By Michael Barr
Note: This article offers practical guidance for attorneys engaging technical experts. It is general commentary, not legal advice. Consult counsel about any specific matter.
An artificial intelligence expert witness is being asked, more and more often, to explain a system that its own builders cannot fully account for. This is the newest field in this series and the one where credentials are least reliable, because the population claiming expertise has grown faster than the population with any.
They distinguish the model from the system around it
Almost nothing in dispute is the model alone. It is the training data and how it was assembled, the preprocessing, the prompt or feature construction, the thresholds someone chose, the human review step, and the product decision to ship at a given confidence level. Harm and infringement usually live in that surrounding system.
An expert who can only discuss architectures will describe the least contested part of the case.
They can testify about training data provenance
Where the data came from, what license it carried, whether it was scraped, whether it was licensed for this use, and whether a specific work is in it are now central questions in copyright and trade secret matters. Answering them requires the discipline of provenance work: manifests, ingestion logs, dataset versioning, and deduplication records.
An expert who accepts a party's description of its own dataset without examining the pipeline has not tested the thing in dispute.
They know what a model can and cannot be made to reveal
Memorization, extraction, and membership inference are real phenomena with real limits. So is the fact that a plausible output is not proof that a specific input was in training. This field invites overclaiming in both directions, and the credible witness marks the boundary between what an experiment demonstrates and what it merely suggests.
They can run reproducible experiments
Where behavior is contested, the answer is usually an experiment: a held-out probe set, a controlled prompt suite, a measured error rate across a defined population. That needs to be documented well enough for another expert to repeat it, with seeds, versions, and configuration recorded.
An expert who reports what a system did when they tried it a few times has produced an anecdote.
They handle bias and performance claims quantitatively
Assertions that a system performed worse for a group, or below its marketed accuracy, are measurable and belong in numbers with confidence intervals and a stated population. An expert who argues these qualitatively will be met by one who does not.
They are candid about the state of the art at the relevant time
This field moves fast enough that the standard of care shifts within the life of a case. What a reasonable developer would have done in 2021 differs from 2026, and applying today's practice to an earlier system is the same dating error that damages medical device and automotive testimony.
They can teach it without either mystifying or trivializing
Juries arrive with strong impressions of this technology and few accurate ones. The witness has to displace the impression without lecturing, and without reaching for the two failure modes: making it sound like magic, or making it sound like a lookup table.
Expert profiles in this field
Anonymized profiles. Names follow a conflicts check.
- Speech recognition and natural language processing
- Speech and voice recognition
- Social media misinformation detection
- Cluster and supercomputing
- Source code analysis and distributed systems
How we help
We vet in this field harder than in any other, because the gap between claimed and actual expertise is widest here. That means an engineer reading a candidate's own work and asking them to account for it, which is the process reflected in our guide to how expert witness firms vet and prequalify experts.
Our artificial intelligence and machine learning expert witnesses work across patent, trade secret, copyright, and product matters, and our source code review practice covers training and inference pipelines under protective order.
The search is free and confidential for the attorney. Our fee is built into the expert's rate, so we are paid only if you retain. Tell us about your case.