Large language models can draft chronologies in minutes. They can also invent citations, dates, case names, and privilege boundaries. For funders underwriting non-recourse capital, that gap is a budget and waiver problem — not a novelty to hand-wave away. This article maps where hallucinations hurt commercial litigation, controls funders require, and how dossier silo architecture reduces exposure.

Keywords: legal AI hallucination, litigation chronology AI, funder controls, waiver risk

Where hallucinations hurt litigation

• Chronologies — wrong timestamps break settlement corridors and credibility with opponents
• Disclosure lists — phantom documents trigger costly rework and waiver debates
• Issue lists — counsel spends days correcting machine output instead of arguing merits
• Case citations — invented judgments surface in drafts until red-team catches them
• Privilege calls — models guess boundaries; wrong guesses create disclosure risk

Cost of uncorrected hallucination

Controls we require before funding

1. Human sign-off — qualified lawyer approves any AI output used in pleadings or disclosure
2. Source pinning — every asserted fact links to Bates or native file hash
3. Versioning — prompts, model IDs, and outputs stored in the data room
4. Red-team pass — second reviewer hunts for invented cases and quotes
5. Silo export gate — no LLM reads full corpus until counsel approves — see silo screening

What claimants should expect

AI can shrink first-pass review cost on large corpora. It cannot replace counsel judgment on privilege, strategy, or settlement authority. Merits scores from screening are inputs to diligence, not substitutes for lawyer advice — automation boundaries.

FAQ

Can we ban AI entirely? Unrealistic — govern it with protocol.

Does screening hallucinate merits? Production uses rubric + counsel gate; demo auto-merits is placeholder only.

ChatGPT for chronologies? Never on privileged corpus — use silo path.

UK disclosure? Extra care on lists — UK checklist.

Protocol checklist for counsel

1. Define allowed AI use cases (chronology draft yes; privilege calls no)
2. Require Bates/hash on every exported paragraph
3. Store model version and prompt with output in data room
4. Red-team sample 10% before disclosure reliance
5. Block consumer chat upload — silo only

Funder underwriting impact

Matters with ad-hoc AI use and no protocol get higher waste haircut on review curve assumptions. Matters with documented controls proceed at standard bands.

Regulatory and professional conduct angle

Law societies and bar guidelines increasingly address generative AI in practice. Funders ask whether firm AI policy exists before Tranche 2 on matters using heavy doc review. Absence of policy ≠ automatic decline — but requires expedited protocol build.

Insurance and indemnity

Professional indemnity insurers ask about AI use on matters. Claimants and counsel should notify brokers when AI-assisted workflows touch pleaded documents — funder diligence may request confirmation.

Extended FAQ

Zero AI policy? Start with silo + sign-off minimum — silo.

Opponent uses AI? Does not lower your disclosure duty — UK checklist.

Training data vs your matter

Public models may have seen public judgments — not your privileged bundle. Never infer that AI "already knows" your case. Screening on silo keeps payer corpora out of consumer training paths.

Glossary

• Hallucination — confident false model output
• Source pinning — Bates/hash link
• Red-team — adversarial review pass

Related reading on this site

Continue with linked pillars in this article and the blog index. Machine-readable catalogue: llms.txt. Cost stress-test: burn calculator. Live screening: agents section. Questions: [email protected]. All content educational — engage qualified counsel for your matter.

See screening demo, review cost curve, screening limits.