Too Reckless or Too Cautious? Legal AI and the New Standard of Care
For the last few years, most legal ethics commentary about artificial intelligence has focused on one obvious danger: lawyers using AI recklessly.
That concern is real. Lawyers have filed briefs citing cases that do not exist. Confidential information has been placed into tools without a clear understanding of where the data goes, whether it is retained, or whether it can be used to train future models. Courts, clients, and bar authorities have all warned lawyers that generative AI does not change the duties of competence, confidentiality, candor, supervision, communication, and reasonable billing.
The message has been simple enough: if lawyers use AI, they remain responsible for the work product.
But that is only half of the emerging standard of care.
The harder and more interesting question is whether lawyers may also face professional risk from being too cautious. If a secure and well-supervised AI tool could reduce cost, find a key document, flag a privilege issue, identify an inconsistency, summarize a record, or improve litigation preparation, can a lawyer simply refuse to consider it? At what point does avoiding AI stop looking prudent and start looking professionally unreasonable?
The answer is not that lawyers must use AI in every matter. They should not.
Some tools are not appropriate for sensitive client information. Some use cases are too unreliable. Some workflows create unnecessary privilege, confidentiality, or accuracy risks. But the profession is moving past the simple question of whether lawyers may use AI. The new question is whether a reasonable lawyer can explain when AI should be used, when it should not be used, and what safeguards are required when it touches client work.
That is the new standard-of-care problem. Legal AI now presents two kinds of ethical risk: the existing risk of misuse and the evolving risk of non-use.
The first risk: using AI badly
The first side of the risk is familiar because courts have already seen it.
Generative AI can produce authoritative-sounding legal analysis that is plain wrong. It can invent cases, misstate holdings, fabricate quotations, confuse jurisdictions, and cite sources that do not support the proposition offered. That is not merely a technology problem. In litigation, it becomes a Rule 11 problem, a candor problem, a competence problem, and potentially a sanctions problem.
Recent court-rule developments reflect that shift. Florida amended its statewide court rules in 2026 to require lawyers and self-represented litigants to certify that legal authorities cited in filings exist and are accurately cited, and to authorize sanctions for fabricated or inaccurately cited authorities. New York similarly adopted a statewide rule governing AI use in court submissions, emphasizing that attorneys and parties who use AI tools must understand their capabilities and limitations and must independently ensure that papers contain no fabricated or fictitious cases, statutes, or other material.
The important point is not that Florida or New York prohibit AI use. They do not.
The important point is that bar regulators (and courts) are becoming accountability focused. The lawyer who signs the filing owns the filing. The fact that a hallucination originated in an AI tool does not make it less sanctionable. In some ways, it may make the failure easier to explain and harder to excuse: the risk is now widely known.
That lesson was reinforced in a recent Mississippi federal case in which a judge disqualified attorneys on both sides of a dispute after unverified AI-generated legal research led to fabricated citations. The sanctions ($8,000 in all) reached not only the lawyers directly involved in the defective filings, but also local counsel who had served as filing counsel, disqualifying every lawyer who had appeared in the case. The ruling illustrates an important point for firms, clients, and co-counsel arrangements: the duty of AI supervision is not limited to the person who typed the prompt. If a lawyer signs, submits, sponsors, or adopts AI-assisted work, that lawyer must verify it.
This is the easiest AI ethics lesson to state: lawyers may use tools, but they may not outsource their judgment.
The second risk: using the wrong AI environment
Accuracy is only one part of the misuse problem. The other is data control.
A lawyer who uploads client documents, deposition transcripts, medical records, discovery materials, or privileged communications into a public or consumer-grade AI tool is creating confidentiality and privilege problems that are separate from whether the output is accurate. The issue is not whether the lawyer had good intentions. The issue is whether the lawyer understood the tool.
Where is the data stored? Is it retained? Can vendor personnel access it? Is it used for model training? Can it be disclosed in response to a subpoena, warrant, or court order? Is the client’s data segregated from other users’ data? Does the tool have zero data retention? Are there contractual confidentiality obligations? Does the vendor offer audit rights, encryption, access controls, or deletion commitments? Has the firm approved the tool for client data?
Those questions are not technical trivia. They go directly to a lawyer’s duty to protect client information.
Recent privilege disputes involving AI tools highlight the risk. Courts and commentators are now confronting whether prompts, outputs, and AI-generated analyses may be discoverable, privileged, or treated as work product. In some cases, litigants have argued that AI exchanges were part of legal strategy or preparation. But where a user voluntarily shares information with a third-party AI platform, especially one whose terms do not promise confidentiality, the privilege argument becomes much harder.
The practical lesson is this: lawyers should not treat all AI tools as interchangeable. A consumer chatbot, an enterprise AI environment, a litigation-specific review tool, and an internally hosted model present very different risk profiles for privilege and client confidentiality. The ethical question is not simply “Did you use AI?” It is “What AI did you use, for what task, with what data, under what contractual and technical safeguards, and with what human review?”
For sensitive client work, the defensible middle points increasingly toward tools with strong privacy and security commitments, including zero data retention and no-training configurations, encryption, access controls, documented deletion rights, vendor diligence, and clear restrictions on disclosure. For litigation teams, those safeguards should also be reflected in protective orders, discovery protocols, vendor agreements, and internal AI policies.
The third risk: charging as if AI does not exist
AI also changes the billing conversation.
The ABA’s Formal Opinion 512 made clear that lawyers using generative AI must consider not only competence, confidentiality, communication, supervision, meritorious claims, and candor, but also fee reasonableness. If a lawyer bills by the hour uses AI to complete a task more quickly, the lawyer cannot simply bill for the time the task would have taken without AI. Hourly billing must reflect actual time spent. Flat fees and other arrangements may remain permissible, but the fee must still be reasonable considering the nature of the work, the value delivered, and the efficiencies created.
That point is gaining force as state guidance grows. The Supreme Court of Rhode Island’s 2026 interim guidance, for example, emphasizes that generative AI does not change lawyers’ ethical duties but does require attention to competence, confidentiality, supervision, and billing. Rhode Island also amended its competence commentary to state that lawyers should keep abreast of the benefits and risks associated with existing and developing technology.
This is where the standard-of-care debate becomes more than a sanctions issue. AI is not only a risk to be controlled. It is also a potential efficiency that corporate in-house teams are beginning to expect outside counsel to evaluate.
If a lawyer uses AI to create a first-pass chronology, summarize deposition testimony, organize medical records, or identify inconsistent witness statements, the client should not be charged as though all that work was done manually from scratch. And if a lawyer refuses to consider tools that could safely reduce cost, the client may eventually ask why.
The overlooked risk: not using AI at all
The legal profession is beginning to talk openly about the other side of the problem: whether failing to use AI can itself become malpractice risk.
That theory is not yet settled law. As of this writing, there does not appear to be a clear, reported malpractice standard requiring lawyers to use generative AI in ordinary practice. Experts have noted that the standard of care remains undefined, and causation will be difficult in many cases. A client would have to prove not just that AI that could do the job was available, but that a reasonable lawyer would have used it, that the tool would likely have found something important or reduced cost, and that the failure caused harm.
But the absence of a malpractice standard today does not mean the issue is hypothetical. Standards of care often evolve before every lawyer adopts a new technology. The question is not whether every lawyer uses the tool. The question is whether a reasonable lawyer under the circumstances should have known the tool was available, suitable, and useful.
That is why the “failure to use AI” issue will likely arise first in concrete, document-heavy, cost-sensitive settings. Consider a few examples.
In a large document review, AI-assisted workflows may help surface key documents, cluster themes, detect duplicates, flag privilege indicators, and identify anomalies. If opposing counsel finds the critical document that a reasonable AI-assisted review would likely have surfaced, the client may ask why the tool was not used.
In deposition preparation, AI may help organize prior testimony, flag admissions, build exhibit lists, and identify contradictions across transcripts and documents. If counsel misses an impeachment point buried in the record, the client may ask whether the review process was reasonable.
In litigation budgeting, AI may reduce the cost of tasks that historically consumed dozens or hundreds of hours. If the client pays for a traditional manual process when a secure AI-assisted process would have been faster and cheaper, the issue may be framed not as technological curiosity but as fee reasonableness.
None of these examples suggest that AI should make the final decision. It should not. But they show why non-use can become difficult to defend in the right case. The reasonable lawyer of the near future may not be required to use AI for everything. But that lawyer may be expected to know enough to decide when AI belongs in the workflow.
The standard is not “AI-first.” It is risk-calibrated.
The emerging standard of care should not be anti-AI or pro-AI. It should be risk-calibrated.
Some uses of AI are low risk and high value. A lawyer may safely use an approved tool to reformat non-confidential text, create a draft agenda, summarize public information, generate internal brainstorming questions, or improve the readability of a document that contains no sensitive client information.
Other uses require stronger safeguards. Analyzing confidential discovery materials, extracting facts from medical records, drafting pleadings, evaluating deposition testimony, or preparing settlement materials is appropriate only in secure environments with clear data controls and attorney supervision.
Still other uses may be inappropriate altogether. A lawyer should not upload privileged communications into any unapproved tool. A lawyer should not cite AI-generated cases without independently verifying them. A lawyer should not rely on an AI-generated legal conclusion without checking the law. A lawyer should not allow AI to make strategic decisions without human judgment. And a lawyer should not represent to a court, client, or adversary that work has been reviewed when no meaningful human review occurred.
What professional-grade AI use should look like
For lawyers, firms, and legal departments, the responsible path is not to debate AI in the abstract. It is to build a practical governance model.
A defensible AI program should answer at least these questions:
First, what tools are approved for client work? Firms should identify approved tools and prohibit unapproved tools for confidential or privileged material.
Second, what security commitments apply? The firm should understand retention, training, encryption, access control, deletion, audit, and disclosure terms.
Third, how will outputs be verified? Legal authorities should be checked against primary sources. Factual content should be verified against the source documents. By humans.
Fourth, when should clients or courts be told? Disclosure may be required by court rule, judge-specific order, client guideline, engagement letter, protective order, or the nature of the work.
Fifth, how will efficiency be reflected in fees? A legal team that uses AI should be prepared to explain how time was recorded, how value was delivered, and how the client benefited.
These are not barriers to AI adoption. They are what make AI adoption professionally defensible.
The reasonable lawyer in the AI era
The next standard of care will ask not just whether a lawyer used AI. It will ask whether the lawyer made a reasonable, informed, client-centered decision about AI.
AI will never replace the lawyer’s duty of judgment. But it will increasingly shape what competent judgment requires. The lawyer who uses AI recklessly violates the existing standard of care. The lawyer who refuses to consider AI may eventually do the same.
The future is not AI at all costs. And it is not AI avoidance at all costs.
It is professional-grade AI: secure, supervised, verified, and used when it serves the client.
Schedule a demo to see how esumry helps defense teams turn case documents into structured, accessible litigation knowledge that supports trusted, faster, consistent pretrial work.
About the Author
James Chapman is a co-founder of esumry and a defense litigator. He writes about the intersection of AI, litigation strategy, and legal operations.
Using esumry, privilege is protected with ZDR (zero data retention), and case analysis is fast, strategic, and secure. Create timelines, tag testimony, assess credibility, and get ahead of how the other side will use the record—before they do.
