AI in Legal Practice: Professional Judgment, Prompts, and Liability
The professional value of the attorney is shifting, and this displacement demands a response more sober than technological marketing and more rigorous than reflexive rejection.
I have spent several years using generative artificial intelligence for parts of my work in industrial property and amparo litigation. My primary uses involve stress-testing theories of the case, refining drafting, auditing argumentative lines, and building comparative maps that manually would take an entire afternoon. What I will state here is driven neither by enthusiasm nor fear, but by a simpler reality: the professional value of the attorney is shifting, and this displacement demands a response more sober than technological marketing and more rigorous than reflexive rejection.
For decades, much of an attorney’s prestige lay in accessing information that others lacked (case law, doctrine, precedents, forms, tools). That asymmetry still exists, but it is closing. A system trained on millions of legal texts and connected to specialized databases can now produce, in seconds, a draft that would have taken a junior associate hours. What the system cannot do—yet—is take responsibility for what it produces. The differential between generating a text and assuming its content and legal consequences is where professional practice finds its new challenge today.
I. The Error of Confusing Information with Judgment
During my legal education, my professors insisted on a maxim that initially seemed like mere relief from the study burden, but which I understood from my first years in practice: the attorney does not need to memorize the law. Statutes, codes, and regulations are amended, abrogated, or repealed with a frequency that renders memorization useless. The central lesson was that professional rigor resides in legal judgment; that is, in the dexterity to interpret legal texts, anticipate their effects, and articulate them strategically in favor of the interests being defended.
I revisit that premise because the public discussion on the adoption of AI in law often centers precisely on the same error: it takes for granted that a system capable of retrieving and reorganizing legal information exercises, by extension, legal reasoning. It does not. Language models predict probable sequences of words based on their training corpus. They imitate the discursive genre of a judgment, a closing argument, or a legal opinion with remarkable fluency, but that syntactic articulation lacks normative comprehension and judgment.
This distinction is relevant because the law does not operate with the raw information contained in a provision, but rather with what the provision means in a given context. For example, knowing that Article 167 bis of the Regulation on Health Inputs governs the pharmaceutical linkage system, or that the Federal Law for the Protection of Industrial Property (Ley Federal de Protección a la Propiedad Industrial, LFPPI) amended in 2026 expanded the powers of the Mexican Institute of Industrial Property (IMPI), is merely available information that the machine retrieves better than anyone. However, what the machine does not decide is whether it is advisable to trigger that mechanism in a specific case, what evidence to offer, how the Federal Commission for the Protection against Sanitary Risks (COFEPRIS) will react, how much risk the client assumes by initiating the proceeding, or at what moment it is better to negotiate than to litigate. That plane—the plane of judgment—remains human for an essential reason: the system does not bear the consequences of its outputs.
Seen in this light, the displacement hypothesis is more precise than usually framed. AI does not replace the attorney; rather, it redefines the value of what the attorney did when merely brokering information, and heightens the value of what the attorney does when supervising, validating, and answering for partially automated processes. Those who fail to make this transition will end up competing, in a saturated market, for the tier of work where the price becomes indistinguishable from that of a model. I have argued previously that Mexico carries a deeper technological strategy deficit than its legislative void regarding AI, and that same deficit, applied to legal practice, reappears as a lack of judgment on how, when, and under what conditions to delegate.
II. The Cognitive Infrastructure of the Modern Attorney
The integration of artificial intelligence into law firms transcends the adoption of a simple tool; it is an infrastructure layer placed between the attorney and their work. The announcement of Claude for Legal by Anthropic in May 2026 (over twenty MCP connectors, twelve plugins per practice area, and launch partners including Freshfields and Accenture) indicates the trajectory: AI ceases to be an external chatbot and becomes the governance interface for legal work, integrated with document repositories, e-discovery platforms, contract managers, and precedent databases. Westlaw and other historical catalogs are consenting to be consumed from within these interfaces, reordering the hierarchy of providers that sustained legaltech for several years. I developed this point in my analysis on Claude for Legal and the new cognitive infrastructure of Mexican law.
For Mexico, this infrastructure will arrive; what is in question are the conditions. Current models are trained on Anglo-American law and international corporate practice, meaning that Mexican evidentiary practice, the particularities of the amparo trial, IMPI criteria on distinctiveness and descriptiveness, administrative litigation practices, or specific contradictions between IMPI and COFEPRIS appear in their corpora at a much lower resolution. A firm that adopts these tools without implementing local validation layers (proprietary taxonomies, internal databases, review controls) will inherit superficial efficiency and dangerous fragility.
There is, additionally, a formative problem that law schools and bar associations will primarily need to address. The portion of work that AI is absorbing most easily (procedural chronologies, standard contract review, first drafts of simple pleadings, precedent searches) is exactly what served for years as training for young attorneys. If that friction simply disappears, someone will have to purposefully design how to forge the judgment that was previously formed there:
A generation of attorneys armed with potent tools and fragile judgment would be worse than one without tools.
III. The Prompt as a Professional Instrument
I have argued elsewhere that prompt engineering is not a computer trick to gain followers; it is a contemporary form of the old dialectical exercise, and I repeat the argument here because I find it central. A strategic legal prompt is not a colloquial question to the model. It is a densely coded instruction that loads, within the context window, the exact factual basis, the applicable jurisdiction, the theory of the case already constructed by the attorney, the requested analytical operation (expansion, contrast, red teaming, extraction), and the constraints: what the model must not do, what it must not invent, and in what format it must deliver.
Asking a system to "act as an expert attorney in litigation, M&A, contracts, industrial property, administrative law, labor law, property law, and criminal law" is not the same as asking it to "subject to argumentative stress, in tabular format, a defense theory built upon three precisely delimited axes." The difference lies in structure: the first prompt treats it as a kind of oracle; the second, as what it is—a tool. Only the second is compatible with honest practice.
Some jurisprudence is beginning to move in that direction. The isolated thesis II.2o.C.9 K (11a.), titled "ARTIFICIAL INTELLIGENCE APPLIED IN JURISDICTIONAL PROCEEDINGS. MINIMUM ELEMENTS TO BE OBSERVED FOR ITS ETHICAL AND RESPONSIBLE USE WITH A HUMAN RIGHTS PERSPECTIVE," issued by the Second Collegiate Tribunal in Civil Matters of the Second Circuit and published in the Semanario Judicial de la Federación on August 22, 2025 (digital register 2031010), established the minimum elements for such use in Mexico: proportionality and harmlessness, protection of personal data, transparency and explainability, and human supervision and decision-making. The thesis mandates that "the adjudicator shall not only report the use of such tools but must also set forth the methodology, the data employed, and how the result was reached," warning that supervision must ensure "that the technology operates as an auxiliary and not as a substitute, keeping deliberation and decision-making strictly within the jurisdictional domain."
The notable aspect of that precedent lies in the method, not merely the fact that a tribunal used AI. Magistrate Juan Jaime González Varas did not ask the machine to resolve the appeal. He provided the reasoning, defined the calculation operation (adjusting a procedural guarantee based on cadastral value, INEGI inflation, Banxico interbank interest rates, and estimated trial duration), and asked the system to execute it with traceability documented within the judgment itself. The result was lower than originally set, but verifiable and reasoned. It is, in practice, a strategic legal prompt applied by a federal body:
The human provides the judgment, the machine processes, and traceability allows the result to be audited.
Any other architecture failing to meet these parameters culminates in exactly what the precedent seeks to prevent.
IV. Privacy, Professional Secrecy, and Governance
One of the most complex and severe problems in using these tools is found in the handling of client information. When an attorney pastes client information (a draft contract, a procedural strategy, a risk matrix, an expert opinion, internal emails) into a chat window, they are processing personal data subject to the Federal Law on Protection of Personal Data Held by Private Parties (Ley Federal de Protección de Datos Personales en Posesión de los Particulares, LFPDPPP) and, if the provider's servers are outside Mexico, to international transfer obligations. Prior filtering or pseudonymization mitigates the risk but does not erase it. As I explained in my analysis on privacy filters in artificial intelligence, a filter eliminates obvious identifying evidence, not the system's capacity to infer from what remains: position, sector, location, medical condition, or specific labor conflict.
The Samsung incident in April 2023 serves less as a corporate anecdote than as a warning. Bloomberg documented that the company detected three separate information leaks in under twenty days: one engineer pasted confidential semiconductor source code to debug errors, another uploaded a defect-detection algorithm to have it optimized, and a third transcribed an internal meeting and fed it to the system to generate minutes. The company ultimately banned generative AI on its devices. If highly qualified engineers behaved this way toward a tool that seemed harmless to them, I see no reason to assume the average law firm will fare better, absent explicit internal governance.
A single ill-conceived prompt is enough to violate professional secrecy.
This is why the opinions recently published by bar associations converge on a point that should be taken seriously. The ABA Formal Opinion 512, from the Standing Committee on Ethics and Professional Responsibility (July 29, 2024), requires the attorney to maintain active technological competence under Model Rule 1.1, obtain informed client consent under Rule 1.4 when the system is "self-learning," protect confidentiality under Rule 1.6, uphold candor toward the tribunal under Rule 3.3, supervise under Rules 5.1 and 5.3, and charge reasonable fees under Rule 1.5. The Florida Bar Ethics Opinion 24-1 (January 19, 2024) follows the same path, adding that a chatbot interacting with clients must identify itself as an automated system and not pass as an attorney.
In Mexico, the Guidelines of the Mexican Bar Association for the Responsible Use of Artificial Intelligence in the Legal Profession are among the first of their kind in Latin America. They establish concrete minimum competencies: knowing the capabilities, limits, biases, and "hallucinations" of the system; requesting client consent when use involves sharing information with external systems; disclosing use to tribunals in certain scenarios; actively mitigating biases; and never delegating professional judgment. The liability guideline leaves no doubt: "Final responsibility for the quality, accuracy, and adequacy of all legal services rendered rests exclusively with the individuals practicing law, regardless of the AI tools utilized."
On a broader plane are the Chapultepec Principles, presented by SECIHTI and ATDT on January 29, 2026. They are non-binding but orientative: "Artificial intelligence must expand rights, never reduce them"; "Every decision supported by AI must have human parties responsible"; "If a decision cannot be explained, it should not be automated."
Comparatively, the European AI Act, generally applicable since August 2026, raises the standard by classifying systems assisting in legal interpretation as high-risk, imposing accuracy metrics and mandatory human oversight. Although this regulation does not govern Mexico, it marks the international trend toward heightened scrutiny in technical documentation, accuracy evaluations, human supervision, and system auditing.
V. When AI Errs: Liability and Reliance
The case of Mata v. Avianca, resolved by Judge P. Kevin Castel of the Southern District of New York on June 22, 2023, is a notorious fact for any responsible attorney involved with AI. Attorneys Steven A. Schwartz and Peter LoDuca submitted an opposition to a motion to dismiss containing six decisions fabricated by ChatGPT, complete with citations to the Federal Reporter and internal references to non-existent cases. When questioned, rather than retracting immediately, they doubled down and produced an affidavit with excerpts that were also model fictions. Schwartz testified under oath that he acted "under the false belief that this website could not produce completely fabricated cases." Castel found sufficient subjective bad faith under Rule 11 of the Federal Rules of Civil Procedure, imposed a five-thousand-dollar fine on each attorney and the firm, and ordered them to notify each real judge falsely attributed as the author of the invented opinions. The fine was small, but the reputational blow is not repaired with money.
French academic Damien Charlotin maintains a public tracker of cases where a tribunal has detected AI-fabricated content in legal briefs. As of June 1, 2026, his database reports approximately 1,459 cases worldwide—a figure growing daily—up from roughly 1,348 in April 2026. He summarized the pace shift: "Before this spring of 2025 we had perhaps two cases per week. Now we are at two or three cases per day." Sanctions have stiffened. In September 2025, in Noland v. Land of the Free L.P., the California Second District Court of Appeal fined attorney Amir Mostafavi $10,000 after finding 21 of 23 citations in his appellate brief were fabricated. In July 2025, in Lindell, the District of Colorado sanctioned Christopher Kachouroff and Jennifer DeMaster $3,000 each. The opinion in Johnson v. Dunn (N.D. Ala., July 2025) went further: the tribunal expressly stated fines were proving ineffective, disqualified the offending attorneys from the remainder of the case, ordered the matter reported to professional associations, and directed the opinion be published in the Federal Supplement.
In February 2026, via order AC739-2026, the Civil, Agrarian, and Rural Cassation Chamber of the Supreme Court of Justice of Colombia fined an attorney fifteen minimum wages (roughly twenty-six million Colombian pesos) for presenting ten non-existent judgments in an extraordinary review appeal. To my knowledge, it is the first pecuniary sanction against an attorney in Latin America for AI-fabricated case law. In Argentina, the Civil and Commercial Appellate Chamber of Rosario rebuked an attorney for invented citations, notifying the Rosario Bar Association "for the purposes of taking cognizance of the problematic situation aroused by the reflexive use of generative artificial intelligence chatbots." In Mexico, I have yet to locate a documented case of sanction on these grounds, though given thesis II.2o.C.9 K, the Bar Guidelines, and the sheer arithmetic of usage, it is hard to believe it will take long.
Regarding chatbot liability, a useful precedent is Moffatt v. Air Canada, 2024 BCCRT 149 (February 14, 2024). The airline's chatbot assured Jake Moffatt he could claim a bereavement fare within ninety days post-travel, while the actual policy—accessible via hyperlink from the bot itself—stated the opposite. Air Canada argued that the chatbot should be viewed as a separate legal entity responsible for its own actions. The tribunal flatly rejected this: "While a chatbot has an interactive component, it is still just a part of Air Canada's website." It applied the standard of negligent misrepresentation and concluded the company breached its duty of care. The damages awarded were $650.88 CAD plus costs: irrelevant in amount, decisive in precedent.
The translation to professional practice is direct: the attorney cannot treat the AI as an agent separate from their own performance. The signature at the bottom of a brief remains what it was before these models existed: a declaration that the signatory read, validated, and assumes everything the document states.
Judge Liebowitz of the Southern District of Florida phrased it well in ByoPlanet:
"Perhaps twenty years from now AI will be flawless; when that day arrives, that flawless brief will only have meaning because of the human signature at the bottom."
I do not know of a more economical way to state what is at stake.
VI. IMPIBot and the Risks of Automated Institutional Guidance
I previously wrote about IMPIBot, the first conversational assistant of the Mexican Institute of Industrial Property, deployed on WhatsApp in late 2025. It allows querying tobacco trademark registrations, COFEPRIS sanitary alerts, and SAT Annex 11 in a single conversation. Its value lies in reducing friction to access data the State already holds, not in generating new information.
What matters is the flip side: if Moffatt makes clear a company answers for its chatbot, the ensuing question is what happens when the guidance comes from the State. The case of MyCity, launched by Mayor Eric Adams in October 2023 on Microsoft Azure infrastructure, serves as a warning. Investigations revealed the bot systematically stated falsehoods that were, in several cases, illegal (e.g., that employers could keep worker tips or landlords could discriminate against housing vouchers). Adding a "beta version" badge after the fact does not repair the reality of an official channel disseminating illegal guidance. The incoming mayor eventually shuttered the system, calling it "functionally useless."
Brought to Mexico, the issue is serious. What if IMPIBot, or the next assistant deployed by an authority, misadvises a merchant? And if a consumer reports an authentic product as counterfeit relying on outdated system information? Who answers then—IMPI, the developers, or no one—is exactly what we currently lack clarity on. Disclaimers stating "this response is not binding" do not erase the legitimate trust generated by a state channel. The principle of unity of administrative action, the same that makes it absurd for one State authority to contradict another (as I argued regarding patent registry publicity), compels us to view with suspicion any attempt to partition liability among the operating body, the training provider, and the inquiring citizen.
When a private company fails, damage can be contained in contract or consumer venues. When an authority fails, the cost is spread over the citizens who relied upon it and the institutional legitimacy it wields. IMPI's operational acceleration following the LFPPI reform, which I commented on with caution in Mexico’s Industrial Property Office Has Started to Move. Now We Need to See Whether It Can Sustain It., should see its AI modernization measured by the same caution: the solidity of the institutional layer supporting the tool (data quality, update protocols, error metrics, appeal avenues, and clear assignment of accountability).
VII. Professional Value in an AI-Assisted Practice
The value of the attorney, as stated initially, is migrating from access to information toward the capacity to supervise, validate, and answer for partially automated processes. This has practical consequences for practice in Mexico.
The first is that technical competence in AI use is no longer optional. The duty to remain abreast of relevant technology has been redefined. Knowing what a model does and does not do, where it hallucinates, how to contain its behavior, what confidentiality risks it introduces, and how to audit its outputs is no longer lateral knowledge: it is part of the minimum standard of diligence.
Then there is the money. Billing models will have to adjust. The 2025 Generative AI in Professional Services Report notes that 40% of professionals expect alternative fee arrangements to increase due to generative AI. Jack Newton of Clio framed it starkly: "Something that used to take you five hours will now take you five minutes, and you will need to justify those four hours and fifty-five minutes you’ve given up because they no longer reflect value." In Mexico, where hourly billing still dominates, that conversation is just beginning.
And the most delicate consequence remains the algorithmic capture of judgment. When plugins codify escalation thresholds, risk tolerance matrices, drafting styles, and negotiation heuristics, they convert the firm's most valuable intangible into an executable flow. This may preserve house knowledge efficiently, but if that judgment is objectified in external infrastructure, the firm ceases to control its core asset. The substantive work is no longer deciding whether to use AI, but erecting proprietary validation layers, local taxonomies, audit criteria, and professional secrecy controls that prevent a suffocating dependency.
Conclusion
Whoever emerges best from this stage understands something neither frontal rejection nor unmitigated enthusiasm grasps: the signature at the bottom of the brief remains the sole point where the legal system imputes consequences, and that fact orders everything else. AI can expand searches, stress-test theories, refine drafting, and organize unmanageable volumes of data.
What it cannot do is answer for what it produces.
My stance, after years of research and intensive use, is deliberately demanding of those who practice this profession. If we use these tools, we must know them internally, contain their exposure to sensitive information, document our validation criteria, notify the client when use is non-routine, and accept that the duty of diligence is not diluted by an intervening model—it becomes more rigorous. The professionally and ethically indefensible position is the middle ground many firms currently inhabit: using AI silently, without internal policies, serious verification, or training, hoping nothing happens until it does.
In industrial property, where I conduct most of my practice, all of this becomes highly concrete. IMPI office actions, COFEPRIS contradictions, administrative declarations, and amparos against registration denials already coexist with tools churning out plausible drafts. The line between a useful brief and a sanctionable one will increasingly ride on prompt quality, verification traceability, and the honesty with which the attorney assumes what they delegate.
Technology will continue to advance with more precise models and denser connectors. None of that resolves the underlying issue of who answers when something fails. Competing by efficiently processing information against a statistical model is futile. Technological refinement will not alter our profession's foundational premise: as long as the legal system demands a subject to impute consequences upon, the core of legal work will remain the inalienable exercise of judgment.
Lecturas relacionadas