Texas Bar Journal • June 2025

AI Evidence in Texas Litigation

Admissibility challenges in the age of generative AI.

Written by Heather L. King and Tom Daley

A stressed man leaning up against his office wall

Generative artificial intelligence is, at its core, software. Understanding the distinct differences between generative AI and the traditional software that we’ve all become used to is a key to mastering the unique evidentiary challenges generative AI presents to litigators. A primary difference lies in how generative AI algorithms process information versus traditional software algorithms. An algorithm is a procedure or set of rules for solving a problem. Traditional algorithms follow fixed, human-programmed rules that consistently produce the same output for identical inputs. In contrast, generative AI algorithms can modify themselves based on new information, identifying patterns in data and adapting their processing accordingly. This difference is crucial for evidentiary purposes. While traditional software produces consistent, predictable outputs that are relatively straightforward to authenticate, generative AI systems may produce different outputs over time as they learn from new data. This adaptability raises important questions about reliability, transparency, and authentication under Texas Rules of Evidence (TRE) 901,1 and thus admissibility. For instance, generative AI models generate text based on patterns in their training data. The more frequently a phrase or concept appears in the data, the more likely the model is to reflect it in its outputs. This is critical in legal contexts, where overrepresentation of certain authorities or jurisdictions may unintentionally bias the generative AI’s responses. As an example, if a generative AI legal research tool is trained predominantly on Texas caselaw, it may generate responses favoring Texas legal principles, even when a certain aspect of the case arises under a different state’s laws. This does not mean the AI is incorrect—it simply reflects the weight of its training data.

A litigator can use his or her own legal research experience as an illustration of the potential evidentiary challenges presented by generative AI. An attorney using a traditional search engine for legal research is engaging in retrieval, pulling up existing cases, statutes, or secondary sources, ensuring that the result is grounded in prior legal authority. The goal being to guide users to ground truth, the search engine returns the closest available existing answer to the query, sourced from published materials. Generative AI, on the other hand, does not retrieve information, but rather generates it, creating responses based on patterns in its training data. While its answers may sound cogent, they can also be hallucinated (plausible but incorrect). This is not a generative AI failure, it is a fundamental difference in purpose, as generative AI is designed to predict the most statistically likely answer, not necessarily the most factually accurate one. Generative AI provides the statistically most likely next word in the sequence, without consideration for the accuracy of the final output.

Among the most serious concerns related to generative AI algorithms and their outputs is the potential for hallucinations, incorrect or misleading results generated through a generative AI’s internal processing. These errors can stem from various factors such as insufficient or biased training data, incorrect assumptions and overfitting (the AI may become too specialized with respect to its training data, leading to poor performance with new inputs). Litigators must be prepared to address potential hallucinations and explain how the reliability of the output has been verified. Another major concern involves deliberate manipulation of generative AI systems, either through spoofing or adversarial techniques. Spoofing occurs when the source of an internet communication is disguised to appear as if it comes from a known, trusted source. Adversarial AI involves using one generative AI system to deliberately manipulate or corrupt another. This practice can significantly impact the reliability of AI-generated evidence.

A significant part of the generative AI evidence admissibility analysis is resilience, the degree to which a generative AI system resists both intentional and unintentional efforts to cause machine-learning models to fail. Resilience speaks directly to reliability, a critical factor in determining whether evidence should be admitted. When presenting AI-generated evidence, attorneys should be prepared to establish the resilience of the underlying system, addressing how the system has been tested against both accidental failures and deliberate attacks.

Since generative AI programming is not common knowledge, TRE 6022 will typically apply, requiring the authenticating witness to have personal knowledge of how the generative AI technology functions or to be established as an expert in the field. Importantly, because generative AI usually involves both machine learning and generative elements, a single witness may be insufficient for authentication purposes. Different aspects of the system—data preparation, algorithm development, and implementation—may require testimony from multiple witnesses with distinct areas of expertise. While Texas courts have not yet established a comprehensive framework specifically for generative AI evidence, existing caselaw regarding technology-generated evidence provides valuable guidance. This standard suggests that for generative AI evidence, Texas courts will likely require proof that the AI system represents accepted technology within its domain, evidence that the system was functioning properly when the evidence was generated, and testimony regarding the system’s overall reliability and error rates. Many forms of AI evidence may be subject to the reliability standards established in E.I. du Pont de Nemours & Co. v. Robinson,3 which adopted a modified version of the federal Daubert standard. Under Robinson, scientific evidence must be relevant to the issues in the case, based on a reliable foundation and helpful to the trier of fact.4

Given the technical complexity of generative AI systems, expert testimony often plays a crucial role in establishing the admissibility and proper interpretation of AI-generated evidence. Under TRE 702-705, such testimony must be provided by a qualified expert with appropriate knowledge, skill, experience, training, or education; help the trier of fact understand the evidence or determine a fact issue; be based on sufficient facts or data; be the product of reliable principles and methods; and involve reliable application of principles and methods to the facts. A key function of expert testimony regarding generative AI evidence is explaining the technology to judges and juries in comprehensible terms. Effective experts will be able to describe complex generative AI processes using accessible analogies, explain in non-technical terms how the system produced its outputs, articulate potential limitations or sources of error, and help the trier of fact appropriately weigh AI evidence against other forms of evidence. Given the complexity of many generative AI systems, courts may require testimony from multiple experts addressing different aspects of the technology, such as data scientists to explain data selection and preparation, AI engineers to explain model architecture and training, domain experts to interpret outputs in specific contexts, and statisticians to address reliability and error rates.

Under TRE 401,5 generative AI evidence must meet this basic relevance threshold, just like any other form of evidence. Even when such evidence is relevant, TRE 403 may exclude it if its probative value is substantially outweighed by the danger of unfair prejudice, confusion, or waste of time. This balancing test is particularly important for generative AI evidence, which may optically appear highly authoritative due to its technological sophistication. Judges must carefully weigh whether the probative value of AI-generated evidence outweighs the risk that jurors will give it undue weight or be confused by its complexity. When authenticating AI-generated evidence, attorneys should be prepared to address the following questions about training data: What data was used to train the AI system? How was this data obtained? Why was it chosen? Where did it come from? What features and weights were chosen for the data? Why were they chosen? Similarly, authentication of AI algorithms should address who programmed the underlying algorithm and how the program was developed. Did the collected data affect how the AI system was programmed? How was the data collected? How is the AI system used? Does it produce valid and reliable results? What was done to determine the validity and reliability of results? What are the chances of error? Do the benefits of the results outweigh the possible errors? Could the possible errors mislead the finder of fact?

As litigators, our responsibility is to develop sufficient technical literacy to effectively authenticate, challenge, and contextualize AI evidence while ensuring that judges and juries understand both its potential value and limitations. This requires ongoing education, collaboration with technical experts, and careful attention to emerging caselaw. As generative AI technology continues to evolve, so too will the legal standards governing its use in litigation. By staying informed about both technological developments and emerging legal standards, Texas attorneys can effectively harness the power of generative AI while ensuring that such evidence meets the fundamental requirements of authenticity and reliability.

Notes

  1. Tex R. Ev. 901(a) (“To satisfy the requirement of authenticating or identifying an item of evidence, the proponent must produce evidence sufficient to support a finding that the item is what the proponent claims it is”).
  2. Tex. R. Ev. 602 (“A witness may testify to a matter only if evidence is introduced sufficient to support a finding that the witness has personal knowledge of the matter. Evidence to prove personal knowledge may consist of the witness’s own testimony. This rule does not apply to a witness’s expert testimony under Rule 703”).
  3. E.I. du Pont de Nemours & Co., Inc. v. Robinson, 923 S.W.2d 549, 556 (Tex. 1995) (“Therefore, we hold that in addition to showing that an expert witness is qualified, Rule 702 also requires the proponent to show that the expert’s testimony is relevant to the issues in the case and is based upon a reliable foundation.”).
  4. Tex. R. Ev. 702 (“A witness who is qualified as an expert by knowledge, skill, experience, training, or education may testify in the form of an opinion or otherwise if 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.”).
  5. Tex. R. Ev. 401 (“Evidence is relevant 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 of consequence in determining the action.”).

Headshot Heather KingHEATHER L. KING is a shareholder in KoonsFuller in Fort Worth, where she practices family law. She is a past president of the Texas Academy of Family Law Specialists and a past president of the Tarrant County Bar Association and the Tarrant County Family Law Bar Association. King was the recipient of the 2023 TexasBarCLE Gene Cavin Award. She is certified in family law by the Texas Board of Legal Specialization.

Headshot Tom DaleyTOM DALEY is a shareholder in KoonsFuller and certified in family law by the Texas Board of Legal Specialization. In addition to practicing family law, Daley researches and writes about the application of artificial intelligence to the practice of law and teaches legal professionals how to use AI ethically, responsibly, and effectively.