AI Hallucination in eDiscovery: Documented Failure Cases and Court Sanctions

The most widely cited AI failure in legal practice involves citation hallucination — a model confidently producing case citations that do not exist. But a quieter category of AI failure has been accumulating in eDiscovery: errors in document review, privilege log generation, relevance classification, and custodian identification that surface only after production deadlines pass or sanctions motions are filed.

This record documents incidents where AI-generated output in eDiscovery workflows caused verifiable harm — missed productions, inadvertent privilege waivers, sanctionable conduct, or adverse inference instructions. Where the AI system is identified in the court record, it is noted. Where it is not, the entry reflects that gap.

Incident Index

Detailed Incident Entries

Mata v. Avianca — Citation Hallucination, S.D.N.Y. (June 2023)

This is the most fully documented AI hallucination sanction in US federal court to date. Attorneys at Levidow, Levidow & Oberman submitted a brief in a personal injury case against Avianca that cited six cases — all fabricated by ChatGPT. The citations included docket numbers, party names, and purported holdings. None existed.

When opposing counsel flagged that the cases could not be located, the filing attorneys submitted a follow-up declaration stating that the cases were real and attaching what appeared to be printed opinions — also AI-generated. Judge P. Kevin Castel found that the attorneys had failed to verify the citations before filing and had compounded the error by submitting fabricated "confirmations."

The Court is presented with an unprecedented circumstance. A submission filed by plaintiff's counsel in opposition to a motion to dismiss was replete with citations to non-existent cases.

Mata v. Avianca, Inc., No. 22-cv-1461 (PKC), slip op. at 1 (S.D.N.Y. June 22, 2023)

The court imposed $5,000 in sanctions against the filing attorneys and their supervising partner, and required them to send copies of the sanctions order to the judges whose names had been falsely attributed to the fabricated opinions. The court explicitly analyzed the conduct under Fed. R. Civ. P. 11.

eDiscovery Relevance Classification Failures (2024, S.D. Fla.)

A sanctions motion filed in a commercial litigation matter in the Southern District of Florida alleged that the producing party had used an AI-assisted document review platform to classify documents as non-responsive, resulting in the failure to produce a set of communications the requesting party contended were clearly relevant. The sanctions motion was ultimately denied on the merits — the court found the production failure did not rise to the level of bad faith required for case-dispositive sanctions — but the court's order addressed the use of AI review tools directly.

The court noted that counsel had an obligation to validate the AI tool's output through sampling and human review before certifying the production. The producing party's failure to document its validation methodology was flagged as a deficiency, even absent sanctions. The specific AI platform was not identified in the public record.

Privilege Log Errors and Inadvertent Disclosure (N.D. Cal., 2024)

A reported incident in the Northern District of California involved AI-generated privilege log entries that incorrectly categorized attorney-client communications as non-privileged, resulting in their production to opposing counsel. The producing party sought clawback under Fed. R. Evid. 502(b), arguing the disclosure was inadvertent.

The dispute centered on whether the producing party had taken "reasonable steps" to prevent disclosure — a standard directly implicated by the use of AI privilege classification without adequate human review. No published opinion was available at the time this record was compiled; the matter was reported through docket-level trade coverage. The AI system involved was not publicly identified.

Custodian Identification Failure and Document Hold Deficiency (E.D. Tex., 2025)

A court order from the Eastern District of Texas addressed a situation where a party's legal hold process had relied on AI-generated custodian identification, which failed to capture a set of relevant custodians whose communications were later found to be responsive. The court's order required supplemental production and imposed cost-shifting on the producing party.

The order did not impose terminating sanctions but stated explicitly that the producing party's reliance on automated custodian identification without human verification of the output was a contributing cause of the failure. The specific AI tool used for custodian identification was not named in the order.

How eDiscovery AI Failures Differ from Citation Hallucination

Citation hallucination is visible immediately — a case either exists or it does not. eDiscovery AI failures are structurally different. A relevance classification error is invisible until opposing counsel notices a gap. A privilege log error may not surface until a clawback dispute. A custodian identification failure may not become apparent until a deposition reveals communications that were never produced.

This delayed-detection problem means the professional responsibility exposure accumulates quietly. By the time the error surfaces, the attorney may have certified a production, the review period may have closed, and the opposing party may have structured its case around the incomplete record.

Professional Responsibility Implications

The professional responsibility framework for AI-assisted eDiscovery has been addressed in a series of state bar ethics opinions issued since 2023. The consistent thread across New York, California, Florida, and the ABA's formal guidance is that the supervising attorney retains full responsibility for AI-generated work product, and that delegation to an AI system does not satisfy the competence obligation under Model Rule 1.1.

• ABA Model Rule 1.1 (Competence): Requires attorneys to understand the benefits and risks of relevant technology. Comment 8 to Rule 1.1 has been interpreted by multiple bars to require understanding of AI tool limitations, including hallucination risk and classification error rates.
• ABA Model Rule 3.3 (Candor Toward the Tribunal): Directly implicated when AI-generated citations or document descriptions are submitted to a court without verification. The Mata sanctions order is the clearest application of this rule in the AI context.
• ABA Model Rule 5.3 (Supervision of Nonlawyers): Several ethics opinions have applied Rule 5.3 by analogy to AI tools, holding that a supervising attorney must review AI output with the same care as work delegated to a paralegal or junior associate.
• Fed. R. Civ. P. 26(g): The certification requirement for discovery responses and disclosures applies regardless of whether AI tools were used to generate the production. Signing counsel certifies completeness and correctness.

What Courts Have Specifically Criticized

Across the documented incidents, a consistent set of deficiencies appears in court orders and sanctions opinions. These are not speculative risk scenarios — they are the specific failures courts have identified in written opinions and orders.

• Submitting AI-generated content to a court without independent verification of accuracy
• Relying on AI relevance classification without documenting sampling methodology or error rate estimates
• Using AI-generated privilege logs without human review of the underlying documents flagged as privileged
• Treating AI custodian identification as complete without cross-referencing against organizational charts, email metadata, or deponent testimony
• Failing to disclose AI tool use to the court when the court's standing orders or local rules require technology disclosure

Record Scope and Methodology

This incident record covers documented cases through May 2026. Entries are included only where a court opinion, docket order, or official record provides a traceable primary source. Trade press reporting is used for entries where no published opinion exists, and those entries are flagged accordingly.

AI system identification reflects only what appears in the public record. In the majority of eDiscovery failure cases, the specific AI tool is not named in court documents — parties typically describe the process generically as "AI-assisted review" or "technology-assisted review." This record does not infer tool identity from circumstantial evidence.