The Admissibility of Artificial Intelligence | New York Law Journal

As the use of artificial intelligence (AI) increases, so does the potential for its role in civil and criminal proceedings. While AI is often viewed as a computer that can match or exceed human performance in tasks requiring cognitive abilities, it is in fact simply software. Software is generally admissible as evidence if it is relevant, material, and competent. However, AI differs from traditional software, perhaps requiring new considerations regarding admissibility.

More specifically, both traditional software and AI contain algorithms. Algorithms are procedures used to solve a problem or perform a calculation. Algorithms act as a step-by-step list of instructions that specify specific actions to be performed using hardware- or software-based procedures. The fundamental difference between AI and traditional algorithms is that AI can change its output based on new input, whereas a traditional algorithm will always produce the same output for a given input.

AI, like traditional software, can create two types of evidence, namely computer records and computer-generated evidence. Computer records do not require analysis or assumptions by programming, while computer-generated evidence does. Computer records are typically printouts made by a computer in a specific way from data. Computer-generated evidence is the output of a computer based on data and assumptions contained in a program.

The admissibility of both computer records and computer-generated evidence is the same as for traditional paper business records and traditional demonstrations. The fact that a computer is involved does not change the standards of admissibility or procedures.

The identification and authentication of both paper business records and computer records are the same as for any other written document (Federal Rules of Evidence 1001(1)(3). Both are no less second-hand evidence than any other event they are intended to prove. Accordingly, they must meet the requirements of the rule against second-hand evidence in order to be admissible. Admissibility requires a proper basis (see, for example, United States v. Catabran 836 F 2nd 453 (1988)). More specifically, courts generally require that the records be produced in the ordinary course of business, be identified by a qualified witness, and that the source, method, and timeliness of preparation suggest that the records are reliable.

There are two types of computer-generated evidence. One is demonstrative, and the other is experimental. Demonstrative evidence is usually static information, such as a computer-aided design of a pipe after it has been broken. Experimental evidence is usually dynamic information, such as the output of a computer model resulting in a simulation of the pipe being broken.

Static representations used in court are almost always subject to the same rule for admissibility, regardless of the computer’s role in creating them. The standard is simply that the representation accurately describes what it is intended to illustrate. Such a representation does not require an expert witness.

Dynamic evidence (such as simulations) typically requires that an expert certify that the simulation, for example, is derived from principles and procedures that are generally accepted scientific standards (see generally, Frye v. United States 293 F 1013 (1938)). In addition, admissibility requires some amount of experimental testing to confirm that the simulation is consistent with reality.

One of the most serious problems with AI-generated algorithms and their results is the lack of proper evaluation. AIs are trained on data and learn to create algorithms, which in turn make predictions by finding patterns in the data. However, if the training data is incomplete or biased, the AI ​​can learn incorrect patterns. This can lead to the AI ​​model making incorrect predictions or hallucinating (see cases related to Roberto Mata v. Avanca Case 1:22-cv-01461-PKC, such as https://www.documentcloud.org/documents/23826751-mata-v-avianca-airlines-affidavit-in-opposition-to-motion?responsive=1&title=1 ).

The admissibility of AI evidence is likely to be challenged by the lack of rigorous testing, since AI algorithms can have significant legal implications. Even when tests of AI algorithms are conducted, they are rarely independent, peer-reviewed, or transparent enough to be properly assessed by people competent to do so. This shortcoming is likely to raise concerns about the admissibility of AI evidence under the Frye standard.

More specifically, the Frye standard requires that the admissibility of computer-based evidence, such as AI evidence, be based on scientific methods that are sufficiently well-established and accepted. Because there are no standards for testing AI algorithms in general or for testing AI products specifically, it will be difficult to obtain expert opinion that AI evidence is admissible because it is “generally accepted” as credible in the relevant scientific community.

Before AI evidence is admissible, courts should require transparency and explainability, both in terms of how the AI ​​system operates and how a decision was made (e.g., whether a pipe will burst), classification (e.g., whether the pipes qualify for a valid simulation), or prediction/inference (e.g., did the test pipe burst as predicted?).

In short, the admissibility of AI evidence related to high-stakes decisions must include explanations that reveal its inner workings and how the AI ​​modifies its algorithms so that the AI ​​algorithms are explainable. The rules of admissibility of evidence typically require that the evidence be relevant, material, and competent. Competent evidence is understood to mean relevant evidence admissible in the action; that is, relevant evidence that is not subject to any exclusionary rule (see People v Brewster, 100 AD2d 134 (1984)).

Evidence is “competent” if it conforms to some traditional notions of reliability (i.e., trustworthy). In relation to AI evidence, such evidence can be considered trustworthy if it can explain the reason(s) for all the results and show that the process by which the AI ​​system generates the results is correct.

Another challenge to the admissibility of AI evidence is robustness. Robustness is the degree to which AI resists both intentional and unintentional efforts to cause machine learning models to fail.

Currently, AI that relies on internet data to learn can develop faulty algorithms due to spoofing. Often spoofing is the act of masking the source of internet communication from an unknown source that came from a known, trusted source. Spoofing can first lead to faulty AI machine learning, then to faulty algorithms, and finally to unreliable AI evidence.

It is becoming increasingly difficult to distinguish between human-generated and AI-generated web content. As a result, AIs are being used to corrupt other AIs. This practice, known as adversarial AI, uses AI to trick machine learning models by providing misleading input. Adversarial AI can be used to modify the output of most AI technologies.

While AI uses technology that may exceed human cognitive abilities, the rules of evidence do not provide a separate standard of evaluation. Thus, evidence obtained from AI should be evaluated by the standard of direct witness testimony, expert witness testimony, or measurement using established technology. In summary, AI evidence is subject to the same rules of evidence as non-AI sources.

Consider Federal Rule of Evidence 401, which defines materiality. Rule 401 states that evidence is material if: (a) it has any tendency to make a fact more or less probable than it would be without the evidence; and (b) the fact is material to the determination of the action.

Rule 401 is typically considered in conjunction with Rules 402 and 403. More specifically, Federal Rule of Evidence 402 reveals that the admissibility of evidence in federal court is typically based on its materiality. Unless specifically prohibited, material evidence is admissible. Rule 403 limits Rule 402 by excluding material evidence if its probative value is outweighed by prejudice, confusion, or waste of time.

With respect to the admissibility of AI evidence, Rule 403 has two important features. First, Rule 403 identifies the trial judge as the decision-maker. Second, Rule 403 provides that the judge may not make a decision unless the party offering the AI ​​evidence is willing to disclose background information. This would include training data, as well as the development and operation of the AI ​​system sufficient to allow the opposing party to challenge it.

Therefore, AI evidence in civil or criminal proceedings should not be admissible unless the information underlying the AI ​​is available. Such information must be sufficient for the party against whom the evidence is presented to establish validity (including the accuracy of the aid) and reliability (i.e. the AI ​​algorithm correctly measures what it purports to measure).

As with non-AI information, the trial judge should give the AI ​​evidence applicant a choice. The applicant can either disclose the underlying evidence (perhaps under an appropriate protective order) or otherwise demonstrate its validity and credibility. If the applicant is unwilling to do so, the AI ​​evidence should not be admissible.

Federal Rule of Evidence 901(a) requires AI evidence to be authenticated before a jury can hear it. Rule 901(b) provides a number of ways a party can achieve this goal. There is no specific exception for AI evidence. A witness with knowledge of AI will testify that the AI ​​is what it claims to be (under Rule 901(a)); and then that witness will describe a process or system under Rule 901(b) and show that it produces an accurate result. Because AI programming is not widely known, Rule 602 would be expected to apply, requiring the authenticating witness to have personal knowledge of the operation of AI technology or to be an expert.

Since AI typically requires both machine learning and generative elements, it is unlikely that a single witness will suffice for admissibility purposes. More specifically, machine AI typically requires a single set of skills to teach the computer to understand specific data and perform specific tasks. Generative AI typically requires a single set of skills to build on that foundation and add new capabilities that attempt to mimic human intelligence, creativity, and autonomy.

Jonathan Bick is of counsel at Brach Eichler in Roseland and chairman of the firm’s patent, intellectual property and information technology group. He is also an adjunct professor at Pace Law School and Rutgers Law School.