91��ɫ

By Jake Okechukwu Effoduh, Assistant Professor, Lincoln Alexander School of Law, Toronto Metropolitan University

Although I am not a computer scientist, my work as a law professor and legal scholar has long engaged with the capacity of legal frameworks to respond to technological disruption. In that room, surrounded by experts who could articulate attention mechanisms and tokenisation with casual fluency, it was evident that I had as much to learn as I had to offer. At the same time, the exchange highlighted that legal training contributes a distinct and necessary perspective. Lawyers are trained to identify risk, anticipate misuse, evaluate harm, and analyze the real-world consequences that may flow from algorithmic failures. The experience therefore underscored a broader insight, one that the AI industry has perhaps been slow to recognise: that effective red teaming exercises need more lawyers.

The Evolution of Adversarial Testing

Red teaming traces its lineage to Cold War military strategy, where designated adversaries (the “red team”) would simulate enemy attacks to expose defensive weaknesses. The practice migrated to in the 1990s, where penetration testers probed corporate networks before malicious hackers could. When applied to AI systems, red teaming evolved from what hobbyists once called “jailbreaking” (the sport of coaxing chatbots into misbehaviour) into a endorsed by governments and embedded in emerging regulatory frameworks.

Today, AI red teams attempt to elicit dangerous capabilities: the generation of misinformation, the production of discriminatory content, the leakage of private training data, or the provision of instructions for genuine harm. The widely discussed and the , both identify adversarial testing as essential to responsible development. Major AI labs now routinely engage external experts to probe their systems before release, a practice that has become standard (even obligatory) in the field.

Yet red teaming is not without limitations. Exercises are necessarily time-bound; testers cannot anticipate every deployment context or user intention. Findings do not always translate into engineering solutions. And crucially, red teams often lack the interdisciplinary breadth necessary to identify the full spectrum of potential harms. A team composed primarily of computer scientists and ethicists will see certain vulnerabilities clearly and remain blind to others entirely.

What Legal Eyes See

During our red team exercise, while my colleagues focused generally on whether a model would generate violent content or produce biased outputs, my attention gravitated toward a different set of questions. When the model confidently asserted false statements about a living person, I immediately thought of : the constituent elements of the tort, variations across jurisdictions, and the practical difficulty of establishing fault when the “speaker” is an algorithmic system. Similarly, when we successfully extracted fragments of what appeared to be training data, my focus shifted to data protection law, in particular the , and the question of whether such extraction might amount to a violation of data protection obligations.

These were not simple peripheral concerns. They were legally cognisable harms, with potential implications for regulatory scrutiny, reputational damage, and liability exposure. Yet without legal expertise in the room, they might have been catalogued simply as “problematic outputs” rather than understood in their full juridical dimension. This experience thus illuminated why lawyers should be integral members of red teams, not as afterthoughts or compliance consultants, but as core participants in the adversarial imagination that safety testing demands.

Five Reasons to Invite Lawyers to Break Your AI

1. Lawyers understand harm with juridical precision

To begin with, legal training brings a distinctive precision to the identification and classification of harm. The law does not treat all harms equally. Defamation, discrimination, privacy violations, and intellectual property infringement each carry specific definitions, evidentiary burdens, and remedial consequences. When a model produces troubling outputs, lawyers can distinguish between the merely objectionable and the actionable, a distinction that fundamentally shapes risk assessment and mitigation priorities. The in the United States illustrates how model failures can rapidly become legal crises.

2. Privacy law requires specialised fluency

A related consideration arises in the domain of privacy, where legal analysis demands specialised fluency. The global patchwork of data protection regimes, the in Europe, the in California, , , and Kenya’s , have created intricate and sometimes contradictory obligations. When red teamers probe for memorization of personal data or attempt to extract training data, lawyers are equipped to assess whether such vulnerabilities implicate specific regulatory requirements and to anticipate how data protection authorities may respond.

3. Lawyers are professionally trained in adversarial reasoning

Equally important is the legal profession’s grounding in adversarial reasoning. The courtroom is, in essence, an institutionalised red team: opposing counsel systematically probes weaknesses in argument, inconsistencies in evidence, and ambiguities in position. This mode of reasoning cultivates a particular cognitive disposition-the habit of thinking from the perspective of an adversary-that translates directly to the work of identifying how malicious actors might exploit AI vulnerabilities. Lawyers are trained, quite literally, to find the holes and the points of failure

4. Legal expertise enables regulatory foresight

The AI governance landscape is shifting rapidly. The imposes binding obligations on high-risk systems. Numerous . Countries across Africa, from to , have developed national AI strategies. Lawyers can stress-test systems against forthcoming requirements, helping developers prepare for regulatory futures rather than merely react to regulatory presents.

5. Lawyers navigate cross-jurisdictional complexity with native facility

A model deployed globally must contend with the legal frameworks of dozens of jurisdictions simultaneously. What constitutes prohibited hate speech in differs from what triggers liability in the United States, which differs again from prohibitions in or . Lawyers versed in comparative and international law can map how a single model output might cascade into disparate legal consequences across borders, a complexity that homogeneous technical teams are ill-equipped to navigate.

An Invitation and an Imperative

To fellow lawyers who have watched the AI revolution from the side lines, uncertain whether our expertise has purchase in this domain: the answer is that it does. The skills we have cultivated through legal training (close reading, adversarial thinking, the anticipation of misuse, the parsing of ambiguity) are precisely what red teaming demands. I don’t think we need to fully understand the mathematics of transformer architectures to recognise that a model’s confident fabrications might constitute negligent misrepresentation, or that its profiling capabilities might violate privacy law or raise constitutional questions under . Legal analysis operates at the level of consequences, responsibility, and rights, precisely where many AI risks ultimately materialise.

To AI developers and safety organisations: we need diversified red teams. The homogeneity that characterises many current exercises (dominated by security researchers and machine learning engineers) leaves some critical vulnerabilities unexamined. Legal scholars, practising attorneys, and even law students possess perspectives that will reveal risks invisible to those who share similar training and assumptions.

The stakes warrant this expansion. As AI systems assume greater roles in consequential domains (employment, credit, criminal justice, healthcare, content moderation), the potential for legally significant harm multiplies. I think that red teaming that excludes legal expertise is red teaming that leaves the courtroom door unguarded. In my own journey from legal academic to participant in AI safety exercises, I have learned that, sometimes, the most valuable insights emerge from disciplinary friction. Questions that may initially appear naive to engineers can illuminate genuine legal and societal risks. This is why we need more of that productive dissonance. We need more lawyers willing to tackle the machines.

---