Artificial intelligence now touches nearly every category of evidence that appears in civil and criminal litigation. AI systems transcribe audio, classify images, flag anomalies in financial records, generate documents, and authenticate or challenge digital media. When any of these outputs become evidence, attorneys face a set of technical and procedural questions that courts are only beginning to address systematically.
The Authentication Problem
Federal Rule of Evidence 901 requires that evidence be authenticated before it is admitted. For traditional documents, authentication is straightforward. For AI-generated or AI-processed evidence, the question is more complex: what does it mean to authenticate an output that was produced by a system whose internal operations are not fully transparent?
Courts have begun to address this question in several contexts. In cases involving AI-generated transcripts, authentication requires not only that the recording was made but that the transcription model performed reliably on the specific audio in question. Word error rates vary significantly across speakers, accents, audio quality, and domain-specific vocabulary. A transcript produced by a general-purpose speech recognition model may be systematically inaccurate for certain speakers or recording conditions, and that inaccuracy may not be apparent from the face of the document.
AI-processed images present similar issues. When an AI system has been used to enhance, denoise, or upscale an image, the output is not a faithful reproduction of the original. It is a reconstruction, and the reconstruction reflects choices embedded in the model's training data and architecture. Authenticating such an image requires understanding what the model did and what it may have introduced or removed.
Chain of Custody for AI-Processed Evidence
Chain of custody requirements apply to AI-processed evidence as they do to any other evidence, but the chain is more complex. For physical evidence, the chain documents who had possession of the item and when. For AI-processed evidence, the chain must also document the version of the AI system used, the input data, the processing parameters, and the output. If any of these elements are not documented, the reliability of the output cannot be independently assessed.
This is particularly relevant in criminal cases where AI tools have been used in the investigative process. Law enforcement agencies increasingly use AI-powered tools for facial recognition, license plate reading, gunshot detection, and predictive analytics. When these tools produce outputs that inform an investigation or prosecution, the chain of custody for those outputs must be established with the same rigor as physical evidence.
Key Authentication Questions for AI Evidence
- What AI system produced or processed the evidence, and what version was used?
- What was the input data, and how was it obtained and preserved?
- What processing parameters were applied, and by whom?
- Has the system's reliability been validated for this type of input?
- Is the output reproducible from the same inputs?
The Deepfake Defense
As synthetic media technology has become more accessible, defendants in criminal cases have begun raising what practitioners call the "deepfake defense": the argument that video or audio evidence purportedly showing the defendant is AI-generated or AI-manipulated. This defense has been raised in cases involving surveillance footage, recorded conversations, and social media content.
The deepfake defense creates a significant evidentiary challenge. Prosecutors must now be prepared to affirmatively establish the authenticity of video and audio evidence, not merely produce it. This requires technical analysis by a qualified expert who can assess the evidence for signs of AI manipulation and explain the methodology and its limitations to the factfinder.
Defense counsel faces the mirror-image challenge: establishing that evidence may have been manipulated requires more than pointing to the general existence of deepfake technology. Courts have been skeptical of deepfake defenses that are not supported by specific technical analysis of the evidence at issue.
AI-Generated Documents and the Hearsay Framework
AI-generated documents raise questions that the existing hearsay framework was not designed to address. When a large language model generates a document, there is no human declarant whose out-of-court statement is being offered. Courts have not yet reached consensus on whether AI-generated text constitutes hearsay, and the answer may depend on how the system was used and what the output is being offered to prove.
More practically, AI-generated documents raise reliability questions that go beyond the hearsay framework. Large language models can produce plausible-sounding but factually incorrect text, including fabricated citations, invented statistics, and inaccurate descriptions of events. When AI-generated documents are offered as evidence of the facts they describe, the reliability of the underlying model is directly at issue.
Expert Testimony and Rule 702
When AI evidence is contested, expert testimony is typically required to explain the technical issues to the factfinder. Under Federal Rule of Evidence 702 as amended in 2023, the proponent of expert testimony must demonstrate by a preponderance of the evidence that the expert's opinion reflects a reliable application of reliable principles and methods to sufficient facts or data.
For AI-related expert testimony, this standard requires that the expert be able to explain not only the general principles of the AI system at issue but also how those principles were applied in the specific case. An expert who can describe how facial recognition systems work in general but cannot address the specific system, dataset, and conditions at issue in the case may not satisfy the Rule 702 standard.
Practical Implications for Attorneys
Attorneys handling cases where AI evidence may be at issue should consider several practical steps. Early in the case, identify all AI systems that may have produced or processed evidence relevant to the matter. For each system, seek documentation of the version used, the training data, the validation methodology, and any known limitations or error rates. Preserve the original inputs and outputs, not just the final processed evidence.
In discovery, request documentation of any AI systems used by the opposing party in the investigation, analysis, or preparation of evidence. This includes not only dedicated forensic AI tools but also general-purpose AI tools that may have been used to process, summarize, or analyze documents.
Retain a qualified technical expert early. The technical issues in AI evidence cases are often complex enough that early expert involvement can shape discovery strategy, motion practice, and trial preparation in ways that are difficult to replicate if the expert is retained only for trial.
AI Expert Witness Services provides technical expert support for attorneys handling matters involving AI evidence, authentication challenges, and digital forensics.
Discuss Your Matter