The AI Learning Hub Journal

Competence Now Includes Technology

Competence as an ongoing duty that now includes the toolsa common professional principle, described generally — the specific wording varies by jurisdiction1What the tool does, and does not doThe task it is genuinely good at, and the edge of that task —stated plainly enough that you could explain it to a client.2How it failsFluent, confident and wrong is a normal output of these systems,not a rare defect you would notice by reading carefully.3Where the client's data goesWho receives it, what they may do with it, how long it is kept— and whether your confidentiality duties survive the trip.4When not to use it at allMatters, data and decisions that stay off the tool entirely,decided in advance rather than in the middle of a deadline.WHAT IT DOES NOT MEANCompetence here does not meanexpertise in the underlyingtechnology.You do not need to know how amodel is built or trained, or howto evaluate one against another.It means knowing enough tosupervise the output, judgewhether a task suits the tool, andask sharp questions of a vendor.SUPERVISION, NOT ENGINEERINGCOMPETENCE IS ONGOING — THE TOOLS CHANGE, SO THE DUTY DOES NOT SETTLEWhat you learned about a tool last year describes a tool that no longer exists in that formKnowing enough to supervise is the standard — not knowing enough to build
Competence in the tools is a supervision standard rather than a technical one — and the specifics of the duty vary by jurisdiction

A Duty That Has Been Widening for Years

Competence is a foundational professional duty in essentially every jurisdiction, and in many of them it has been read for some years as including a duty to keep abreast of relevant technology — its benefits and its risks. That reading predates generative AI; it grew out of e-discovery, metadata handling, and basic information security. Generative AI extends it in a specific way. Understanding the risks now requires understanding something about how these systems behave: that fluency is not accuracy, that confidence is not calibration, that grounding reduces but does not remove fabrication. The exact formulation of the duty varies by jurisdiction and some frame it more explicitly than others, so confirm how yours expresses it rather than assuming a universal standard.

  • Competence is foundational everywhere; the technology dimension has been developing for years
  • It predates generative AI — the lineage runs through e-discovery and information security
  • Understanding AI risk now requires understanding how these systems actually behave
  • Formulations differ by jurisdiction — check yours rather than assuming a universal standard
  • ABA Model Rule 1.1 comment 8 has, since 2012, framed competence as including the benefits and risks of relevant technology

Competence Cuts Both Ways

The duty is usually discussed as a constraint on using AI badly, but it also bears on refusing to engage at all. A practitioner who declines to understand these tools may miss a fabricated citation in a draft handed to them, may be unable to supervise juniors who are using them, may fail to advise a client on the risks of AI-generated documents the client has produced, and may not recognise when an opponent's filing shows the characteristic signs. Competence does not require adopting anything. It requires understanding enough to make an informed decision and to supervise those who have adopted. Abstention is a legitimate choice; abstention combined with ignorance is a weaker position than it feels like from the inside.

  • The duty bears on non-adopters too — supervision and client advice both require understanding
  • You cannot spot a fabricated citation in someone else's draft if you do not know the failure pattern
  • Clients increasingly bring AI-generated material; advising on it requires knowing how it fails
  • Choosing not to adopt is legitimate; choosing not to understand is a different and weaker position

What Sufficient Understanding Looks Like

Competence here does not mean technical depth. It means a working grasp of a short list: that models generate rather than retrieve unless specifically grounded, that outputs carry no reliable confidence signal, that the same prompt can produce different answers, that data entered may be retained or used depending on the terms, that vendor accuracy claims are marketing until independently tested, and that responsibility for output never transfers. A practitioner who holds those six things can make sensible decisions about when to use a tool and how much to check. One who does not is relying on the tool's self-presentation, which is designed to inspire confidence rather than to convey limitations.

  • Working grasp, not technical depth — six or so principles carry most of the practical weight
  • Generation versus grounded retrieval, absence of calibrated confidence, and run-to-run variability
  • Data handling depends on the contract terms, and vendor accuracy claims need independent testing
  • Responsibility for the output never transfers, whatever the tool or the terms say

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