AI sanctions land on whoever signs, so your review tool must make checking cheap
Doug Austin's eDiscovery Today has summarised one trend from Exterro's whitepaper, Prompts, Privilege, and Preservation in 2026: the duty to verify AI output cannot be delegated. His examples are Pauliah v. Univ. of Mississippi Med. Center, where an attorney was sanctioned over AI-generated deposition summaries containing fabricated verbatim quotes. In Whiting v. City of Athens, the Sixth Circuit imposed $15,000 punitive sanctions per attorney for hallucinated citations. In American Council of Learned Societies v. NEH, a federal judge refused to let an organisation blame ChatGPT for a flawed grant-screening process.
Who is liable when AI output is wrong?
The human or organisation that put the tool to work. In the rulings cited, courts placed responsibility on the signing attorney or the organisation that chose the AI platform, not the software.
In Pauliah and Whiting, the flawed material went to court and the signers paid. In the NEH case, Austin reports, the court treated the platform as the organisation's "chosen instrument", so the organisation owns its outputs. None of these is a review-platform dispute, and the summary is silent on vendor contracts. Applying the logic to review software is our inference, though expect it to be argued. Do not assume your vendor's terms move the exposure off you.
What do the $15,000 figure and this roundup prove?
Little on their own. The sanction is a single data point, and the rulings were selected by Exterro, a sponsor of the site that summarised them.
The figure comes from one appellate decision, applies per attorney, and punishes fabricated citations in briefs. That is a different failure from a tool mis-coding documents, and amounts turn on the court and the conduct. Austin discloses that Exterro sponsors his site, and Exterro chose which rulings to feature and how to group them. We have seen only Austin's summary, not the opinions, so read the whitepaper itself before citing it in policy.
Which review workflows share the deposition-summary risk?
Any feature that summarises, quotes or paraphrases evidence, such as deposition summaries, chronologies and document-level explanations, because a fabricated quote in those looks identical to a real one.
A deposition summary is the same kind of output as a chronology or an answer about the record: generated prose a busy reviewer is tempted to paste into work product. Austin's advice to check line by line only works if the tool lets a reviewer jump from each statement to its exhibit. Otherwise verification means rereading the transcript. The test for any RAG product is not how rarely it errs but how fast an error is caught.
Our report on the Broward judge duped by a deepfake made a related point: whoever relies on material answers for it.
What should I ask my vendor before renewal?
Ask whether every generated statement links to its source passage, whether quotes are checked against the text automatically, and whether reviewers' verification is logged. Get answers in writing.
Press on how any accuracy claim was produced: who defined a hallucination, which documents were tested, who measured. An average error rate means little to a judge reading your filing. Ask whether an unsupported quote is blocked or merely flagged, and whether the system declines when it finds no source. The whitepaper's title puts prompts and preservation in scope, though Austin does not say what courts held. Ask whether prompts and outputs are logged and exportable, and whether privileged material ever trains a model.
Frequently asked questions
Does the NEH ruling apply to review platforms?
Not directly; it concerned grant screening. The reasoning, that you own the output of the AI you choose, may be argued by analogy. That is our reading of a second-hand summary.
Is the full whitepaper in Austin's post?
No. He summarises a single trend and points readers to Exterro for the rest. We have not reviewed it.
Sources: eDiscovery Today (Doug Austin), summarising Exterro's whitepaper.
The original report
Industry headlines from other publications. Each links to the original reporting on the publisher's own site.