How to Challenge AI-Generated Evidence in California Courts Under the California Evidence Code and Due Process Rights
Californians can challenge AI-generated evidence by attacking authentication, relevance, reliability, and prejudice under the California Evidence Code, and by invoking due process when the defense cannot meaningfully test the system. As AI outputs appear in criminal and civil cases—from deepfakes to automated “risk” scores—courts are being asked to decide what qualifies as trustworthy proof. This article explains the strongest California objections, motions, and hearing strategies to exclude or limit AI evidence.
What Counts as “AI-Generated Evidence” in California Court?
“AI-generated evidence” is not a defined statutory category in the California Evidence Code. In practice, it describes any exhibit, testimony, or data output created or materially altered by machine-learning or generative systems. Common courtroom examples include:
Deepfake video or audio (synthetic media purporting to show a person speaking or acting); AI-enhanced images (resolution “upscaling,” face enhancement, object removal); automated risk or threat scores (pretrial risk assessments, fraud scoring, “propensity” outputs); AI-written reports (summaries drafted by a model based on records); and pattern-recognition outputs (voice identification, gait analysis, or other classifier results).
The core litigation issue is usually the same: whether the proponent can establish a reliable foundation that the AI output is what it claims to be and whether its probative value is outweighed by the risks of confusion, unfair prejudice, or misleading the factfinder.
The Governing Framework: California Evidence Code + Constitutional Due Process
Challenging AI-generated evidence in California typically proceeds on two tracks:
(1) Statutory evidence objections under the California Evidence Code (e.g., authentication, relevance, hearsay, expert foundation, and Evidence Code section 352); and
(2) Due process arguments—especially in criminal cases—when the defense cannot meaningfully test or confront the AI system’s reliability due to opacity, trade-secret claims, withheld training data, or lack of access to model documentation.
Courts are often receptive to targeted, rule-based challenges. Even when evidence is not excluded entirely, judges may restrict the scope of AI evidence, require limiting instructions, or order discovery that exposes flaws.
Step One: Demand Proper Authentication (Evidence Code §§ 1400–1402)
Authentication is frequently the fastest path to exclusion or significant limitation. Under Evidence Code section 1400, the proponent must produce evidence sufficient to sustain a finding that the writing (broadly defined) is what it purports to be. For AI content, this means more than “it came from a computer.”
Authentication problems unique to AI outputs
Unknown provenance: Who generated the output? What prompt, settings, model version, and input materials were used? Was it edited? Without a chain of custody and process description, authenticity is weak.
Model/version drift: Many AI tools update frequently. If the exhibit was created on “Version X” but tested on “Version Y,” the proponent may not be able to reproduce results—undermining authentication.
Undetectable manipulation: AI-generated media can appear “real” but be entirely synthetic. A witness’s “it looks like the defendant” is not necessarily authentication of origin.
Practical objections and foundational demands
In deposition and at trial, force the proponent to establish:
1) the identity and competence of the operator; 2) the system used (vendor, model, version); 3) the precise input data; 4) the prompts/settings; 5) the workflow and any post-processing; 6) storage, integrity checks, and chain of custody; and 7) reproducibility.
If these pieces are missing, object for lack of authentication under Evidence Code sections 1400–1402 and, where appropriate, argue that the proponent’s foundation is speculative.
Step Two: Attack Relevance and Speculation (Evidence Code §§ 210, 350)
AI evidence is often offered with an implied claim: “the model says it, therefore it’s true.” Relevance under Evidence Code section 210 requires a tendency in reason to prove or disprove a disputed fact, and section 350 bars irrelevant evidence.
Challenge the logical link between the AI output and the ultimate fact. Examples:
AI-generated “summary” of records: If the model’s summary is not tied to admissible source documents, it may be irrelevant or misleading; the underlying records—not the model’s narrative—are the probative evidence.
Risk or propensity score: A score may be untethered to the legal standard at issue (e.g., “risk of reoffense” offered to prove intent in a specific incident). Argue that it invites improper character reasoning and lacks a valid relevance theory.
Step Three: Use Hearsay Rules Against AI Narratives (Evidence Code §§ 1200 et seq.)
Many AI outputs contain assertions—about what happened, what someone “intended,” or what a dataset “shows.” If offered for their truth, they implicate hearsay (Evidence Code section 1200).
Key distinctions:
Machine-generated vs. human statements: Traditional hearsay focuses on “statements” by a “person.” AI complicates this because a model’s output may not be a “person’s” statement—yet it still functions like an out-of-court assertion. Expect litigation over characterization.
Workaround attempts: Proponents may offer AI output as “not for truth,” or as a “business record.” Challenge whether the output is genuinely routine, whether it was created in anticipation of litigation, and whether the method and time of preparation indicate trustworthiness.
Even when the court is uncertain whether classic hearsay doctrine fits, you can still press reliability and section 352 arguments to prevent an untested AI narrative from becoming a substitute for admissible proof.
Step Four: Force Expert Standards—Opinion Foundation, Methodology, and Kelly/Frye Issues
AI evidence frequently comes in through an expert. California permits expert opinion testimony when it will assist the trier of fact, but it must rest on a reliable basis and not be speculative. Where the AI method is novel or scientific, California’s Kelly test (often described as the Kelly/Frye framework) may apply, requiring general acceptance in the relevant scientific community.
When to argue Kelly applies
Consider a Kelly challenge where the proponent relies on a scientific technique that is new, opaque, or presented as determinative—such as AI-based voice identification, automated “deepfake detection,” or proprietary predictive analytics offered as scientific truth.
Cross-examination targets for AI-driven expert opinions
Even outside a full Kelly hearing, attack the opinion’s foundation:
Training data and bias: What data trained the model? Does it underperform on certain demographics? Was it validated on similar populations and conditions?
Error rates and confidence: What is the known false-positive/false-negative rate? Were thresholds chosen scientifically or for business convenience?
Validation and reproducibility: Can the expert reproduce results? Are there independent studies? Or is the expert relying on vendor marketing?
Human-in-the-loop editing: If a person curated inputs or selected the “best” output, the process becomes subjective—undercutting claims of objectivity.
Step Five: Evidence Code § 352—Your Most Flexible Tool Against AI Hype
Evidence Code section 352 allows exclusion when probative value is substantially outweighed by the probability of undue prejudice, confusing the issues, misleading the jury, undue consumption of time, or cumulative presentation.
AI evidence is particularly vulnerable under section 352 because jurors may overvalue a “computer says so” exhibit. Concrete section 352 arguments include:
Misleading aura of certainty: Generative tools can “hallucinate” plausible but false details; predictive scores can be mistaken for proof of conduct.
Mini-trial risk: Litigating model training, validation, and error rates can consume excessive time relative to marginal relevance.
Inflammatory impact: A realistic deepfake or AI-enhanced image may provoke emotional reactions disproportionate to its reliability.
Ask for exclusion or, at minimum, limitations (e.g., no numerical “risk score” displayed to jurors; descriptive testimony only; or admission conditioned on production of validation materials).
Step Six: Deepfakes and Synthetic Media—Specific Authentication and Integrity Attacks
When a party offers video/audio that could be synthetic, shift the burden to rigorous foundation. Useful litigation steps:
1) Chain-of-custody and metadata: Demand original files, device identifiers, creation dates, codec information, and transfer logs. Gaps support authenticity objections.
2) Forensic examination: Seek court orders permitting an independent forensic expert to examine native files for editing traces, compression artifacts, and generation markers.
3) “Best evidence” instincts: While California’s secondary evidence rule is nuanced, judges still respond to the common-sense point: a copy of a copy—especially one processed through AI—creates integrity risks.
Example: In a workplace harassment civil case, the plaintiff offers an audio clip purportedly capturing a supervisor’s threats. Defense learns it was “cleaned up” with an online AI tool. Move to exclude for lack of authentication and section 352, and request an evidentiary hearing requiring the proponent to establish the full enhancement workflow and produce the unaltered original for comparison.
Step Seven: Discovery and Motions to Compel—Don’t Litigate Blind
AI disputes are often won in discovery. If the other side plans to rely on AI outputs, pursue:
Requests for production: model cards, validation studies, accuracy testing, bias audits, prompt logs, version history, vendor contracts, and user manuals; underlying datasets used in the case; and all intermediate outputs.
Interrogatories
