparent to the expert themselves.
Forensic experts are expected to offer opinions that reflect independent professional judgment. Ethical guidance from the American Academy of Psychiatry and the Law emphasizes objectivity, methodological transparency, and independence from advocacy pressures. These standards were developed with human cognitive biases in mind. Generative AI introduces a new category of influence that existing frameworks did not anticipate.
Unlike traditional reference materials, AI systems do not simply provide information. They generate explanations and narratives that may frame an expert's analysis before independent evaluation has occurred. The influence is structural, not merely informational, and it may be most pronounced precisely when the expert believes they are thinking independently.
The risk is not that experts will outsource their opinions to artificial intelligence. The risk is that AI will quietly shape the reasoning that produces those opinions.
The question of how experts should use artificial intelligence is, at its core, a question about methodology. For forensic psychiatrists, that methodology has always required more than clinical familiarity with a diagnosis or a treatment record. It requires a structured, transparent, and defensible process for forming opinions, one that can withstand cross-examination and satisfy the evidentiary standards courts apply to expert testimony. Artificial intelligence does not change that requirement.
A sound forensic psychiatric evaluation begins before any record is summarized and before any AI tool is consulted. It begins with a clear formulation of the forensic question, the specific legal issue the evaluation is designed to address. Competency to stand trial, criminal responsibility, emotional distress causation, psychiatric disability: each of these questions carries its own evidentiary framework, its own relevant literature, and its own methodological demands. The forensic psychiatrist must enter the evaluation with that framework already in place, not derive it from whatever an AI system happens to emphasize in a document summary.
From that foundation, the evaluation proceeds through direct examination of primary sources. In psychiatry, this may mean the clinical interview, the review of treating records, collateral information where available, and relevant psychological testing. These sources are not interchangeable, and their relative weight depends on the forensic question being addressed. An AI system can assist with organizing voluminous records or flagging entries that warrant closer review. It cannot replicate the evaluator's judgment about which findings are clinically significant, which reported symptoms are consistent with observed presentation, and which aspects of the history bear most directly on the legal question at hand.
Clinical impressions should be formed independently before AI-generated summaries or interpretations are introduced into the analytical process. Once an AI system has proposed a diagnostic formulation or a causation narrative, that framing becomes the default against which the evaluator's own reasoning is measured. The appropriate role for AI is to stress-test those independently formed impressions, to identify gaps, surface literature, or flag alternative interpretations the evaluator can then assess on their merits. The evaluator's independent reasoning remains the foundation. AI functions as a check on that reasoning, not a participant in forming it.
The use of AI in literature review warrants similar discipline. Generative AI systems can identify relevant research efficiently, but the evaluator must read and assess that literature independently. A system that summarizes a study's conclusions without conveying its methodological limitations, its sample characteristics, or the state of the surrounding literature may produce a confident-sounding but incomplete picture. In forensic contexts, where the strength and reliability of the underlying science is itself subject to scrutiny, that incompleteness can be consequential.
Forensic psychiatric opinions must also be transparent about their basis. Under the ethical guidelines of the American Academy of Psychiatry and the Law, forensic evaluators are expected to be honest about the sources and limits of their opinions. This obligation extends to the use of AI tools. An evaluator should be able to describe, with specificity, how AI was used in the evaluation process, at what stage, and how the evaluator's independent judgment was preserved. If that account cannot be given clearly, it is worth asking whether the methodology was sound.
Finally, forensic psychiatric opinions are always subject to cross-examination. The evaluator who has followed a structured, transparent methodology, forming opinions from primary sources and independent clinical reasoning, is in a position to defend those opinions on their merits. The evaluator whose reasoning was quietly shaped by AI-generated framing may find that defense more difficult, not because the opinion is necessarily wrong, but because the process that produced it cannot be fully reconstructed or explained.
Artificial intelligence will not replace sound forensic psychiatric methodology. But it will test whether evaluators actually have one.
None of this suggests experts must avoid artificial intelligence. These tools will continue to offer genuine value in managing complexity, identifying relevant literature, and supporting the administrative dimensions of expert work. These questions are no longer theoretical for those working in our legal system. The challenge is ensuring that final opinions remain the product of independent professional reasoning. As AI becomes more deeply integrated into legal and professional practice, the question will not simply be whether these tools were used. The harder question, and increasingly the one courts will ask, is whether the expert's opinion is genuinely their own.
Gaube S, Suresh H, Raue M, et al. Do as AI say: susceptibility in deployment of clinical decision-aids. npj Digital Medicine. 2021;4:31. ↩︎
Jabbour S, Fouhey D, Shepard S, et al. Measuring the Impact of AI in the Diagnosis of Hospitalized Patients: A Randomized Clinical Vignette Survey Study. JAMA. 2023;330(23):2275-2284. ↩︎
Goddard K, Roudsari A, Wyatt JC. Automation bias: a systematic review of frequency, effect mediators, and mitigators. Journal of the American Medical Informatics Association. 2012;19(1):121-127. ↩︎
Expert Testimony in the Age of Generative AI: Recent Case Developments. Greenberg Traurig LLP. ↩︎
