Today’s guest post is from Nick Dellefave, an up and coming Holland & Knight litigator.  The Blog has rolled out a few posts on the latest edition of the Reference Manual on Scientific Evidence.  Nick adds to this opus with a dive into the intersection between scientific evidence, the role of trial judges, and this thing called “artificial intelligence,” which the Blog has also touched on from time to time.  As always, the guest poster deserves all of the credit and/or blame for the post.  (At a minimum, Nick gets some credit for allowing us to not have to write a post while on vacation.)

**********

As we covered last month, the new Reference Manual on Scientific Evidence is here, complete with a new chapter on Artificial Intelligence.  While the entire chapter should be required reading for litigators practicing in 2026, today we are focusing on one key subsection, titled “Judges as AI Gatekeepers.”  It offers a window into how courts will evaluate AI-generated evidence, and more importantly, where the pressure points are for both sides.  We thought it would be useful to pull out the key takeaways for practitioners who may be dealing with AI evidence sooner than they think.  This subsection has not drawn the attention of the climate science chapter, but it may end up being more impactful.

The Familiar Rules, Applied Unfamiliarly

Here is the good news: AI evidence will be evaluated under the evidentiary rules we all know and love.  The not-so-good news is that applying those rules to AI is anything but straightforward.  Under Federal Rule of Evidence 401, AI applications are essentially probability assessment tools, theoretically making them inherently relevant when offered to demonstrate whether something is “more or less probable.”  But Rule 403 gives courts plenty of room to exclude AI evidence when it risks unfair prejudice, confusion, or misleading the jury.  This is likely to be a prime battleground in determining the admissibility of AI evidence, as litigators and judges will rightfully ask whether the AI’s probative value is substantially outweighed by these dangers.

If you are challenging AI evidence, be ready to argue that certain outputs are too inaccurate or biased to be relevant, or that they are simply a poor fit for the purpose offered.  Courts may exclude AI evidence under Rule 403 precisely because juries tend to assume that if a computer said it, it must be true.  One need only look around at the epidemic of advertisers boasting “AI-powered” products to see this concept in action.  The imprimatur of “science” or “technology” can lend AI evidence a false authority it may not deserve.  Of course, with changes in jury attitudes about science and experts, it is hard to be sure how AI evidence will be received when its source is revealed.  If you are opposing AI evidence, hammer this point.  If you are the proponent, be prepared to show that the evidence is reliable and appropriately scoped for the purpose for which you are offering it.

Discovery Is Where the Action Is

The foundation for admissibility fights over AI evidence will be laid in discovery, which promises to generate its own fights.  Opponents seeking to challenge AI evidence will need access to the underlying algorithm, the training data, and knowledge of what is happening inside that machine-learning black box.  If you are challenging AI evidence, go after this information aggressively.  If you are the proponent, decide early on how you will address those requests, because rest assured, they are coming.  For more on this, see our recent discussion on Discovery of Artificial Intelligence Prompts. 

Of course, there is tension here with trade secret protection.  AI developers are not exactly eager to open up their proprietary algorithms for public inspection.  Courts have tools to navigate this: they can seal records, use protective orders, and exercise their general power to oversee how evidence comes in.  The Defend Trade Secrets Act (18 U.S.C. § 1835) also gives courts specific direction to protect trade secrets in relevant proceedings.  Attorneys should be prepared to negotiate protective orders and in camera review procedures.

Daubert, But Make It Complicated

Daubert as spelled out in the enhanced Rule 702 is, of course, the go-to framework for expert testimony, but applying its factors to AI is trickier than it sounds.  Every AI application is different: different algorithms, different machine learning methodologies, different training data.  This means AI issues are generally not going to be resolved the way DNA analysis was.  There will not be a single landmark case that settles the question of when AI evidence comes in.  Instead, expect adjudication for each application and each context.

The question of testability—the hallmark of Daubert/Rule 702 analysis—is complicated by the multiple inputs into an AI engine system.  The threshold question is: what exactly are we testing?  The sensors?  The algorithm?  The math?  The training data?  The system as a whole?  Similarly, credible peer review requires access to the underlying algorithm, parameter weights, and training data.  If the AI evidence being offered has not been meaningfully peer reviewed, and if that review did not include the underlying code and data, that is a vulnerability to exploit.

Error rates are another minefield.  Attorneys must ask whether error rates vary depending on whether the AI was tested using the relevant local population to which it will be applied, as opposed to a national population or idealized lab database.

Context and Bias: The Low Hanging Fruit

Attorneys should scrutinize whether the AI application is a good “fit” for the purpose for which it is proffered.  Is the AI application actually designed for the purpose it is being used for?  Some algorithms are built for one thing and then repurposed for another—with predictably dicey results.  The Manual’s authors cite the example of criminal risk assessment algorithms, which might be designed to determine who would benefit from alternatives to incarceration, but are probably inapt for use in sentencing determinations.  More applicable to readers of this blog, algorithms designed to model drug or device efficacy might be inapt for use in modeling the incidence of adverse effects.

Even when an algorithm is being used for its intended purpose, it may perform worse in certain contexts because of how it was trained.  An AI trained on one population may not achieve the same accuracy when applied to a different demographic.  If there is a mismatch between the training population and the population at issue in your case, that disparity is ripe for attack.

There is also the risk that neural networks rely on inapt factors—i.e., things they should not be considering—in making their predictions.  Judges are going to want to know whether inappropriate or unconstitutional factors were included, and whether any factors are working as proxies for suspect categories like race.  Demand disclosure of what factors are in the model and how they are weighted.  You may not get it easily, but in many cases, this is a fight worth having.

Deepfakes and the Importance of Witness Authentication

The Manual’s authors devote several pages to the problems posed by deepfakes.  Deepfake technology is only getting better, cheaper, and more accessible.  Practitioners should prepare for increased litigation over the authenticity of photos, videos, and audio recordings, as the mere availability of deepfake technology throws the authenticity of all types of evidence into doubt.

Under Rule 901, the proponent has to produce evidence sufficient to show the item is what it claims to be.  But here is the problem: your authenticating witness may not be able to tell whether the evidence is genuine or fake.  Deepfakes can shade human memory.  Archive systems can be hacked.  Digital forensic experts are going to be in high demand, and rightly so.

On the authentication front, there are some tools worth knowing about.  Federal Rules of Evidence 902(13) and 902(14), adopted back in 2017, provide for self-authentication of certified electronic records and certified data copied from electronic devices.  Items with cryptographic hashes may be authenticated under these rules, subject to the adverse party’s “fair opportunity to challenge.”  Although litigators may be accustomed to stipulating to the authenticity of electronic evidence, in this age of deepfakes, they should be increasingly wary of waiving these challenge opportunities without consulting a digital forensic expert.

So What Should You Actually Do?

If you are offering AI evidence, line up your expert testimony now.  As we have argued, admission of AI evidence likely requires the proponent to offer an expert who can speak credibly to the AI’s design, training, validation, and accuracy.  Your goal should be to simplify the foundational requirements and make the evidence as easy to admit as possible.  If you are opposing AI evidence, your job is the opposite: attack relevance and reliability, demand discovery of algorithms and training data, and prepare to cross-examine the software engineers who built the system.

We expect that AI evidence will continue to be an area where the rules are being written in real time.  Practitioners should understand that courts are going to face layered adjudicative challenges every time AI evidence comes up.  As the chapter’s authors put it, “the most important thing courts can do is ask careful and informed questions.”  That means the most important thing you can do as an advocate is be ready to answer them well.