Legal Evidence AI Verification

Provenance comes first. FauxLens reads C2PA content credentials and EXIF/XMP metadata. A credential or generator signature from an AI tool ends the scan, and an authentic camera record can clear a photo. Missing metadata is never treated as evidence of anything, because messaging apps and social platforms routinely strip it from files.

Next, FauxLens reads the invisible watermark that OpenAI and Google embed in the images their AI tools produce (SynthID). A detected watermark identifies the vendor family only, not the specific model, account, or prompt that created the image, so it cannot by itself show who made a piece of evidence.

A specialist third-party detection model then scores the pixels and judges whether the image is real, partly changed by AI, or generated by AI. The report turns that into one of four verdicts (AI-generated, likely AI-generated, likely authentic, or no AI detected), with an evidence line when only part of the image was changed. This is the layer that addresses images from Midjourney, Stable Diffusion, Flux, and other tools that carry no credential or watermark.

If the specialist model cannot return a result, a multimodal vision model reviews the image as a fallback and describes any visible signs of generation or editing it finds.

Techniques such as Error Level Analysis (ELA), PRNU sensor noise matching, Fourier frequency analysis, and shadow geometry belong to a certified forensic examiner's toolkit. FauxLens does not perform them, which is one reason its result is a screening input rather than an expert finding.

The EXIF metadata check looks for AI generator signatures and camera records; it does not test GPS coordinates or camera model claims against the scene, so those questions still need an examiner.

The output is a verdict, a confidence score that is not a calibrated probability, and plain-language evidence lines. Treat it as supporting evidence that can justify requesting expert examination, not as expert testimony.

FauxLens is appropriate for: preliminary evidence review at intake, identifying which submitted items warrant deeper investigation, supporting, alongside other grounds, a discovery request that questions evidence authenticity, and providing a starting brief for a retained forensic expert. A FauxLens report that flags an image as high-confidence AI-generated gives your forensic expert a documented starting point and shows which check (content credentials, watermark, or detection model) produced the finding.

FauxLens is not appropriate as the sole basis for: sworn expert testimony, a motion in limine excluding evidence, or any argument where the authenticity determination will be contested by opposing counsel with their own expert. For these purposes, engage a certified digital forensic examiner (someone credentialed by ISFCE (Certified Computer Examiner), IACRB (Certified Computer Forensics Examiner), or holding EnCE or ACE certification) who can apply proper chain-of-custody procedures, document their methodology to NIST standards, and provide sworn testimony subject to cross-examination.

The criteria for escalating from FauxLens to a certified examiner are: the evidence will be formally presented in court; opposing counsel has retained their own forensic expert; chain-of-custody documentation is required for the evidence to be admissible; the FauxLens result is inconclusive but the evidence is high-stakes; or the case involves criminal charges where the standard of proof demands the strongest possible authentication.

At the federal level, existing rules of evidence, primarily FRE 901 (authentication) and FRE 902 (self-authentication), apply to AI-generated evidence without specific AI carve-outs. The burden remains on the proponent to authenticate evidence and on the challenger to demonstrate inauthenticity. Even so, multiple federal courts have issued local rules specifically addressing AI disclosure. The Southern District of New York, the Northern District of California, and the Fifth Circuit have all issued standing orders or local rules requiring disclosure when AI was used in the preparation of court filings or evidence. Violations have resulted in sanctions ranging from cost awards to disciplinary referrals.

At the state level, the patchwork is significant. Texas enacted HB 4337 in 2025, requiring parties in civil proceedings to disclose when AI tools were used to generate or materially alter any submitted photograph, video, or document. Colorado SB 205 (2025) established authentication requirements for AI-generated evidence similar to those for other electronic records. California AB 2602 (2024) required disclosure of AI-generated media in political advertising and created a private right of action for undisclosed AI election content, establishing precedents that civil litigators are citing in non-election contexts.

The American Bar Association issued Formal Opinion 512 in 2023 addressing attorney competency obligations in using AI tools, including for evidence preparation. The opinion establishes that attorneys have a duty of competence that includes understanding the limitations of AI tools they use and supervise. Using AI to prepare visual evidence without verification creates potential professional responsibility exposure.

For practitioners: treat AI evidence disclosure as a prophylactic measure regardless of jurisdiction. Document your forensic screening process. Retain FauxLens reports as part of your evidence file. When opposing counsel produces visual evidence, a forensic challenge grounded in documented analysis is increasingly available and increasingly accepted.

At intake, run FauxLens on every submitted photograph and video segment before any attorney reviews the evidence for substantive content. This prevents anchoring bias, the tendency to interpret ambiguous forensic signals in light of the story you already believe the evidence tells. Document the FauxLens report in your file management system alongside the original evidence file.

Preserve the original in unaltered form. Before any processing, compute a cryptographic hash (SHA-256) of the original evidence file and record it with a timestamp. This hash is the foundation of chain-of-custody documentation - any modification to the file will produce a different hash, proving tampering. Use a forensic duplicate tool (FTK Imager, Autopsy, or equivalent) rather than copying the file manually.

When FauxLens returns an AI-generated or likely AI-generated verdict on any item, escalate to a DFIR (digital forensics and incident response) professional before proceeding. The FauxLens report (the verdict, the confidence score, and the evidence lines showing which check produced it) becomes the brief for the expert's deeper analysis. This structured handoff is more efficient than a general request for "forensic analysis" and produces a more focused expert report.

For evidence you are producing, submitting photographs or videos in support of your client's case, verify your own evidence with the same rigor you apply to opposing evidence. A likely authentic FauxLens result on your own submissions is one documented part of your due diligence, not a defense to an authenticity challenge on its own. Retain the report. If the result is unexpectedly high-confidence AI-generated on evidence you believe is authentic, that is your signal to investigate the source before submission.

Align your workflow with NIST Special Publication 800-86 (Guide to Integrating Forensic Techniques into Incident Response) for the documentation standards that hold up under cross-examination. Courts increasingly expect digital evidence handling to follow recognized standards, and SP 800-86 provides the benchmark most forensic experts cite.