The Legal AI 'Ghostwriter Liability' Problem: Why Partners Are Signing Off on Work Product They Cannot Fully Reconstruct
There is a quiet crisis building inside the upper floors of large law firms, and most of the partners contributing to it do not fully realize it yet. They are signing off on memos, briefs, due diligence summaries, and contract analyses that were substantially drafted...
There is a quiet crisis building inside the upper floors of large law firms, and most of the partners contributing to it do not fully realize it yet. They are signing off on memos, briefs, due diligence summaries, and contract analyses that were substantially drafted by AI systems — and when pressed, they could not walk a client, a judge, or a disciplinary board through the reasoning chain that produced the work. That is not a technology problem. It is a professional responsibility catastrophe waiting to be triggered.
What 'Competent Supervision' Actually Requires
Model Rule 5.1 and its state equivalents require supervising attorneys to make reasonable efforts to ensure that work performed under their oversight conforms to professional obligations. For decades, that standard was applied to associate attorneys and paralegals — humans whose reasoning you could probe in a conference room. The rule was never designed with a large language model in mind, but it applies anyway.
Competent supervision of an AI system requires more than reading the output and deciding it looks right. It requires understanding the methodology well enough to defend it. If a partner cannot explain why the AI flagged a particular clause as high-risk, cannot reconstruct the comparative analysis that produced a risk rating, and cannot identify whether the model was trained on data current enough to reflect the applicable legal standard, then the partner has not supervised the work. They have endorsed it. Those are materially different things under Rule 5.1.
The ABA's Formal Opinion 512, issued in 2024, addressed generative AI use in legal practice and was clear that responsibility for AI-assisted work product rests entirely with the attorney of record. What it did not do — and what the profession has been slow to confront — is specify what supervision actually looks like when the drafter operates probabilistically rather than deterministically.
Where Reconstruction Failure Becomes a Liability Event
The risk is not theoretical, and it clusters around four specific scenarios.
Adversarial depositions of the work product itself. In complex commercial litigation, opposing counsel increasingly asks producing parties to identify the methodology behind legal analysis. If a memo was AI-assisted, and the partner cannot explain the analytical steps, you have created a credibility problem that can metastasize into a sanctions motion. Mata v. Avianca in 2023 was an early and spectacular example of what happens when an attorney cannot defend AI-generated citations before a federal judge. The reputational and monetary sanctions in that case should have been a sector-wide reckoning. They were not taken seriously enough.
Transaction closings where the AI was wrong. A due diligence summary generated by AI that missed a material contractual obligation — because the model hallucinated an irrelevant precedent or misread a cross-reference — becomes a malpractice trigger the moment the client suffers loss. The partner's inability to show they verified the specific analytical chain compounds the exposure. You cannot prove reasonable care through general statements that you "reviewed the output."
Regulatory submissions to agencies with technical staff. The SEC's enforcement division, for instance, has sophisticated legal staff who will probe the logic of legal arguments in comment letters and disclosure documents. An AI-drafted argument that relies on a mischaracterization of a rule will be caught. The inability to reconstruct your own reasoning will be noticed.
Bar complaints from disgruntled former clients. This is the least discussed vector but potentially the most dangerous long-term. A client who believes they received inadequate representation has every incentive to scrutinize how the work was produced. "We used AI tools" followed by an inability to demonstrate supervisory process is not a defense. It is a confession.
How California, New York, and Texas Are Signaling Their Positions
The three most consequential bar jurisdictions in the country are moving, at different speeds, toward enforceable positions.
California's State Bar published its Practical Guidance for the Use of Generative AI in late 2023 and has been building toward a formal opinion that will treat AI supervision failures as competence violations under Rule 1.1. The California Bar's Technology Task Force has been explicit in public remarks that signing off on AI work product you cannot independently verify is not competence — it is abdication.
New York is moving more cautiously but no less seriously. The NYSBA's Task Force on Artificial Intelligence released its final report in 2024 recommending that attorneys maintain documentation of AI use sufficient to demonstrate meaningful review. New York's disciplinary framework, built around the concept of reasonable professional judgment, will not excuse the "I reviewed it" defense without evidence of what that review entailed.
Texas adopted Formal Ethics Opinion 705 addressing supervisory duties generally, and the State Bar's guidance updates have specifically included AI tools within the ambit of attorney supervision obligations. Texas is also an opt-out jurisdiction on ethics opinions, which means formal guidance lands with practical weight.
Documentation Habits That Actually Reduce Exposure
The fix is not abandoning AI tools. The fix is building a supervisory record.
Partners using AI for substantive work product should, at minimum, maintain a contemporaneous log that identifies the AI system and version used, the specific prompts or queries submitted, the key assertions in the output that were independently verified, the source materials consulted to confirm accuracy, and the judgment calls made by the attorney in modifying or rejecting AI-generated analysis. This does not have to be elaborate. It has to exist.
Firms should also require that any AI-assisted work product include an internal cover notation — not in the client-facing document, but in the matter file — identifying the AI's role and the verification steps taken. Some firms are building this into their matter management systems. Most are not.
The Accountability Gap Cannot Be Delegated Away
The legal profession has spent two years debating AI ethics at a philosophical level while individual partners have been quietly accruing liability exposure they cannot see. The ghostwriter problem is not about disclosure to clients, though that matters. It is about whether the attorney whose name is on the work can actually stand behind it under pressure.
If you cannot reconstruct the reasoning, you did not supervise it. And if you did not supervise it, you should not have signed it. The bar associations are arriving at that conclusion. The plaintiffs' malpractice bar will get there faster.