1. What is the situation or context you are addressing?
+Alternate dispute resolution is a practice-based profession but the core practice skills such as mediation, negotiation, client counselling and oral argumentation cannot be taught through coursework alone. They require repetition, failure, and exposure to unpredictable human behaviour in real time. The structural problem is that there is no scalable way to deliver this. Faculty time is limited, legal clinics are unevenly distributed across institutions, and ADR competition circuits which currently serve as the primary training ground are accessible to a small fraction of students.
+The result is that most law students encounter the real pressure of a negotiation or mediation for the first time during internships or early practice, where the cost of failure is no longer pedagogical. This is not unique to law. Medicine, counselling, and conflict resolution share the same bottleneck which means that the skills that matter most are the ones that depend on practice access, and practice access does not scale under the current model.
+This gap is visible even among students who compete internationally in ADR. Representing JGLS in mediation competitions, we noticed that the technical work like drafting opening statements and setting agendas was consuming the cognitive bandwidth that should have gone toward reading the room, noticing when a party's language shifted, deciding whether to follow an emotional thread or hold the structure and managing silence. The procedural scaffolding was crowding out the human skill that mediation actually depends on.
+2. What is your critical analysis of this situation?
+Two responses to this problem are common, and both are insufficient.
+The first is to increase supervised practice hours. This works pedagogically but fails at scale. A single faculty member running a mediation simulation can engage perhaps 6–8 students per session. Multiplied across a cohort, this requires resources most law schools do not have and will not have.
+The second is to use AI as a general-purpose study tool that can help summarise case law, generate practice questions and draft documents. This addresses productivity but does nothing for the practice gap. Knowing the law and applying it under interpersonal pressure are fundamentally different competencies, and current AI integration in legal education conflates them.
+A harder question sits underneath both responses. If the bottleneck is access to realistic practice, and AI can now simulate conversational partners with adjustable behaviour, can it serve as a practice environment for human skills like negotiation and mediation? And if so, what are the precise boundaries of what it can and cannot develop?
+This is not a settled question. There is a genuine risk that practising against a language model builds pattern recognition that holds in simulation but breaks in the room where tone of voice, body language, power dynamics, cultural context, and genuine emotional stakes are all present. A student who performs well against a simulated party may develop false confidence that collapses on first contact with a real counterpart. This risk is not hypothetical; it is the central design constraint of any AI-assisted practice system for interpersonal skills.
+There is also a second risk that is less discussed, which are AI tools used for legal training can easily become tools for outsourcing cognition rather than sharpening it. If a student uses AI to generate their negotiation strategy rather than develop it themselves, the tool has replaced the learning it was meant to support. The line between cognitive enhancement and cognitive replacement is not always obvious, and system design determines which side students land on.
+3. What perspectives were discussed and how were they debated?
+Our team debated three tensions that shaped every design decision.
+The first tension is between simulation fidelity and pedagogical value. One position held that the AI must simulate human behaviour as realistically as possible with emotional volatility, irrational anchoring and cultural nuance or it teaches the wrong instincts. The counter-position argued that for early-stage learners, a controlled and predictable simulation is more useful than a realistic one, because it allows students to isolate and practise specific skills before facing the full complexity of a real interaction. We resolved this through difficulty calibration: the system supports a range from procedurally structured scenarios for beginners to emotionally volatile, multi-issue disputes for advanced practitioners. But we remain honest that even the highest difficulty level cannot replicate genuine interpersonal stakes.
+The second tension is between AI-driven feedback and reflective learning. An AI system can identify what a student did, the moves they made, the sequence and the language used. It cannot reliably assess why a particular move worked or failed in context, because that judgment depends on reading the other party's internal state, which the AI is also simulating. We debated whether detailed AI-generated feedback would encourage students to treat the debrief as authoritative correction or as a starting point for their own analysis. This led to a design decision wherein the debrief engine provides structured observations on practitioner moves on what was said, what options were available and what the simulated party's stated interests were, paired with open reflective prompts rather than evaluative scores. The goal is to provoke thinking, not replace it.
+The third tension is between scalability and supervision. AI practice is only valuable if it feeds into, rather than substitutes for, supervised human interaction. But if the system requires faculty oversight to be effective, it does not solve the scalability problem it was designed to address. Our current position is that the AI practice layer must be explicitly framed within the system itself as preparation for supervised sessions, not a replacement. The system should tell learners what it cannot simulate. That honesty is a design feature, not a disclaimer.
+4. What contribution are you proposing, and under what conditions could it be implemented?
+We propose a domain-specific AI practice environment for negotiation and mediation training, which we are calling the ADR Practice Simulation Framework. It has three components.
+Current stage and early observations. A working prototype has been built and tested informally with law students at JGLS who have prior ADR competition experience, during preparation for their competitions. These were not controlled experiments but practitioner-led sessions where the team used the tool and observed its behaviour. Three observations emerged. First, the simulation was most useful for practising procedural moves like agenda sequencing, reframing techniques and interest-mapping where students could repeat and refine specific skills in a low-stakes environment. Second, the simulation was least convincing when emotional dynamics were central to the scenario; the AI's emotional responses followed predictable patterns that experienced students quickly learned to exploit rather than genuinely navigate. Third, the debrief output was more useful when students treated it as a prompt for self-assessment than when they read it as evaluative feedback suggesting that the framing of the feedback matters as much as its content.
+These observations are preliminary. They do not constitute validation. What they do provide is a set of testable hypotheses for the next stage.
+Implementation conditions and research agenda. The prototype can be deployed at any law school with internet access and no additional infrastructure. The next stage requires cross-institutional testing with a structured evaluation design. The core research question. Does AI-assisted practice, combined with supervised human interaction, produce measurably better negotiation and mediation practitioners than supervised human interaction alone? This requires pre/post measurement of specific competencies, controlled comparison across student cohorts with varying levels of prior ADR exposure, and a real-world transfer measure that tests performance in live human interactions, not just in-simulation metrics. We are also investigating what feedback architecture produces genuine reflection rather than passive correction, and where the system needs to be transparent with learners about its limitations.
+We are not proposing a finished product. We are proposing a research-grounded practice tool, with a working prototype, a set of early observations, and a clear experimental design for answering a question the field has not yet seriously addressed.
+
Share