Artificial intelligence systems increasingly influence decisions that may later be challenged in court or investigated by regulators. Prompts, outputs, model versions, system instructions, audit logs, and human-review records can become material evidence.
Organizations use artificial intelligence systems to create, rank, recommend, predict, and document decisions that may later face litigation or investigation.
The question is whether evidence retention has become a central AI compliance function. U.S. federal civil litigation is the primary forum for this analysis. European Union law supplies a limited comparison. The answer is yes, with limits.
Federal preservation law does not require indefinite retention of every AI record. It requires reasonable preservation of relevant information when litigation becomes reasonably anticipated. The process must also respect proportionality, privilege, confidentiality, privacy, and applicable retention limits.
Executive Summary
- [U.S. federal] Rules 26, 34, and 37 apply to relevant AI records as electronically stored information. The rules contain no AI exception.
- [U.S. federal] Preservation begins when litigation is reasonably anticipated or pending. Routine schedules may operate before that trigger.
- [U.S. federal] A hold notice alone is not enough. The organization must stop relevant deletion, preserve sources, monitor compliance, and test results.
- [U.S. federal] Rule 37(e)(1) permits proportionate cures for prejudicial loss. Rule 37(e)(2) reserves severe sanctions for intent to deprive.
- [U.S. federal] Relevance, proportionality, accessibility, privilege, and protective orders limit discovery. Preservation does not make every retained record producible.
- [Evidence] Model, configuration, provenance, and audit records may support authentication under Rules 901 and 902.
- [European Union] Regulation (EU) 2024/1689 imposes logging and documentation duties for covered high-risk systems when the relevant provisions apply.
- [Operational] A defensible program needs an AI record map, matter triggers, system holds, vendor controls, preservation testing, and documented release.
AI records are electronically stored information
Relevant AI records fall within Rule 34’s broad coverage of electronically stored information. Rule 34 covers writings, images, other data, and data compilations stored in any medium. Fed. R. Civ. P. 34(a)(1)(A). It does not distinguish a model output from another electronically stored business record.
Relevance depends on the claims and defenses. A discrimination case may turn on inputs, ranking outputs, version history, and human review. A contract case may turn on prompts, generated text, approvals, and delivery records. A product case may turn on test results, incident logs, corrective changes, and warnings. These categories are not automatic preservation requirements. They identify likely sources when the disputed decision involved the system.
Other relevant sources may include system instructions, model identifiers, configuration records, retrieval sources, timestamps, access logs, and review history. Training data, model weights, and source code require separate treatment. They belong in a hold only when the issues make them relevant and preservation is reasonable.
Rule 34 reaches records in a party’s possession, custody, or control. Vendor hosting does not remove a record from scope when governing law and contractual rights place it within the party’s control. Fed. R. Civ. P. 34(a)(1). The governing forum’s interpretation and the contract therefore matter.
The preservation duty is event-driven
Federal law does not impose indefinite litigation retention merely because an organization uses AI. Rule 37(e) addresses information that should have been preserved in anticipation or conduct of litigation. Fed. R. Civ. P. 37(e). The trigger remains fact-specific.
A demand, incident report, regulator inquiry, internal complaint, or contractual dispute may make litigation reasonably anticipated. The event must be assessed against the claim’s credibility, specificity, and likely forum. A remote possibility does not require a perpetual hold.
Once the duty attaches, the organization should identify the disputed conduct, likely claims, custodians, systems, vendors, dates, versions, and deletion paths. The hold should interrupt routine deletion for relevant sources. Unaffected information may remain under ordinary schedules, unless another law or contract requires retention.
The organization should revisit scope as the dispute develops. New claims may add systems or custodians. Dismissed claims may narrow the hold. The organization should release the hold when the matter ends and no related duty remains.
Volatile systems require system-level action
Short retention periods and user-controlled deletion require technical action. A notice to employees cannot preserve records that administrators leave subject to automatic deletion.
In Google Play, relevant chats disappeared when users left history off. Google left preservation decisions to employees and did not monitor compliance. The court found that Google failed to take reasonable steps. It also found prejudice and intent, then authorized a permissive adverse inference. In re Google Play Store Antitrust Litigation, 664 F. Supp. 3d 981, 992–95 (N.D. Cal. 2023); In re Google Play Store Antitrust Litigation, No. 21-md-02981-JD, Dkt. 700, at 3–4 (N.D. Cal. Oct. 20, 2023).
Where a platform permits centralized retention, the organization should use that control for relevant accounts and channels. It should not rely only on users to recognize legal relevance. Where centralized holds are unavailable, administrators should export or snapshot relevant content before expiration. They should record the limitation, method, date, and validation result.
Skanska confirms that a hold without implementation may support severe sanctions. In-house counsel initiated an active hold after a known incident. The company did not preserve known custodians’ phone data and allowed routine deletion. The Eleventh Circuit upheld the bad-faith finding. It also held that failure to act can violate Rule 37(e). Skanska USA Civil Southeast, Inc. v. Bagelheads, Inc., 75 F.4th 1290, 1311–15 (11th Cir. 2023).
Routine deletion did not excuse the loss in Skanska. The decision does not require perfect preservation. It requires reasonable action directed at known sources and deletion risks.
Taylor prevents overgeneralization. A company-wide chat problem in another case did not prove relevant loss in the case before the court. The movants needed a matter-specific factual basis. Taylor v. Google LLC, No. 20-cv-07956-VKD, Dkt. 160, at 4–7 (N.D. Cal. Dec. 3, 2024). Prior failures may justify focused inquiry. They do not establish automatic sanctions or unlimited discovery.
Rule 37(e) calibrates sanctions
Rule 37(e) has three threshold conditions. Relevant ESI should have been preserved. The party failed to take reasonable steps. The lost information cannot be restored or replaced through further discovery. Fed. R. Civ. P. 37(e).
If loss prejudices another party, Rule 37(e)(1) permits measures no greater than necessary to cure that prejudice. If the party intended to deprive another party, Rule 37(e)(2) permits a presumption, adverse instruction, dismissal, or default.
Skanska treats intent to deprive as bad faith in the Eleventh Circuit. Negligence, including gross negligence, is not enough for Rule 37(e)(2). 75 F.4th at 1311–12. Other courts apply the federal rule within their own circuit precedent.
A missing AI artifact does not automatically support an adverse inference. A vendor archive, audit store, recipient copy, or other source may restore or replace it. The substitute must preserve the material information, not merely provide a partial account.
Reasonable steps depend on system design, cost, accessibility, notice, and the information’s expected importance. A documented decision may defeat an inference of indifference. A paper hold with no technical execution will not.
Preservation does not equal production
Preservation protects potentially discoverable information from loss. Production remains limited by relevance, proportionality, accessibility, privilege, and court orders.
Rule 26(b)(1) limits discovery to nonprivileged matter relevant to a claim or defense and proportional to the case. Rule 26(b)(2)(B) addresses sources that are not reasonably accessible because of burden or cost. Fed. R. Civ. P. 26(b)(1), 26(b)(2)(B).
A company may preserve a full log segment yet produce only relevant fields and time periods. Model weights, training data, and source code should not be demanded or produced by default. Their relevance, availability, burden, and confidentiality require separate analysis.
Rule 26(c)(1)(G) permits protection for trade secrets and confidential research, development, or commercial information. Rule 26(b)(5) supplies procedures for privilege claims. A Rule 502(d) order can protect against waiver from disclosure in the federal proceeding. Fed. R. Civ. P. 26(b)(5), 26(c)(1)(G); Fed. R. Evid. 502(d).
Rule 26(f) requires parties to discuss preservation and develop an ESI discovery plan. An AI-specific protocol should identify systems, log fields, version identifiers, native formats, retention settings, and export methods. It should also state disputed sources and proposed safeguards.
Vendor architecture must support legal holds
AI services often place evidence outside the customer’s direct infrastructure. Contracts and technical design determine whether the customer can preserve, retrieve, and authenticate that evidence.
At procurement and renewal, legal and technical teams should address seven functions. The provider should suspend deletion for a defined legal hold. It should provide metadata-preserving exports. It should identify model, configuration, and log versions. It should give advance notice of deletion or material format changes. It should bind relevant subprocessors. It should report incidents affecting retained data. It should provide reasonable preservation and certification assistance.
The agreement should also identify default retention periods and available hold settings. A service that keeps prompts for thirty days creates a known deadline. A contract that forbids export may create a proof problem before any dispute begins.
Vendor possession does not always equal the customer’s control under Rule 34. The answer depends on forum law and the customer’s rights. The inability may still affect Rule 37(e)’s reasonable-steps analysis if the customer created or ignored the limitation after the duty attached.
A business using a service with no hold or export function should capture critical records locally for uses with material legal exposure. If local capture is infeasible, the business should document the limitation and decide whether the use should continue.
Provenance supports authentication and merits proof
Evidence retention must preserve enough provenance to connect a record with the disputed system and event. Saving only a screenshot may omit facts needed to authenticate or interpret it.
Rule 901(a) requires evidence sufficient to support a finding that an item is what its proponent claims. Rule 901(b)(9) recognizes evidence describing a process or system that produces an accurate result. Fed. R. Evid. 901(a), 901(b)(9).
When a party offers an AI output, material provenance may include the original output, prompt, timestamp, system instruction, model identifier, configuration, retrieved sources, and human changes. The proper set depends on the proposition for which the output is offered.
Rule 902(13) permits certification of records generated by an electronic process or system. Rule 902(14) addresses copied electronic data authenticated through digital identification. Fed. R. Evid. 902(13)–(14). Native exports, hashes, chain-of-custody records, and qualified certifications can support those routes.
Authentication does not prove substantive accuracy. An authentic output may remain unreliable, irrelevant, privileged, hearsay, or unfairly prejudicial. Provenance records permit the parties and court to address those separate questions.
Privacy, privilege, and trade secrecy constrain retention and access
The presence of personal data is not a reason to destroy relevant evidence after a legal duty attaches. Litigation is also not a reason to retain every user interaction forever.
The GDPR requires storage limitation. It also limits the right to erasure when processing is necessary for the establishment, exercise, or defence of legal claims. Regulation (EU) 2016/679, arts. 5(1)(e), 17(3)(e). These provisions support targeted legal holds, not indefinite blanket retention.
A cross-border hold should define purpose, scope, access, location, security, and release. Separate analysis may be required before EU personal data is transferred for U.S. discovery. Regulation (EU) 2016/679, arts. 44–49.
Privilege and work-product protection require separate controls. Legal advice in prompts or review notes should be segregated where feasible. Access should be limited to personnel with a matter role. Production review should use Rule 26(b)(5) and, where appropriate, a Rule 502(d) order.
Trade-secret concerns also affect access and production. Model details, source code, testing records, and security data may require tiered confidentiality terms. A protective order can restrict use, recipients, storage, and return or destruction.
The EU AI Act creates independent records for covered high-risk systems
Regulation (EU) 2024/1689 creates specific record duties for covered high-risk AI systems. Article 12 requires technical capability for automatic event logging over the system’s lifetime. Article 18 requires providers to keep listed documentation for ten years after the system is placed on the market or put into service. Articles 19 and 26(6) require providers and deployers to keep controlled logs for an appropriate period of at least six months, subject to other applicable law. Regulation (EU) 2024/1689, arts. 12, 18, 19, 26(6).
These periods are not universal retention rules for all AI. Coverage depends on system classification, actor role, control, purpose, and the applicable date. The retained material also differs by article. Article 18 documentation is not a command to keep every prompt or output for ten years.
The Council adopted final legislative text that would delay key high-risk duties. It sets 2 December 2027 for systems classified under Article 6(2) and Annex III. It sets 2 August 2028 for systems classified under Article 6(1) and Annex I. PE-CONS 30/26, art. 1(40) (June 18, 2026). Article 4 makes entry into force depend on Official Journal publication. That publication was not verified as of July 15, 2026. Current status must be checked before relying on the delayed dates.
AI Act logs can later become litigation or investigation evidence. A six-month regulatory minimum is not a universal litigation ceiling. A triggered hold may require longer preservation, subject to applicable privacy, secrecy, and transfer rules.
A minimum defensible control set
A workable program assigns named owners and repeatable actions before a dispute begins.
Record map. Identify each AI system, business purpose, owner, model, vendor, data sources, outputs, logs, locations, default periods, export methods, and deletion controls.
Trigger route. Route demand letters, claims, incidents, regulator contacts, internal complaints, and credible disputes to legal review. Record the trigger decision and its factual basis.
Hold scope. Define issues, custodians, systems, vendors, dates, record categories, and technical actions. State exclusions and the reason for each material exclusion.
Technical preservation. Suspend deletion, preserve relevant accounts, export native data, capture metadata, record hashes, and keep chain-of-custody records. Preserve the operative model and configuration when they matter.
Validation. Obtain custodian acknowledgments and administrator confirmation. Test exports. Check that deletion stopped. Repeat checks when systems, claims, or personnel change.
Release. Document matter closure, related-duty review, hold release, and return to ordinary schedules. Delete records when no legal, regulatory, contractual, or business basis remains.
