Where machine learning meets the legal system — admissibility of AI-derived evidence, discovery into models and training data, algorithmic accountability, and the fast-moving law of AI governance.
Machine-learning systems break several assumptions the rules of evidence and discovery quietly depend on: that a process repeats, that a record can be read, and that provenance can be traced. Courts are now resolving those questions with tools written for filing cabinets.
The same prompt to the same model can produce different outputs. A single exhibit therefore rarely establishes what a system 'does' — reproducibility has to be demonstrated through a designed protocol on agreed inputs, not assumed from an example. This is the point at which ordinary exhibit practice stops working and testing methodology becomes the disputed issue.
A model's behaviour is encoded across millions of parameters rather than in readable rules, so reviewing the source code alone does not explain an output. Explaining one requires structured testing of the trained system. Courts should be alert to the difference between a party that will not explain its model and a party whose model cannot be explained by code review — the remedies are not the same.
Training corpora can span billions of records from many sources, which makes 'what went in' expensive rather than simple to answer. Recent training-data and model-inspection orders have turned on the ordinary Rule 26(b)(1) relevance and proportionality analysis rather than on anything novel about the technology. Counsel who frame these demands as unprecedented tend to get worse outcomes than counsel who frame them as scope.
Rule 901 requires evidence sufficient to support a finding that an item is what its proponent claims. Synthetic audio, video, and documents put pressure on that standard because the ordinary indicia of authenticity can now be generated. Proposed Rules 901(c) and 707 address the question directly, and the drafting debate is worth following closely by anyone litigating with machine-derived exhibits.
A neutral with command of the systems can design reproducible tests, trace the provenance of disputed training data, inspect code and weights under a protective order, and evaluate competing measurements of accuracy, bias, or copying — then report findings in terms the bench can use while reserving questions of law to the court. That preserves both the proprietary material and the record, which party-to-party discovery in these matters struggles to do at once. What such a reference is scoped to reach, and what the order of reference has to say, is set out on the AI and algorithmic special master page linked below.
Recent training-data and model-inspection orders show that the ordinary rules of relevance and proportionality, not novelty, decide how far discovery reaches into a machine-learning system.
Read the analysis →How courts are applying Rules 901 and 902 to synthetic audio, video, and documents — and what proposed Rules 901(c) and 707 would change.
Read the analysis →Machine-learning systems break the assumptions behind ordinary discovery. A technical special master gives courts a way to examine models, training data, and outputs that the rules were not built to reach.
Read the analysis →For counsel weighing a reference
Learn how to propose one to the court, review model appointment orders, or request a conflicts check — without obligation.
Also on the practice: the AI and algorithmic special master, who pays a special master, and the neutral’s credentials.
Stay current