Early in the pretrial phase of litigation, the court should ask the parties whether the case is likely to involve AI-generated evidence. If so, the court and the parties can address the parameters of potential discovery related to the AI system, as well as its input and output data. They should also discuss whether a Daubert hearing will be necessary.

Discovery of AI systems is likely to implicate requests for training data, source code, and/or other information that a party may deem to contain or reflect proprietary trade secrets. To address these concerns, the parties (or the court) can craft suitable protective orders. 

An important evidentiary issue that can impact acknowledged AI-generated evidence is bias. Bias can affect the output of GenAI systems in different ways. ‘Statistical’ or ‘mathematical bias’ refers to an AI system that generates non-random error; for example, when GenAI is asked to provide a random number between 1 and 100, it may respond with ‘42’ due to a pop-culture reference in The Hitchhiker’s Guide to the Universe. ‘Discriminatory bias’ refers to an AI system that produces systemic errors related to protected groups. One example of this is when facial recognition systems misidentify people of color.

The judge must decide any preliminary questions concerning the admissibility of AI-generated evidence (Fed. R. Evid. 104(a)). As a general rule, witnesses called to authenticate AI-generated evidence must have personal knowledge of the authenticating facts (Fed. R. Evid. 602).  Under Fed. R. Evid. 901(a), the proponent of the AI-generated evidence must produce evidence through witness testimony sufficient to support a finding that the item 'is what the proponent claims it is.' The proponent may use one of the non-exclusive methods listed in Fed. R. Evid. 901(b). The most common methods are: (i) testimony of a non-expert witness (901(b)(1)) and (ii) evidence describing a process or system that shows it produces an accurate result (901(b)(9)). Authentication must be established by a preponderance of evidence.

A court’s analysis of challenges to AI-evidence is likely to be similar to challenges of other types of scientific or highly technical evidence.  As a result, some evidentiary challenges are likely to implicate expert testimony. When addressing challenges to expert testimony involving AI evidence, courts will apply the familiar framework of Rule 702 and Daubert v. Merrell Dow Pharmaceuticals Inc., 509 U.S. 579 (1993). Rule 702 requires expert evidence to be the product of reliable principles and methods, and many Ryle 702 challenges are therefore likely to consider whether the AI system used to produce the evidence was tested, had a known error rate, or incorporated bias. Relatedly, courts may have to consider the methodology an expert used to produce the evidence from the AI system: was the methodology subject to peer review, was the methodology generally accepted in the scientific community, and were standard procedures used?  

Because disputes over AI-generated evidence and AI systems may require the involvement of experts, a court may encounter the situation where an expert is needed but no expert has been retained.  In such a situation, Fed. R. Evid. 706 authorizes the court to appoint an expert.  If the parties do not have funds to pay for such an expert, however, the court may be unable to invoke Rule 706.  

A more in-depth exploration of how the Federal Rules of Evidence and Daubert factors might apply to AI-generated evidence can be found in Artificial Intelligence section of the Federal Judicial Center’s Reference Manual on Scientific Evidence, beginning on page 1542, Judges as AI Gatekeepers, as well as in the Center’s 2023 Introduction to Artificial Intelligence for Federal Judges, chapters 6 and 7.