title: "E-Discovery Sanctions: When AI Redaction Goes Too Far" description: "In Athletics Investment Group v. Schnitzer Steel (2024), improper redaction triggered discovery sanctions. With AI tools achieving only 22.7% precision, legal teams face real liability." category: legal-tech publishedAt: 2026-03-12 tags:
- e-discovery sanctions
- redaction liability
- AI redaction precision
- document review
- legal technology readingTime: 10
Updated for 2026
Two Ways Redaction Fails
Legal teams face two failure modes. Both create real liability.
Under-redaction exposes privileged data or personal info that must stay hidden. The party discloses material it had a right — and often a duty — to protect.
Over-redaction hides facts that opposing counsel has a right to see. Courts treat this as obstruction. It is a discovery violation subject to sanctions.
AI tools that favor recall over precision cause the second problem by design. An AI engine that blacks out 80% of a doc avoids missing anything. But the result is useless. It may also draw court sanctions.
Both failure modes lead to the same place: a judge, an explanation, and costs.
The Schnitzer Steel Case (2024)
The 2024 case of Athletics Investment Group v. Schnitzer Steel shows how courts handle improper document withholding.
One party produced docs with broad markings. Opposing counsel pushed back. The court looked at the materials. It found the markings went beyond what the law allowed.
The result: sanctions under Federal Rule of Civil Procedure 37. The producing party paid for a flawed process.
Such sanctions are not new. Courts have used them for years. What makes this case stand out is timing. AI-assisted review is now common in litigation. The case raises a key question: have legal teams checked the precision of their AI tools before using them in production?
The answer matters. A tool with poor precision will flag far too much. The attorney who relies on it without checking bears the risk.
For a full case breakdown, see E-Discovery LLC's analysis of relevance-based withholding.
The 22.7% Precision Problem
Presidio is an open-source PII detection engine built by Microsoft. It is widely used in document review tools. Tests on court filings and contracts give it a 22.7% precision rate.
Precision measures how often a positive flag is correct. At 22.7%, about 77 of every 100 flags are false positives. Those items are not sensitive by any applicable standard.
For e-discovery, the math is direct. A set of 10,000 docs processed at that rate will have thousands of baseless markings. The producing party faces the same risk as the Schnitzer Steel defendant: a challenged production, a court review, and possible sanctions.
This figure is for Presidio's out-of-box setup on law firm content. Not all AI tools perform at this level. But this engine is the most widely used open-source option in the field.
The cause is structural. NLP systems train on general text. Courtroom language is different. It uses terms of art, citation formats, and drafting rules that deviate from training data. A tool that works well on medical records may do far worse on deposition transcripts.
What AI Usage Data Shows
Here is a second data point: 27.4% of AI chatbot content is sensitive, per independent analysis of enterprise AI usage.
This describes what employees send during normal tasks. Not data they meant to share — content included by habit or accident. Attorneys using AI to draft letters, review contracts, or summarize depositions send sensitive content to AI servers as a side effect of normal work.
Nearly three in ten interactions involve client data, privileged info, or case strategy. That content reaches the AI vendor's servers in usable form unless controls stop it first.
For law firms checking their AI risk, 27.4% is not a minor issue. It is the base rate. Nearly a third of AI use in a firm involves content that needs protection.
The Liability Chain
Over-withholding and AI data leaks create separate but linked risk paths. Both start with the same decision: deploy an AI tool without proper evaluation.
The discovery path: AI flags content broadly → attorney relies on output without spot-checking → production has unjustified markings → opposing counsel objects → court reviews → sanctions.
The data leak path: Attorney uses AI for case work → AI receives privileged comms → AI vendor has a breach → client data is exposed → malpractice claims follow.
The starting point is the same in both cases. Firms deploy AI tools without knowing what those tools actually do. No controls are set up for the work.
Precision-First Review for Productions
Courts ask a narrow question when they review disputed markings. Was each one backed by privilege, a confidentiality rule, or a court order? Courts do not ask whether the producing party's tool flagged as much as possible.
A marking without a proper basis is a discovery violation. It does not matter if a human or an AI made it. The inquiry is marking-by-marking.
For attorneys, this means AI review tools need to be tested on precision — the share of flags that are truly privileged. Not just recall. A tool that hits 90% recall at 22.7% precision catches more sensitive content. But it creates a review burden for the 77.3% of false flags. When that review does not happen, broad over-withholding follows.
Each marking in a production is a claim to the court. It says: this content is legitimately withheld. After Schnitzer Steel, that claim must hold up.
For more on how anonymization tools differ from standard PII detection, see our guide to AI precision in legal document review. For context on privilege logs and AI tools, see our piece on attorney-client privilege and AI.