The Dual Standard: When AI Reliance Is Negligent and When Non-Use May Become Negligent
Professional Liability in the Age of Generative AI: From Sullivan & Cromwell to the SRA's Competence Consultation
Introduction
On 18 April 2026, Sullivan & Cromwell LLP wrote to Chief Judge Martin Glenn of the United States Bankruptcy Court for the Southern District of New York after Boies Schiller Flexner identified AI-generated errors in a Chapter 15 filing in the Prince Global Holdings proceedings. The reported errors included inaccurate citations, misstatements or misquotations of bankruptcy law and other drafting defects that should have been caught before filing.1 The incident matters not because elite firms are uniquely vulnerable, but because it shows that AI risk has moved from the margins of legal practice into the core workflow of sophisticated firms.
The professional liability question is becoming unavoidable: when does reliance on AI constitute negligence, and when might failure to use AI become negligent? Those questions appear to point in opposite directions. In fact, they are converging. The emerging standard is not anti-AI and it is not pro-AI. It is a standard of competent, supervised and verified use.
The Scale of the Problem
The Sullivan & Cromwell incident is not isolated. It is the most prominent recent example of a wider failure pattern. Damien Charlotin’s AI Hallucination Cases database tracks legal decisions in which generative AI has produced hallucinated content, or where AI use has been addressed by a court or tribunal in more than a passing reference. The database is important, but its methodology matters: it does not purport to capture every false citation in every filing. It captures the court-facing subset that has reached judicial or tribunal attention.2
As at 15 May 2026, Charlotin’s tracker identified 1,450 matters worldwide, including 1,003 in the United States.3 The precise number changes quickly. The direction of travel does not. What began as a few notorious filing failures has become a recurring operational risk in litigation practice.
The origin point in public consciousness remains Mata v Avianca. In June 2023, Judge P. Kevin Castel of the Southern District of New York imposed a $5,000 Rule 11 sanction after lawyers submitted an affirmation containing six non-existent judicial opinions generated by ChatGPT.4 At the time, much of the profession treated Mata as a cautionary novelty. By 2026, that reading is no longer sustainable.
The sanctions environment has hardened. In Couvrette v Wisnovsky, the District of Oregon found that summary judgment briefing contained 15 non-existent cases and eight fabricated quotations. The financial outcome was not a single fine, but a combination of a $15,500 sanctions order and later fee and cost orders that together exceeded $110,000.5 The distinction matters: the case should not be described simply as a ‘$110,000 fine’. Its real significance is broader. Courts are increasingly willing to make the cost of verification failure fall on the lawyers and parties responsible for putting false material before the court.
Court-level requirements are also multiplying. As at May 2026, Legal AI Governance identified 113 active orders and rules binding attorney filings.6 The requirements vary. Some require disclosure of AI use. Some require certification that AI-assisted work has been checked. Others simply restate that Rule 11, candour and professional responsibility duties apply irrespective of the tool used. The common message is that ‘AI did it’ is not a defence.
The Regulatory Response
The American Bar Association responded with Formal Opinion 512 on 29 July 2024, its first comprehensive ethics opinion on generative AI in legal practice. The opinion addresses competence, confidentiality, communication with clients, billing, candour to the tribunal and supervisory responsibilities under Model Rules 1.1, 1.4, 1.5, 1.6, 3.3, 5.1 and 5.3.7 The central proposition is simple: a lawyer may use generative AI, but the lawyer remains responsible for the work product.
Formal Opinion 512 does not require lawyers to understand the model architecture of every system they use. It does require them to understand the capabilities and limitations of the tools sufficiently to use them competently. It also requires lawyers to protect confidential information, communicate with clients where AI use materially affects the representation, verify AI-generated legal analysis and citations before submission, supervise junior lawyers and non-lawyers using AI and avoid billing clients for time spent correcting avoidable AI errors.
California has moved towards a more prescriptive model. In 2026, the State Bar of California sought public comment on proposed amendments to six Rules of Professional Conduct addressing AI: competence, communication, confidentiality, candour to the tribunal, managerial responsibility and supervision of non-lawyer assistants.8 If adopted, the proposals would make verification explicit. A proposed comment to the competence rule states that, when using technology including AI, a lawyer must independently review, verify and exercise professional judgment over the output. A proposed comment to the candour rule specifically requires verification of the accuracy and existence of cited authorities before submission to a tribunal.
That is not yet the same as an enacted rule. The California material should therefore be described as a proposed amendment, not as binding black-letter law. Its significance lies in the direction of regulatory movement: duties that were previously implicit in competence and candour rules are being translated into AI-specific drafting.
Colorado is useful for a different reason. It has both a disciplinary example and a rule-based development. In People v Crabill, a Colorado lawyer received a suspension of one year and one day, with 90 days actively served, after using ChatGPT-generated case law without reading or verifying it, failing to alert the court to sham cases and then falsely attributing the errors to a legal intern.9 Separately, the Colorado Supreme Court approved Rule Change 2026(02), which adds commentary making clear that technology, including AI, does not diminish a lawyer’s professional responsibilities and that a lawyer who uses technology in delivering legal services may be subject to discipline for a resulting rule violation.10
These regulatory materials share a common architecture. They treat AI as a tool, not as an independent professional actor. They impose verification duties. They reject any presumption that AI-generated legal output is reliable. They locate the obligation within existing competence, candour, confidentiality and supervision frameworks rather than creating a wholly new category of professional regulation.
The English Position
The leading English authority is Ayinde v London Borough of Haringey and Al-Haroun v Qatar National Bank, heard together and decided by the Divisional Court in June 2025.11 The judgment is often described as an AI case, but that shorthand needs care. The court was dealing with the actual or suspected use of generative AI by lawyers to produce written legal arguments or witness statements that were not checked, resulting in false information being put before the court.
In the Ayinde matter, the underlying problem was the inclusion of five fake cases in judicial review materials. The provenance of the false authorities was contested; the barrister denied using AI. Ritchie J nevertheless made wasted costs orders of £2,000 each against Ms Forey and Haringey Law Centre and required referral to the Bar Standards Board and the Solicitors Regulation Authority.12 The Divisional Court later held that the threshold for initiating contempt proceedings was met in relation to Ms Forey, but decided not to initiate contempt proceedings or refer the case to the Law Officers.13
The practical guidance from the judgment is more important than the disputed factual mechanics. The court stated that freely available generative AI tools such as ChatGPT are not capable of conducting reliable legal research. They may cite sources that do not exist, quote passages that do not appear in genuine sources and produce confident assertions that are simply untrue. Lawyers who use AI for legal research, or rely on others who have done so, have a professional duty to check the output against authoritative sources before using it in advice or before a court.
The Solicitors Regulation Authority has not adopted an ABA-style AI ethics opinion. Its approach remains principles-based. On 22 April 2026, however, the SRA opened its consultation, ‘Strengthening our continuing competence approach’, running until 15 July 2026.14 The consultation proposes stronger requirements for solicitors to record how they identify and address learning and development needs and to participate in annual ethics discussions. It is not an AI-specific rulebook, but it sits alongside the SRA’s broader AI risk materials and compliance guidance on AI and technology.15
The Law Society’s guidance on generative AI likewise has persuasive rather than binding force. Its September 2025 update expressly added references to Ayinde and Al-Haroun, and it now stresses verification against reliable and authoritative sources, supervision and risk management.16 The English position is therefore less prescriptive than the emerging American model, but not less serious. The courts have made clear that false authorities in court documents may trigger wasted costs orders, regulatory referral and, in an appropriate case, contempt proceedings.
The Emerging Dual Standard
Professional liability in the age of generative AI is developing along two vectors at the same time. The first is negligent reliance. A lawyer who submits AI-generated legal material without proper verification may breach duties of competence, candour, supervision and care to the client. That proposition is now strongly supported by the cases, ethics opinions and regulatory materials.
The second vector is less developed but potentially more disruptive: the possibility that failure to use AI may, in some circumstances, fall below the standard of care. This proposition should not be overstated. No court has yet held a lawyer negligent simply for failing to use AI. Nor is there a general duty to use every available technology. The better formulation is narrower: where a particular AI tool has become a reasonably standard, reliable and proportionate means of improving the relevant task, a professional who fails to consider or deploy it may struggle to justify the omission if the client suffers avoidable loss.
That argument is consistent with orthodox professional negligence principles. The standard is not perfection. It is the standard of a reasonably competent professional in the relevant circumstances. In English law, Bolam and Bolitho provide useful analogies, particularly the idea that a practice accepted by a responsible body of professionals must also withstand logical scrutiny. In the solicitor context, the same essential question appears in a different form: what would a reasonably competent practitioner have done, judged by the standards of the profession at the time?
Anurag Bana’s SSRN paper on artificial intelligence, legal professional negligence and AI-covered indemnity risk articulates the point clearly: as AI becomes prevalent in legal practice, liability may arise both from using AI incorrectly and from failing to use it where its use would have been reasonably expected.17 Recent professional negligence commentary takes a similar position: AI is not generally mandatory, but the real question is whether a reasonably competent professional would have used it in the client’s interests in the particular circumstances.18
The point is easiest to see in verification. A lawyer who asks a public chatbot to produce case law and files the result without checking it is plainly exposed. But consider the inverse case: a firm has access to a reliable citation-checking, document-comparison or disclosure-analysis tool; the tool is widely used for the task; the cost of using it is proportionate; and the error that later causes loss is precisely the kind of error the tool would probably have caught. In that scenario, failure to use AI is not negligence because AI exists. It is negligence, if at all, because the professional failed to use an available and reasonably standard quality-control method.
The practical tension is acute. A lawyer who uses AI without verification may be negligent. A lawyer who refuses to use AI where it has become an ordinary part of competent practice may also become exposed. The safe ground is not abstention. It is disciplined adoption: use AI where it adds value, understand its limits, verify its output, supervise its use and preserve human professional judgment as the final decision-making layer.
Indemnity and Insurance Implications
Professional indemnity insurers now face the dual standard as both a coverage and pricing problem. The negligent reliance vector produces familiar claims in new clothing: failed applications, adverse costs orders, lost procedural opportunities, confidentiality breaches and client losses caused by unverified work product. The failure-to-use vector is more novel: claims alleging that a lawyer failed to use an available technology that would have prevented the loss.
Coverage questions remain unresolved. United States commentary on lawyers’ professional liability policies notes that many policies do not expressly exclude AI use, but coverage may depend on whether the conduct falls within the policy definition of professional services and whether exclusions for intentional acts, fraud, fee disputes or technology failures are engaged.19 Some professional liability insurers have also begun experimenting with AI-specific exclusions or endorsements.20
The hardest cases will sit between negligence and recklessness. An isolated failure to check an AI-assisted draft may be characterised as negligence. Repeated submission of fabricated authorities after warning signs have been raised, or a deliberate refusal to verify citations known to have come from a generative tool, may be characterised very differently. Coverage will turn on policy wording, governing law and the factual findings in the underlying claim.
The NAIC Model Bulletin on the use of AI systems by insurers, adopted in December 2023, is relevant but only indirectly.21 It addresses insurer use of AI in insurance operations, including governance, risk management and compliance with insurance law. It does not solve the professional indemnity question of how insurers should underwrite or respond to claims arising from insured lawyers’ use, misuse or non-use of AI.
For underwriters, the risk is now two-sided. Traditionally, technology risk in professional indemnity was framed as a risk of using defective systems. AI introduces the additional possibility of technology abstention risk: the allegation that competent practice required the use of a tool and the insured failed to use it. That does not mean underwriters should require blanket AI adoption. It means proposal forms, renewal questions and risk engineering will need to move beyond asking whether a firm uses AI and start asking how AI is governed, supervised, verified and documented.
Strategic Outlook
The standard of care is moving. Expected work product quality will rise as reliable AI-assisted methods become normalised. Firms that ban AI entirely face one liability vector. Firms that permit uncontrolled AI use face another. Both approaches are inferior to governed use.
The first strategic imperative is verification. Every firm should have a clear rule that AI-generated legal authorities, quotations, factual propositions and analytical conclusions must be checked against authoritative sources before submission to a court, delivery to a client or use in advice. The rule should apply not only to partners and associates, but also to trainees, paralegals, knowledge teams, external consultants and anyone else contributing to legal work product.
The second imperative is supervision. AI use should be treated like any other delegated work stream. A partner is not excused because the first draft came from an AI tool rather than a junior lawyer. A recent US sanction against a managing partner for a junior lawyer’s AI-related citation error illustrates the same point in operational terms.22 The supervision question is the same: who checked it, against what source, using what process and where is the evidence that the check occurred?
The third imperative is training. The profession has spent too much time debating whether AI should be used and too little time teaching practitioners how to use it responsibly. Competence training needs to cover prompt design, tool selection, confidentiality, privilege, hallucination risk, citation verification, document comparison, disclosure workflows, billing treatment and escalation protocols. The SRA’s continuing competence consultation points in this direction even if it does not prescribe AI-specific rules.
The fourth imperative is insurance engagement. Firms should not wait for a claim before asking how their professional indemnity policy treats AI-assisted legal work. They should understand whether AI use is within the scope of covered professional services, whether any AI-specific exclusions apply, whether sanctions and fee-shifting orders are covered and what notification obligations arise when an AI-related error is discovered.
The dual standard is likely to sharpen over the next three to five years. Courts will continue to sanction negligent reliance. At some point, a failure-to-use allegation is likely to be tested in a professional negligence claim, most probably where a client can show that a readily available verification, disclosure or drafting tool would have caught the error that caused the loss. The claim may or may not succeed. Its arrival should surprise no one.
The profession cannot resolve this tension by choosing one vector over the other. It can resolve it only by occupying the disciplined middle ground: deploying AI tools competently, verifying their output rigorously, supervising their use systematically and maintaining human judgment as the irreducible core of legal practice. Firms that master that discipline will reduce cost, improve quality and protect themselves. Firms that do not will be exposed in both directions: for unverified AI reliance and, in time, potentially for failing to adopt standard AI-assisted checks.
Notes
1. Reuters, ’Sullivan & Cromwell law firm apologizes for AI hallucinations in court filing’, 21 April 2026; Legal Cheek, ’Sullivan & Cromwell apologises after AI hallucinations appear in court document’, 22 April 2026.
2. Damien Charlotin, AI Hallucination Cases Database, methodology note: database tracks legal decisions where the use of AI, whether established or merely alleged, is addressed in more than a passing reference by a court or tribunal; it does not track the wider universe of all false citations or uses of AI in court filings.
3. Damien Charlotin, AI Hallucination Cases Database, last updated 15 May 2026, identifying 1,450 cases worldwide and 1,003 USA matters. The database is live and should be rechecked immediately before publication.
4. Mata v Avianca, Inc., No. 22-cv-1461, 678 F Supp 3d 443 (SDNY 2023), sanctions order dated 22 June 2023.
5. Couvrette v Wisnovsky, No. 1:21-cv-00157-CL (D Or), Opinion and Order on sanctions, 12 December 2025, 2025 WL 4109655; Opinion and Order on fee/cost allocation, 23 March 2026, ECF No. 225; subsequent merits order, 30 March 2026, ECF No. 227.
6. Legal AI Governance, Federal and State Court Orders on AI tracker, identifying 113 active orders and rules binding attorney filings.
7. American Bar Association, Formal Opinion 512, Generative Artificial Intelligence Tools, 29 July 2024.
8. State Bar of California, Proposed Amendments to the Rules of Professional Conduct Related to Artificial Intelligence, public comment material, 2026; public comment deadline 4 May 2026.
9. People v Zachariah C. Crabill, 23PDJ067, Colorado Office of the Presiding Disciplinary Judge, 22 November 2023.
10. Colorado Supreme Court, Rule Change 2026(02), approved 8 January 2026.
11. Ayinde v London Borough of Haringey and Al-Haroun v Qatar National Bank QPSC [2025] EWHC 1383 (Admin), Divisional Court, 6 June 2025.
12. R (Ayinde) v London Borough of Haringey [2025] EWHC 1040 (Admin), Ritchie J; summarised in the Divisional Court judgment at [2025] EWHC 1383 (Admin).
13. Ayinde and Al-Haroun [2025] EWHC 1383 (Admin), discussion of contempt threshold and decision not to initiate contempt proceedings.
14. Solicitors Regulation Authority, Strengthening our continuing competence approach, consultation opened 22 April 2026 and closing 15 July 2026.
15. Solicitors Regulation Authority, Risk Outlook report on the use of artificial intelligence in the legal services sector, updated 23 April 2026; SRA compliance tips on AI and technology, updated 9 February 2026.
16. The Law Society of England and Wales, Generative AI: the essentials, 1 October 2025, Updates: September 2025.
17. Anurag Bana, ’Artificial Intelligence, Legal Professional Negligence and the Rise of AI-Covered Indemnity Risk’ (SSRN, abstract dated 2025; PDF posted 9 April 2026).
18. Cripps, ’When is it negligent for a professional to use or ignore AI?’, 7 May 2026.
19. Reuters Legal News, ’From innovation to exposure: artificial intelligence risks for legal professionals’, 14 July 2025.
20. Reuters Legal News / Westlaw Today, ’Insuring against productive laziness: attorney use of artificial intelligence’, 22 December 2025.
21. National Association of Insurance Commissioners, Model Bulletin: Use of Artificial Intelligence Systems by Insurers, adopted 4 December 2023.
22. Reuters Legal News, ‘US judge says senior lawyers must pay for mistakes by subordinates using AI tools’, 1 May 2026.