Human-in-the-Loop: Navigating AI Ethics and the Thai Supreme Court Guidelines

The debate inside Thai law firms about whether to adopt AI has largely settled. The question in 2026 is not whether AI belongs in a legal practice but who is responsible when it gets something wrong.

Two regulatory frameworks now give a clear answer. In October 2025, the President of the Thai Supreme Court issued guidelines on AI use in legal proceedings, establishing disclosure requirements and reaffirming that accountability rests with the parties and their lawyers, not the tools they used. The Ministry of Digital Economy and Society established Thailand’s AI ethics framework in collaboration with Mahidol University, anchored on transparency, accountability, and human oversight. The direction from both is consistent: AI is permitted as a professional tool; the professional remains fully liable for the output.

For senior partners at boutique Thai law firms, this is not a warning to slow down AI adoption. It is a framework for adopting it correctly.

What the Supreme Court Guidelines Actually Require

The guidelines issued by the President of the Supreme Court on October 9, 2025, address AI use in documents submitted to Thai courts. The requirements are specific.

Parties must disclose when generative AI was used to draft legal documents, including pleadings and other court filings. The obligation is on the party and their counsel to inform the court of AI involvement, not to avoid it. AI-assisted drafting is not prohibited. Undisclosed AI-assisted drafting is a different matter.

Beyond disclosure, the guidelines establish that attorneys must “verify results obtained from AI consistently” with respect to document content, legal citations, and case law references. This is the critical requirement. The court is not asking lawyers to distrust AI output. It is requiring that they read it, check it, and own it before it reaches a judge.

The guidelines also flag data sensitivity: lawyers must consider whether entering client information into AI prompts risks inadvertently disclosing that data to third parties. The instruction is to exercise caution about what enters the prompt, not to avoid AI entirely.

The Accountability Architecture the Guidelines Imply

The Supreme Court guidelines reflect a consensus that has emerged from the Council of Europe, UNESCO, and OECD frameworks that Thailand has aligned with. That consensus has a specific architecture.

AI produces a first draft. A qualified human reviews it, verifies its accuracy, checks its legal citations, and signs off before the output is relied upon. The AI is a production tool. The human is the accountable professional. This is the Human-in-the-Loop model, and it is now the expected standard in Thai judicial proceedings.

For senior partners who have been hesitant about AI adoption on professional liability grounds, the guidelines actually provide clarity rather than adding risk. The liability question under Thai law was always going to resolve the same way: the signing lawyer is responsible. The Supreme Court guidelines formalise that principle in the context of AI-assisted work. Adopting AI with proper review procedures does not change your liability exposure. Adopting AI without review procedures, or adopting it and not disclosing its use when required, does.

The Ministry Framework: Ethics Grounded in Accountability

The Ministry of Digital Economy and Society established Thailand’s AI ethics guideline making Thailand one of the first Southeast Asian nations to formalise AI governance principles at this level. The framework was developed in collaboration with Mahidol University, with Microsoft serving as consultant.

The six principles in the framework are broad, but two are directly relevant to how Thai law firms should think about AI adoption: transparency and accountability. Transparency requires that AI use be disclosed when it affects others, which aligns directly with the Supreme Court’s disclosure requirement. Accountability requires that someone be responsible for AI output. In a professional services context, that someone is the professional whose name goes on the advice.

The framework does not restrict AI adoption. It establishes the conditions under which AI adoption is responsible. Meeting those conditions, disclosure, verification, and accountability, is more achievable than many firms assume.

The Practical Risk Partners Are Underestimating

The senior partners most resistant to AI adoption in Thai law firms typically frame their concern as a liability question: if the AI produces an incorrect legal citation and we rely on it, who is at fault?

The answer under both Thai professional ethics rules and the Supreme Court guidelines is the same as it has always been: the signing lawyer. AI does not change professional accountability. It changes the speed and volume at which drafts are produced. A junior associate who researches incorrectly and a language model that generates a case citation that does not exist create the same professional risk through the same failure mode: the output was not verified before it was relied upon.

The practical risk senior partners are underestimating is not the risk of AI adoption with proper review procedures. It is the competitive risk of not adopting AI while their counterparts do. A firm that has not integrated AI-assisted research and drafting into its workflow by the end of 2026 is operating at a material speed and cost disadvantage relative to firms that have. The verification step required by the Supreme Court guidelines is not an obstacle to adoption. It is a description of how professional practice should work regardless of the tools involved.

What the Verification Step Requires from Software

The Supreme Court’s requirement that attorneys verify AI-generated content consistently places a specific demand on the tools firms use. If an AI drafting tool produces a legal narrative that cites three cases and quotes a clause from a regulation, the attorney reviewing it needs to be able to check each citation without leaving the platform or conducting a separate search from scratch.

Opaque AI output, where the system produces text without indicating which sources were used or where specific claims originate, fails this requirement. The attorney cannot verify what they cannot trace. The standard the court expects is one where the AI produces output with attribution intact, so that verification is a reading and checking task rather than an independent research task performed from zero.

Audit trails matter for the same reason. A firm that can demonstrate, in an enforcement or malpractice context, that its AI-assisted output was reviewed and verified by a named professional on a specific date has evidence of its review process. A firm that cannot reconstruct that record has no evidence even if the review occurred.

FirmFlow and the Human-in-the-Loop Architecture

FirmFlow is built around the principle that AI handles the first operational layer and the professional owns everything that follows. The Document Analyser reads and extracts findings from uploaded contracts and legal documents. The Report Drafting module assembles a first version from the matter record. The Meeting Summariser captures and structures what was discussed. At each stage, the output is presented to the fee earner for review, correction, and approval before it goes anywhere.

Citations in AI-generated analysis trace back to the source document so that the reviewing lawyer can check each claim against the original text. The matter record maintains a log of what was generated, what was reviewed, and what was approved, providing the audit trail that professional accountability requires.

This is the architecture the Supreme Court guidelines describe: AI produces a draft, the professional verifies it, and the record shows that the verification occurred.

The firms that will be best positioned as AI governance in Thai professional services matures are not the ones that adopted AI most aggressively. They are the ones that adopted it with the right procedures in place from the beginning: disclosure where required, consistent verification of AI output, and a record that demonstrates both.