AI is the shiny new tool in every trial lawyer’s toolbox. It can research in minutes what used to take nights and weekends. Like always, you need to check the results lest AI’s desire to please result in your being sanctioned for providing hallucinated results as fact. Nothing new about that. What is more interesting, and potentially game changing, is AI’s ability to provide “expert opinion” without a human sponsor. How courts handle AI’s expert opinions from an evidentiary perspective is a work in process.
An example of the value of AI’s expertise is illustrated by our recent evaluation of a wrongful death case involving pedestrian versus car. Within hours of the tragedy, the state police had three video feeds from nearby security cameras capturing the collision from three separate perspectives. Enter Claude. We took those videos and uploaded them to Claude with some limited instructions about what we were hoping it could tell us. (Notice I said “it” despite my first instinct being to say “he.”) Claude processed the videos and produced a frame-by-frame breakdown that we could flip through like an animation. At ten frames to a second, we could see the disaster unfold one frame at a time—always irrationally hoping somehow the result would change. Claude then analyzed the data to provide information on the vehicle speed at various points leading to the crash. It also calculated estimated stopping distances, reaction times and other key data points given several different assumptions. When prompted, Claude explained its analysis and provided a list of follow-up tasks and potential defenses.
Accident reconstruction is just the tip of the iceberg when it comes to AI’s potential for expert opinion. The several categories of such expert evidence include:
- Forensic identification and comparison — probabilistic genotyping (TrueAllele, STRmix) producing DNA likelihood ratios; facial recognition matches; latent‑print and firearm/toolmark comparison; voice/speaker identification; gunshot‑detection classification and localization; disputed-document examination, including handwriting and signature comparison.
- Financial and economic analysis — AI forensic accounting and fund tracing; anti‑money‑laundering detection; damages and lost‑profits models; algorithmic business or asset valuation.
- Medical and scientific opinion — AI reads of diagnostic imaging (radiology, pathology); prognostic models; algorithmic causation assessments in toxic‑tort and product‑liability matters.
- Accident reconstruction and engineering — physics‑based collision simulations; structural‑failure modeling; 3‑D reconstructions or animations embodying an expert conclusion.
- Risk and behavioral prediction — recidivism/risk scores (e.g., COMPAS) at bail or sentencing; dangerousness predictions.
- Document and authorship analysis — stylometry/authorship attribution; deepfake‑detection outputs; source‑code similarity in trade‑secret or copyright disputes.
- Language and data interpretation — machine translation offered for meaning; sentiment analysis; transcription with speaker attribution; large‑scale pattern extraction.
- Cybersecurity and technical attribution — malware attribution; intrusion and network‑forensic conclusions.
- Digital forensics and geolocation — AI-assisted cell-site and GPS location analysis; device-extraction and data-recovery conclusions; call-detail-record mapping.
- Generative LLM outputs — e.g., the model’s characterization of whether a contract term is “industry standard.”
The question is: Can I put Claude’s opinion into evidence? That question is the subject of a proposed change in the Federal Rules of Evidence to include new Rule 707, which states:
Where machine-generated evidence is offered without an expert witness and would be subject to Rule 702 if testified to by a witness, the court may admit the evidence only if it satisfies the requirements of Rule 702(a)-(d). This rule does not apply to the output of simple scientific instruments.
In other words, where you don’t have a human expert, you might still meet your burden if you can offer the same expert opinion generated by AI. The rule contemplates applying the same criteria regarding reliability to machine-generated opinions as apply to those same opinions when offered by a human expert subject to Rule 702 and the court’s gatekeeping function.
Without an expert to defend the methodology or face cross‑examination, courts have been left to consider the AI evidence under the general rules pertaining to relevance and authenticity set forth at Rules 401, 402, 403, and 901. (Note that a separate proposed amendment, Rule 901(c), deals with authentication of potentially AI-fabricated depictions fabricated depictions, i.e., deepfakes.)
In criminal cases, a further constitutional hurdle may await. The Sixth Amendment’s Confrontation Clause guarantees the accused the right to confront the witnesses against them, yet you cannot cross-examine an algorithm. Whether a machine-generated opinion can be “confronted” at all (and whether testimony from a human who sponsors or explains the model cures the defect) remains an open and consequential question.
If Rule 707 is adopted (or even before its adoption when used as a framework to test the admissibility of AI-generated opinions no matter whether offered directly or sponsored by a human expert witness), counsel should be prepared to consider questions like: Was AI’s forensic identification based on proprietary source code that the other side wants to examine and that you don’t have and cannot get? In the absence of a human expert witness, how can you establish reliability or secure judicial notice of reliability? Because LLMs are probabilistic such that a prompt may yield different outputs on different runs, is that uncertainty disqualifying or is there some degree of uncertainty permissible within the reliability standard? What body of data was the generative AI model trained on and what was the algorithm used to process that data? And is the output even hearsay or, because no human “declarant” made a “statement,” does it fall outside Rule 801 altogether, leaving reliability rather than confrontation or cross-examination as the principal battleground?
In summary, AI has the potential to make trial work less expensive, more efficient and to produce a fairer result. The realization of that promise is yet a ways off and will require a lot more thought and judicial deliberation, part of which will be focused on the admissibility of AI opinions like the one Claude prepared for us.