A growing number of ELT teachers are watching AI reshape their own field, so it’s fascinating to see how another high-stakes profession is grappling with similar questions. A new open-access survey by Dehghani, Dehghani, Naderzadeh Ardebili & Rahnamayan (Humanities & Social Sciences Communications, 2025) maps out how large language models are being used across legal systems worldwide. Although the context is law, the themes echo many of our own debates in language education.
The study
The authors conducted a wide-ranging literature review of recent work on LLMs in legal practice, legal education, judgement prediction & legal text processing. Rather than running experiments, they synthesised findings from dozens of empirical studies, model evaluations & domain‑specific benchmarks. Their scope included:
- applications in legal drafting, research, compliance & client communication
- performance of models like GPT-4, BLOOM & LLaMA on legal exams & reasoning tasks
- domain‑specific models (e.g. Lawyer LLaMA, ChatLaw)
- datasets used to train or evaluate legal LLMs
- risks, limitations & ethical concerns
In other words, the report is a snapshot: what LLMs can currently do in law, where they fail, & what this means for the future.
The findings
Several patterns stand out:
- LLMs are already useful for routine legal tasks. Drafting simple documents, summarising case law & answering basic queries are areas where models perform reliably. In some studies, GPT‑4 reached or exceeded passing thresholds on components of the US Bar Exam.
- Performance drops sharply when tasks require deep, domain‑specific reasoning. For example, models struggle with “issue spotting” (identifying legally relevant details in a scenario) or applying jurisdiction-specific case law.
- Hallucinations remain a serious risk. The paper cites real cases where lawyers submitted fabricated citations generated by AI.
- Bias is a documented problem. Studies show persistent demographic bias in outputs, with implications for fairness in legal decision‑making.
- Ethical & regulatory gaps are significant. Issues include confidentiality, accountability, authorship, transparency & the risk of over‑reliance on automated reasoning.
Why this matters for ELT
Admittedly, law is not language teaching by any measn, but the parallels with language education are striking. Both professions rely on interpretation, judgement, contextual nuance & ethical responsibility. The legal field’s experience offers a preview of what happens when LLMs enter a domain where language is the medium, the tool & the product.
Three reflections for language educators:
- Fluency ≠ accuracy. A model that “sounds right” can still be wrong. This is as true for grammar explanations as it is for legal advice.
- Domain knowledge matters. Just as legal models need legal corpora, ELT-oriented tools need pedagogically sound, learner-appropriate data.
- Human oversight is non-negotiable. The legal sector’s cautionary tales remind us that AI should support, not replace, professional judgement.
Teacher takeaways?
- Treat AI outputs as drafts, not decisions. Encourage learners to question, verify & compare.
- Use AI to model reasoning, not just produce answers. Prompting for explanations (“Why?” “How do you know?”) mirrors legal prompting strategies like Chain-of-Thought.
- Discuss AI ethics explicitly. The legal field shows how quickly bias, misinformation & overconfidence can cause harm.
How are you navigating the balance between AI support & human judgement in your own teaching?
Leave a Reply