The AI Learning Hub Journal

Supervision, Candour, and the Tribunal

Supervision and candour over AI-assisted workeducational orientation only — not legal advice1A draft is producedby a junior, by a tool, or by a junior using a tool —from here on the origin changes nothing about the duty2A supervising lawyer reviews the substancenot whether it reads well: whether it is right, whether it issupported, and whether it is the advice this matter needs3A named person accepts responsibilitythe signature is a statement about the work, not a recordof who assembled it or what assisted them4It is filed or sentwith candour to the tribunal, including any disclosure aboutthe use of such tools that the applicable rules requireREVIEWING SUBSTANCE, NOT FORMIs every proposition actually supported?Has each authority been opened and read?Is anything material missing or overstated?Is this the advice we would give unaided?Would it survive being questioned closely?WHAT CANNOT BE DELEGATEDA tool holds no duty to the clientA tool holds no duty to the tribunalA tool cannot be sanctioned or struck offA tool cannot explain its own reasoning on oathSo the responsibility never leaves the personA tool assisted; a person is answerableWhere the rules or the tribunal require it to be said that such tools were used, saying so is part of the duty, not an admissionUnreviewed work does not become reviewed because something confident produced it
Every link in the chain ends at a person — the tool changes how the draft is made and nothing at all about who answers for it

Supervising a Tool Is Not Like Supervising a Person

Supervision duties over subordinates are long-established and translate awkwardly to AI. A junior has a stable error pattern, learns from correction, tells you when they are unsure, and can explain their reasoning. A model has none of these properties: its errors are unpatterned, correction does not persist across sessions, it expresses no genuine uncertainty, and its stated reasoning is generated text rather than an account of what actually happened internally. The consequence is that supervision cannot rely on trust built over time. It has to be systematic — sampling output regardless of past performance, and treating a long run of good results as a reason for continued vigilance rather than reduced checking.

  • Model errors are unpatterned and do not diminish with correction across sessions
  • Stated reasoning is generated text, not a faithful account of the internal process
  • Supervision cannot be trust-based because trust cannot be earned by something that does not learn from you
  • A long run of good output is a reason for sustained sampling, not reduced checking

Candour to the Tribunal

Duties of candour and of not misleading the court are near-universal in some form. AI engages them in a way that is easy to underestimate: submitting a document with citations you have not verified risks misleading the court whether or not you knew the citations were false, and in several reported instances the failure characterised by courts has been the absence of verification rather than any intention to deceive. Candour also engages after the fact. Once a doubt is raised about a citation, the obligation is to check immediately and correct on the record. Delaying while you consider your position, or defending the citation without verifying it, has consistently made outcomes worse in the reported pattern.

  • Filing unverified citations can mislead the court regardless of intent
  • Courts have characterised the failure as absence of verification rather than deliberate deception
  • The obligation is continuing — a doubt raised triggers immediate checking and correction
  • Delay and defence-before-verification have consistently aggravated outcomes

Disclosure Requirements Are a Per-Forum Question

Whether you must disclose AI use in a filing depends entirely on the forum. Some courts and individual judges require disclosure of any generative AI use, some require certification that a human verified all citations, some restrict use in specified filing categories, and many say nothing at all. These requirements vary between jurisdictions, between courts in the same jurisdiction, and between judges in the same court, and they have been changing. There is no universal rule to memorise. What there is, is a checklist item: before filing, check the current standing orders and practice directions for this court and this judge. Firms should keep a register and assign someone to maintain it, because individual practitioners will not track it reliably.

  • Requirements vary by jurisdiction, by court, and by individual judge — and keep changing
  • They range from disclosure of use, to human-verification certification, to category restrictions
  • No universal rule exists; the control is a pre-filing check of the specific forum
  • Keep a maintained firm register — individuals will not track this reliably on their own

Prefer slides, quizzes, and saved progress? Read this lesson in the library — free, no sign-up.