Usability of AI-based Evidence

Question

I'm a computer forensics graduate student doing research in blockchain currency (e.g. Bitcoin) forensics. I'm wondering if there are any rulings or decisions regarding the usability and admissibility of AI-generated evidence. For example, if I have an algorithm which identifies suspicious wallets likely associated with ransomware activity and another which helps identify the locations of Bitcoin wallets, how much can I do with this? Can I only use this in the context of furthering an ongoing investigation? Would I be able to obtain a search warrant with this information alone or how much further justification would I need.

While I'm particularly interested in ransomware attack attribution, some of this applies to blockchain currency forensics as a whole. As a toy example, suppose I'm working with a team in Los Angeles tracking illegal weapon shipments into the United States and blockchain analysis points us to a distributor likely working out of Wyoming. Could we use this

I know there are rulings regarding digital software such as EnCase; since this is approved by the IEEE and the digital forensic community as a whole its use is generally admissible in court. It gets a bit more dicey when the basis for your work is not well-known. For example, while the PageRank and DeepWalk algorithms for graph analysis are pretty well-known, the use of them together to identify suspicious Bitcoin wallets is novel and would likely require some justification.

Are there any relevant rulings, decisions, etc. that you could point me towards?

Answer 1

In principle, the evidence could be admitted. Federal Rule of Evidence 702 allows an expert to testify if

(a) the expert’s scientific, technical, or other specialized knowledge will help the trier of fact to understand the evidence or to determine a fact in issue;

(b) the testimony is based on sufficient facts or data;

(c) the testimony is the product of reliable principles and methods; and

(d) the expert has reliably applied the principles and methods to the facts of the case.

A precedent for admitting AI evidence comes from Bertuccelli v. Universal City Studios. The court determined that

Dr. Griffor has experience with algorithmic reasoning for artificial intelligence-enabled driving systems, including facial recognition technology and is considered an expert in the field of facial target recognition. The Court finds Dr. Griffor's methodology reliable given that he conducted an artificial intelligence assisted facial recognition analysis of the King Cake Baby and Happy Death Day mask to determine whether the use of mathematics and target facial recognition algorithms comparing the two works would find that human perception would view the works as substantially similar. Accordingly, the Court finds Dr. Griffor is qualified to testify as an expert in this case.

Note that some person has to testify: an machine printout can't just be entered as proof of guilt or innocence. It does not automatically follow that a particular AI method is admissible, but it is at least potentially admissible.

Then if some evidence is admissible in court, it may also form the basis for searches and seizures requiring reasonable suspicion or probable cause. See also this article on admissibility of machine learning evidence.

Answer 2

Your question is about the admissibility of expert opinion evidence. The applicable law in U.S. courts is called the Daubert standard. Under this standard, the judge is required to play a “gatekeeper” role by “ensuring that an expert’s testimony both rests on a reliable foundation and is relevant to the task at hand” before admitting it. Although the details vary, most jurisdictions apply a similar rule intended to exclude pseudoscience from the courtroom.

The various decisions about TrueAllele show how these standards have been applied to a novel algorithm. The software vendor publishes a list of decisions, dating back to 2009, upholding the admissibility of expert evidence based on TrueAllele. These include rulings of the Pennsylvania Superior Court (2012), New York Supreme Court (2019), Nebraska Supreme Court (2019) and Florida Court of Appeal (2021).

Although TrueAllele is not “AI,” it is proprietary software, which has raised many of the “black box” concerns associated with machine learning systems. See Pishko J, The impenetrable program transforming how courts treat DNA evidence, WIRED (29 November 2017). Machine learning evidence: admissibility and weight (2019) 21:3 University of Pennsylvania Journal of Constitutional Law 919 is a detailed scholarly analysis of the issues with machine learning evidence in particular.

Answer 3

In most cases, the kind of AI you are talking about in the OP would be useful as part of a larger set of facts used to establish probable cause that a crime was committed, in order to justify an arrest or search warrant. But merely fitting a profile alone is neither sufficient for probable cause or to convict, and the profile would probably be inadmissible as character evidence in a trial on the merits.

You would probably need something in addition to the AI to corroborate your AI conclusion to establish probable cause to arrest or get a search warrant.

You probably could use AI to make the equivalent of a "Terry stop" (to the extent that such action would even make sense in the cybercrime setting) which involves a brief non-custodial interruption of someone's activities to ask a few questions or to ask someone to identify themselves, based upon a reasonable suspicions of a crime being committed (a lower standard than probable cause), or to issue a subpoena, something that a grand jury or in many states, a prosecutor without a grand jury, has the authority to issue, since a subpoena does not have to be supported by probable cause.