Legal systems unprepared for risks of AI-driven decisions

Artificial intelligence is rapidly moving from legal research assistants into roles that directly shape legal reasoning, drafting, and decision-making. Courts, law firms, regulators, and legal educators are increasingly turning to large language models to manage growing workloads, reduce costs, and accelerate access to legal information. Yet as adoption accelerates, so do concerns about accuracy, accountability, and the risk of delegating core legal functions to systems that generate language without understanding law in the human sense.

A new comprehensive review published in Humanities and Social Sciences Communications warns that the legal sector is entering a critical transition phase. While large language models offer substantial efficiency gains, their limitations pose serious risks in high-stakes legal environments where errors can affect rights, liabilities, and the legitimacy of institutions.

The research, Large Language Models in Law: A Comprehensive Survey of Applications, Challenges, and Future Directions, underscores that the challenge facing legal systems is no longer whether to use AI, but how to do so without undermining trust, fairness, and professional responsibility.

How large language models are reshaping legal work

The study shows that large language models are already embedded across much of the legal workflow. Their most widespread adoption has occurred in tasks traditionally viewed as labor-intensive but low-risk, such as document review, legal research, summarization, and drafting. Law firms use these systems to analyze contracts, extract clauses, identify risks, and produce first drafts of legal documents at scale. Courts and regulators are experimenting with AI-assisted case law retrieval, enabling faster access to precedents and statutory interpretations.

In legal education, the models are being used to support writing, generate study materials, and simulate exam questions, changing how students engage with legal texts. The research notes that these applications have expanded rapidly, driven by the models’ ability to process large volumes of legal language with speed and apparent fluency.

More advanced uses are also emerging. Predictive analytics powered by language models are being tested to forecast case outcomes, assess litigation risk, and support strategic decision-making. Client-facing legal chatbots are increasingly used to provide preliminary legal guidance, particularly in areas such as consumer law, employment disputes, and regulatory compliance. These systems promise to widen access to legal information, especially for individuals and small businesses that cannot afford traditional legal services.

However, the study makes clear that productivity gains come with significant trade-offs. Large language models do not reason about law in the way human lawyers do. They generate responses based on statistical patterns in training data rather than legal logic, precedent hierarchy, or jurisdictional nuance. This distinction becomes critical as models move from support roles into tasks that influence legal judgment.

One of the most severe risks identified is hallucination, where models generate confident but false legal information. The survey documents multiple cases in which AI systems produced fabricated case citations, misrepresented statutes, or invented legal principles. In professional settings, these failures have already resulted in court sanctions and reputational damage, demonstrating that the risks are no longer theoretical.

The research also highlights jurisdictional fragility. Legal systems vary widely across countries, states, and courts, yet most large language models are trained on mixed and uneven datasets dominated by a small number of legal traditions. This creates a high risk of incorrect reasoning when models are applied outside their dominant training context. Even within a single jurisdiction, rapidly evolving case law can render model outputs outdated or misleading.

Ethical, legal, and institutional risks come into focus

The study brings to the fore the ethical and institutional concerns raised by the integration of large language models into law. Bias remains a central issue. Because these models are trained on historical legal texts and broader societal language, they can reproduce and amplify existing inequalities embedded in legal systems. The research points to risks of discriminatory outcomes, particularly in areas such as sentencing analysis, risk assessment, and legal advice affecting marginalized groups.

Transparency and explainability emerge as major obstacles. Legal systems rely on reasoned justification, traceable sources, and accountability. Large language models, by contrast, operate as opaque systems whose outputs cannot be easily explained or audited. This lack of interpretability clashes directly with legal norms that require decisions to be justified and contestable.

Accountability is another unresolved challenge. When AI-assisted legal work leads to harm, responsibility remains unclear. The study shows that existing legal frameworks struggle to assign liability when errors result from a combination of human oversight and machine-generated content. Questions around authorship, professional responsibility, and malpractice exposure are becoming increasingly urgent as AI-generated text enters official filings and legal advice.

Confidentiality and data protection also present significant risks. Legal work often involves sensitive personal, commercial, and state information. The use of cloud-based language models raises concerns about data leakage, unauthorized reuse of inputs, and compliance with privacy regulations. The research stresses that many legal practitioners remain unclear about how their data is processed, stored, or incorporated into future model training.

Intellectual property issues further complicate adoption. The survey reviews ongoing disputes over whether AI-generated legal texts can be copyrighted and who owns outputs produced through human-AI collaboration. These unresolved questions create uncertainty for law firms, publishers, and clients relying on AI-assisted drafting.

The study also examines the growing ecosystem of legal datasets used to train and evaluate large language models. While specialized legal corpora improve performance, they remain fragmented, unevenly labeled, and heavily skewed toward certain jurisdictions. Data scarcity in non-English and less-resourced legal systems raises concerns that AI adoption could deepen global legal inequalities rather than reduce them.

Why human oversight and regulation remain essential

Despite these risks, the study does not argue against the use of large language models in law. Instead, it calls for a clear reframing of their role. The authors emphasize that these systems should be treated as assistive tools rather than autonomous legal actors. Human oversight, professional judgment, and institutional safeguards are presented as non-negotiable requirements for responsible deployment.

The research highlights several technical approaches that can reduce, but not eliminate, risk. Retrieval-augmented generation, which grounds model outputs in verified legal sources, significantly reduces hallucinations and improves accuracy. Domain-specific fine-tuning enhances performance in targeted areas of law. Hybrid systems that combine symbolic legal reasoning with language models show promise, though they remain at an early stage of development.

However, the study stresses that technical fixes alone are insufficient. Governance frameworks must evolve alongside technology. Regulatory bodies, courts, and professional associations need updated standards that define acceptable uses of AI, mandate transparency, and clarify liability. Training for legal professionals must also adapt, equipping lawyers and judges with the skills needed to critically evaluate AI-generated content rather than treating it as authoritative.

The survey points to a growing consensus that explainability, auditability, and accountability should be built into legal AI systems by design. Without these features, the legitimacy of AI-assisted legal processes may be questioned by courts, clients, and the public.