The best second-year attorney submitted a brief with four citations she had never even opened. She ran a search query, pasted the results, and moved on. The brief reads smoothly, but one gets the sense that her thinking stopped somewhere around page two.<\/p>\n
She isn\u2019t lazy in the sense that partners usually complain about. The tool did exactly what it was built to do, in a matter of seconds, and therein lies the trap: the faster a machine provides a plausible answer, the less effort the mind expends to verify whether the answer is trustworthy. This phenomenon can be called the \u201cshortcut curve,\u201d because as retrieving information becomes easier, judgment quietly slips out the door. Speed can be brought back into the firm\u2019s results only if it is guided by rules rather than impressions.<\/p>\n
The solution is not to ban the software. It lies in establishing, from the outset, what a lawyer must be able to do without any assistance and in integrating an AI assistant for lawyers (AI assistant for lawyers<\/a>) within these boundaries, without circumventing them. A drafting tool only improves the work of someone who can already recognize a weak draft. With the boundary set correctly, a junior lawyer\u2019s brief becomes faster and better; with the boundary set incorrectly, an entire generation of lawyers is taught to trust the first result displayed.<\/p>\n No one loses the ability to type a search query. What erodes is the unappealing middle ground of legal work: formulating a hypothesis, pursuing a counterargument, reading the court\u2019s full decision rather than the summary. It is in this middle part that professional judgment is formed, and it is the first thing to go when the answer arrives in four seconds.<\/p>\n The author has observed this pattern in both small defense firms and firms with twenty attorneys. The lawyer who once spent an afternoon wrestling with a bad summons now spends four minutes accepting a good one. She is not wrong about the case, but she never had to win it through hard work, so she never learned where it might falter.<\/p>\n According to the American Bar Association (American Bar Association<\/a>), competent representation includes staying current with the benefits and risks of relevant technology. The wording is worth reading carefully. The obligation concerns understanding the tools well enough to catch them when they fail, without mandating their use. A law firm that installs software and calls this competence has misread the rule.<\/p>\n The \u201cshortcut curve\u201d has a counterpart in experimental research. Betsy Sparrow, Jenny Liu, and Daniel Wegner published the study \u201cGoogle Effects on Memory\u201d in the journal Science in 2011, based on the concept of transactive memory previously proposed by Wegner. The basic idea is that people do not retain everything they know on their own, because they share the burden of memory with other members of a group and remember, above all, who knows what. In the 2011 experiments, participants who believed that a piece of information would be saved on a computer recalled less of its content and more of where it could be found. The authors concluded that the internet has become a transactional memory partner, treated by the mind as a colleague who knows the answer.<\/p>\n Applied to legal research, this finding accurately describes the situation of the lawyer mentioned in the introduction. She remembered where to find the answer\u2014in the legal instrument\u2014without retaining the reasoning that leads to it. For simple information, externalization is efficient and harmless, since no one gains anything by memorizing phone numbers. Legal judgment, however, is more than just stored information: it is formed through the repeated exercise of weighing conflicting arguments and cannot be retrieved from an external system by someone who has never constructed it.<\/p>\n The difference from a human colleague matters here. In a group, transactional memory works because each member knows how reliable the others are in a given field and adjusts their trust accordingly. An artificial intelligence assistant produces equally fluent responses regardless of whether they are correct or incorrect, so the lawyer does not receive the signals of hesitation that they would receive from an uncertain colleague. The three working rules described below compensate for precisely this shortcoming, as they reintroduce verification where the tool offers no indication of its own reliability.<\/p>\n The framework has limitations that must be acknowledged. A large-scale replication study coordinated by Colin Camerer and published in 2018 in *Nature Human Behaviour*, which revisited several social science experiments originally published in *Science* and *Nature*, failed to replicate one of the experiments from the 2011 study, and another replication study, published in 2020 in *PeerJ*, reached a similar conclusion for the same experiment. The findings regarding the prioritization of location over content were obtained from college students using general knowledge statements\u2014a far cry from the work of a motivated professional. The research thus supports a plausible hypothesis about memory externalization, without proving that legal research tools impair judgment, and the test proposed in the following section has the merit of directly measuring this in one\u2019s own practice.<\/p>\n The appropriate exercise for any managing partner concerned about this issue is as follows. A closed case with a known outcome is selected, and a junior attorney is given only the facts of the case. The junior lawyer researches and drafts the basic analysis without any artificial intelligence tools, against the clock. The same task is then repeated with the AI assistant activated, and the two briefs are compared line by line.<\/p>\n Speed is not the criterion being assessed. Three specific indicators matter: whether they identified the same binding case law each time, whether they addressed the most compelling contrary ruling or only the one highlighted by the tool, and whether they can explain aloud why they rejected the rulings they rejected. The third question decides everything. A lawyer who cannot explain their reasoning once the tool is turned off has learned nothing and has merely transcribed the text.<\/p>\n This test is worth repeating twice a year for every junior attorney, with notes kept on record. It is not a trap; it is a fitness test for the very skill that software is imperceptibly eroding, and it requires only an afternoon, at no cost.<\/p>\n This argument is not a call for nostalgia. The assistant is the right choice for tasks where the machine performs better, and the stakes of a mistake are low: first-read summaries of a large volume of documents, extracting all the data from a deposition transcript, drafting standard evidentiary forms that have already been filed forty times, or compiling a client intake summary from a disorganized questionnaire.<\/p>\n Work in which judgment carries significant weight remains human: strategic decisions, witness preparation, and the argument that would be made if the opposing side throws a curveball. This distinction is not a moral judgment on technology; it is a matter of sorting between errors that are inexpensive to catch and those that cost a trial.<\/p>\n One more thing needs to be said. The job market is pushing law firms toward efficiency, whether they like it or not. Data from the Bureau of Labor Statistics (Bureau of Labor Statistics<\/a>) show approximately 863,700 lawyer positions in the United States by 2025 and an estimated 5% growth through 2035, with about 28,700 new positions opening each year. When every law firm can draft documents faster, clients begin to expect speed. Tools will be necessary to remain competitive, and savvy lawyers will be needed to make the firm worth hiring.<\/p>\n Verification rules have a solid empirical foundation. Varun Magesh and his colleagues at Stanford published the first pre-registered study in 2025 in the Journal of Empirical Legal Studies on commercial AI-based legal research tools, which vendors promote as free from hallucinations. Tests showed that the tools offered by major legal database providers produced incorrect answers or citations that did not support the claim in 17% to 33% of cases. The rate is lower than that of general-purpose models, but it remains far too high for a document filed in court without being read.<\/p>\n The type of errors matters just as much as their frequency. A completely fabricated ruling is relatively easy to spot, since it does not appear in any database. Much harder to spot is the citation of a real ruling that does not support the claim for which it is invoked, either because it was subsequently overturned or because the tool attributed to it a conclusion that the court did not actually reach. Such errors slip past a superficial check\u2014in which the attorney merely confirms the case\u2019s existence\u2014and can only be detected by reading the relevant passage. The rule of reviewing each citation, proposed in the following section, must therefore be understood as an obligation to read the court\u2019s ruling, beyond simply confirming that the link works.<\/p>\n The consequences are already measurable. In the case of Mata v. Avianca, tried in 2023 in New York, two attorneys were fined $5,000 after filing a brief containing rulings fabricated by a chatbot. As of early June 2026, the database maintained by researcher Damien Charlotin listed 1,058 cases in the United States in which parties or attorneys had improperly used generative artificial intelligence in documents filed with the courts, with approximately 40% of those cases involving attorneys. The American Bar Association responded in July 2024 with Formal Opinion 512, which applies existing obligations regarding competence, confidentiality, client communication, and supervision to generative tools.<\/p>\n Confidentiality deserves its own rule in any internal policy. Case data entered into a generative tool may be stored, used to train the model, or made accessible to other users, depending on the provider\u2019s contractual terms; the opinion requires attorneys to understand these terms before use and, for tools that learn from the data entered, to obtain the client\u2019s informed consent. A prudent law firm sets forth in writing which tools are approved, which categories of information may not be entered under any circumstances, and who is responsible for periodically reviewing the contractual terms, which may change from one version of the product to another.<\/p>\n An assistant must earn their place through their workflow, not simply by receiving approval. The rules below are worth putting in writing.<\/p>\n The number three was chosen intentionally. Two seem too few, and four tend to be ignored by the second month, whereas three fit on a single index card and hold up through a busy Monday.<\/p>\n The decision criterion that matters more than any comparison of functions is simple. Before renewal, each junior attorney is asked to present a case file in which the assistant changed his or her conclusions. The case in question is one in which the tool caused the lawyer to question their initial reading, to double-check a ruling they might have overlooked, and to arrive at a better conclusion; a case in which the tool merely saved time does not answer the question.<\/p>\n If no one can describe such a situation, the firm has purchased a sophisticated search tool. If half the team can, things are heading in the right direction.<\/p>\n It is worth noting how the rules of professional conduct address supervision. The Department of Labor (Department of Labor<\/a>) publishes occupation-level data for the entire economy, which is useful when a managing partner requests figures for staffing decisions. The clearer argument, however, comes from the profession\u2019s model rules, which hold partners and supervising attorneys accountable for the work of subordinate attorneys and non-attorney assistants. The coordinating attorney is responsible for the work bearing his or her name, regardless of whether the initial draft was produced by a person or a program, and this responsibility cannot be delegated.<\/p>\n The rules above address internal verification within the firm, but the client still faces a similar problem when viewed from the outside. They cannot tell whether the brief they are paying for was read line by line or simply accepted from the screen, and they lack the necessary training to evaluate the citations on their own. Their position vis-\u00e0-vis the firm resembles that of the attorney vis-\u00e0-vis the tool: they receive a polished result and must decide how much trust it deserves.<\/p>\n A few questions asked at the first meeting reduce this asymmetry. The client can ask whether the firm uses artificial intelligence tools, who verifies their results before filing, whether case data is entered into external systems, and how the time saved is reflected in the fee. Formal Opinion 512 explicitly addresses the last two issues, so a U.S. law firm that cannot answer them has not read its own professional rules. A lawyer\u2019s status can be verified separately in the public registry of the bar association in the state where they practice, where any disciplinary sanctions are also listed.<\/p>\n The question regarding fees has a clear basis in the same opinion. An attorney who bills by the hour may bill only for time actually worked; thus, a brief drafted in fifteen minutes using a tool cannot be billed as if it had taken three hours, and the time spent by the attorney learning to use a general-purpose tool is not, as a rule, charged to the client. A client who is familiar with these principles can review a billing statement more carefully and, where appropriate, request an explanation for the hours billed for research and drafting.<\/p>\nWhat Deteriorates When Research Becomes Instantaneous<\/h2>\n
What Cognitive Psychology Says About Externalized Memory<\/h2>\n
Reviewing a case without assistance: a test you can try this week<\/h2>\n
Where Speed Brings Real Gain<\/h2>\n
What the data shows about tool errors<\/h2>\n
Three Rules to Keep Junior Attorneys on Their Toes<\/h2>\n
\n
The Question to Ask Before License Renewal<\/h2>\n
The Same Problem, as Seen by the Client<\/h2>\n