AI citation failures, agentic legal workflows, private-model infrastructure, and automated due diligence

Two courts respond to unverified AI output as legal teams assess long-horizon agents, private models, and automated due diligence.

Edition
eDiscovery Decoder Newsletter

Sent September 14, 2026

Lead story

New Mexico Supreme Court Sanctions Lawyer Over AI-Fabricated Witness Testimony

The New Mexico Supreme Court held a defense lawyer in direct contempt after an AI-assisted murder-appeal brief included testimony attributed to witnesses who did not exist.

What happened
The New Mexico Supreme Court held a defense lawyer in direct contempt after an AI-assisted murder-appeal brief included testimony attributed to witnesses who did not exist.
Why it matters
The order extends the legal-AI verification problem beyond nonexistent citations to invented record evidence. Litigation teams need controls that trace assertions to the actual record, confirm quotations and witnesses, document human review, and prevent generated text from bypassing filing-quality checks.
Watch
The news report was published September 13, while the underlying court order was filed September 9. The order concerns one lawyer's conduct and does not prohibit responsible use of AI in legal work.

Sourcespymnts.comcdn.arstechnica.net

Briefing item

Supio Describes Long-Horizon Agents for Multi-Step Legal Workflows

Supio is positioning persistent agents to coordinate legal work across case data, communications, external providers, and firm systems over days or weeks.

What happened
Supio is positioning persistent agents to coordinate legal work across case data, communications, external providers, and firm systems over days or weeks.
Why it matters
Persistent agents create governance requirements that are broader than those for a single prompt or document review. Discovery teams will need bounded permissions, source attribution, audit trails, role-based access, escalation rules, and controls over communications or actions taken outside the review platform.
Watch
The article is an analyst's account based partly on a Supio briefing and customer statements. Product performance, the reported customer outcome, security controls, and the completeness of agent audit records were not independently tested for this brief.

Sourcessiliconangle.comsupio.com

Briefing item

D.C. Court Strikes Entire Brief After Lawyer Cites Four Nonexistent AI-Generated Cases

The District of Columbia Court of Appeals struck Deutsche Bank's brief after counsel acknowledged that four cited authorities generated through AI did not exist.

What happened
The District of Columbia Court of Appeals struck Deutsche Bank's brief after counsel acknowledged that four cited authorities generated through AI did not exist.
Why it matters
Citation existence checks alone are insufficient for AI-assisted legal work. Review protocols should also compare quoted language, holdings, procedural posture, and the proposition asserted against the authoritative source, with supervision and sign-off recorded before filing.
Watch
The underlying foreclosure appeal remained undecided at the time of the report. Any disciplinary investigation or sanction beyond striking the brief was not yet resolved.

Sourceswjla.com

Briefing item

Latham's GPU Strategy Raises Questions About Private Legal-AI Infrastructure

A new analysis examines Latham & Watkins' investment in Nvidia GPU capacity for open-weight models and tests the cost, confidentiality, and data-sovereignty rationale.

What happened
A new analysis examines Latham & Watkins' investment in Nvidia GPU capacity for open-weight models and tests the cost, confidentiality, and data-sovereignty rationale.
Why it matters
Architecture choices change the discovery and governance surface. Firms considering private models need documented rules for client-data authorization, corpus selection, retention, tenant separation, model evaluation, incident response, and the preservation or reproducibility of outputs used in matters.
Watch
The September 14 source is analysis rather than a new firm announcement and builds on reporting published September 11. It does not provide Latham's technical architecture, costs, training corpus, deployment controls, or evaluation results.

Sourcesartificiallawyer.comLegal IT Insider

Briefing item

Xapien Raises $56 Million to Expand AI-Assisted Counterparty Due Diligence

Xapien announced a $56 million investment round to expand its automated people-and-company due-diligence platform, including a larger U.S. presence.

What happened
Xapien announced a $56 million investment round to expand its automated people-and-company due-diligence platform, including a larger U.S. presence.
Why it matters
Automated counterparty research can accelerate investigations and matter intake, but its findings may influence high-stakes decisions. Teams should retain source-level evidence, record search scope and timing, distinguish verified facts from risk signals, and provide a review path for identity resolution and adverse findings.
Watch
Funding terms beyond the announced amount were not disclosed. Client counts, revenue mix, platform capabilities, and due-diligence coverage are company-reported and were not independently verified for this brief.

SourcesLegal IT Insiderxapien.com

Educational only. Not legal advice. Review sources and matter-specific obligations before relying on any workflow decision.

eDiscovery Brief: AI Citations, Long-Horizon Agents, and Private Models | eDiscovery Decoder Newsletter