Regulation, AI accountability, and the human layer.

Legal practice is one of the most regulated forms of professional work in any common-law jurisdiction. Rules on conduct, conflicts, confidentiality, client funds, advertising, supervision, and admission all exist because the consequences of bad legal work fall on people who are not in a position to evaluate the work themselves. A client cannot reliably tell whether a contract clause protects them, whether a statute was correctly applied, or whether a case citation actually exists. The apparatus exists because the asymmetry exists.

In Sri Lanka, the shape of that apparatus is familiar. The Supreme Court controls admission to the roll of Attorneys-at-Law. The Bar Association of Sri Lanka maintains the rules of professional conduct. The Judicial Service Commission and the courts supervise the conduct of practitioners. Personal responsibility runs through every layer: a practitioner is not free to outsource it, and there is no version of the practice in which a software tool absorbs the duty. The same structural logic - personal accountability of the admitted lawyer - runs through the English, Australian, Indian, and Singaporean regimes that Sri Lankan practice is in conversation with.

AI-assisted legal work sits outside that apparatus. There is no admission process for a model. There is no professional rule prohibiting a model from acting outside its competence. There is no court that supervises the model's day-to-day output. The model has no duty of confidentiality. It has no duty of candour. It owes no fiduciary obligation to the client. It has no duty at all.

This is not a problem of capability. Capability is improving steadily. It is a problem of accountability.

When a model produces a confident wrong answer, the question of who is responsible has only one defensible answer: the human who used it. That answer holds regardless of how impressive the output looked, and it holds well before any regulator has formalised the rule, because the duty was already on the practitioner under the existing apparatus.

The early case law is small but consistent. Each reported decision treats AI use as a delegated activity that does not relieve the lawyer of the underlying professional duty.

Professional-conduct opinions have begun to formalise the same principle in writing. ABA Formal Opinion 512 (July 2024) addresses generative AI specifically and concludes that the existing duties of competence, confidentiality, communication, supervision, and reasonable fees all apply to AI-assisted work. The model's involvement does not reduce any of them. Where AI assistance is material to the work, lawyers may need to disclose its use to the client. Where the AI is non-lawyer assistance, the lawyer's supervisory duties extend to it.

Several US state opinions - Florida, California, New York, New Jersey - have reached compatible conclusions, often with more concrete operational guidance. The consistent thread is that AI is treated as a category of delegated work, the kind that has always been governed by the duty of supervision (Model Rule 5.3 in the US; analogous provisions in most Commonwealth codes). The result is structural: the lawyer remains the responsible person.

The legislative picture is moving the same way, more slowly and more bluntly. Three frameworks are doing the early work.

There is a separate accountability surface around confidentiality. Sending privileged matter to a third-party AI service can risk waiving privilege, depending on the jurisdiction and the vendor's terms. Several state bar opinions in the US have flagged this directly: lawyers must understand where their inputs go, who can see them, whether prompts are retained for training, and whether the vendor's terms of service preserve the duty of confidentiality. Most public-facing AI products do not preserve it without specific enterprise terms.

The choice of tool is itself a discharge of the duty of competence under the rules of conduct, not an operational footnote.

This is one of the reasons the document-execution infrastructure I deploy is self-hosted. It is also why the legal-AI reference corpus excludes anything that is live-client privileged material until it has been properly anonymised and cleared for reuse.

The most interesting next move for AI-assisted legal work is not technical. It is procurement. Standard AI-vendor terms of service routinely disclaim accuracy, fitness for purpose, and any liability for downstream loss. Most professional indemnity policies for lawyers do not exclude AI-related errors - but they do not absorb the vendor's disclaimer either. The exposure is real and it sits with the practising firm.

Mature legal-tech contracts will, in time, have to allocate that risk explicitly: warranties on training data provenance, accuracy testing protocols, evidence retention requirements for human review, indemnity carve-outs for hallucinated authorities, and discovery treatment of AI-assisted drafts. Where the regulator is slow, the contracts and the insurance policies will do the early work. That is also where a reviewer-gated, audit-recorded workflow becomes practically defensible: it is easier to insure, easier to disclose, and easier to defend than an unstructured prompt-and-paste process.

Concretely, in the systems described elsewhere on this site, the human layer is not a courtesy. It is a control. The drafting workflow ends at a reviewer's signature. The operating record exists to put the relevant facts in front of the person making the judgment, in the order they need to see them. The execution workflow refuses to treat a document as signed until the version is locked, the attachments are complete, the signer authority is verified, and the audit trail is being kept. The reference corpus captures reviewer edits rather than confident outputs, because the edits are what carry judgment.

A model with no duty is a useful tool. It is not a substitute for the person who carries the duty.

These are not separate design choices made for separate reasons. They are the same choice expressed in different places. The systems are built around that distinction because the distinction is structural - and because every framework moving through the major jurisdictions in 2024 and 2025 has reaffirmed it.

A note on what this is, and isn't. I am reading for Attorney-at-Law in Sri Lanka, expected admission 2027. Where regulation is settled, follow it in the relevant jurisdiction. Where it is unsettled, the conservative posture is to design as if the human-duty regime were already binding - because for every reported case so far, it has been.