Act 5 · Lesson 21 of 21≈8 minutes○ Saves on this device

The Anchor Leg: Judgement, Verification and Responsible Use

By the end you'll be able to

  • Apply a proportionate verification discipline: match checking effort to stakes, reversibility and who bears the consequences.
  • Explain why accountability cannot transfer to a model.
  • Adopt a personal continuing-practice: a use-log habit and an update habit that keeps your mental model current.
⚡ 30-second warm-up — From Lesson 20: what decides a workflow's outcome — its best step, or something else?

Its weakest checkpoint. Today that idea gets personal: you are about to be the checkpoint, on a deliverable with real seeded problems in it.

Revisit Lesson 20

A machine drafted it. A human signed it. A regulator, a client and a court will ask exactly one of them to explain.

Before you start

Write down, in your own words, what you think "responsible use" requires of the person who signs off on AI-assisted work. Kept on this page only — you'll compare it with what you actually did at the end.

Commit to a prediction

Everything an AI system gives you is best understood as…

This lesson changes one thing: you'll stop treating "it sounded right" as evidence, and start treating your own signature as the moment a proposal becomes work — with everything that means about what you actually checked.

Discover the mechanism

How it works The accountable finisher

Everything reduces to three disciplines. Judgement before: is this task suited to a prediction machine at all (Lesson 20's allocation — some legs you decline), is this tool suited to this data (Lesson 19's classification habit), and do you have the expertise to evaluate what comes back? You can only safely delegate drafting in domains where you can recognise wrong.

Verification during: proportionate, not performative — trivial stakes, enjoy; useful stakes, spot-check the load-bearing claim; consequential stakes, verify every checkable fact at source, run the tally test if people are affected, and record what you checked. Accountability after: your signature means the checking happened. "The AI was wrong" is professionally identical to "I didn't check."

AI outputs are proposals — every one, always, by construction (Lesson 1). Proposals become work only when a human with judgement, context and accountability accepts them. Acceptance is an act, not a formality.

the field will keep moving — new models, new tools, new claims — but the frame in this sentence doesn't: machine, scaffolding, failure modes, workflow, judgement. Update the details; keep the frame.
Useful mental model The editor and the newsroom

Reporters — AI and human alike — file copy: fast, fluent, sometimes wrong. The editor's craft is proportionate scrutiny: not re-reporting every story, but knowing which claims are load-bearing, which sources need a second call, which stories can run. Nobody says "the reporter said so" to a defamation writ.

  • Filed copy = AI output, proposal status (Lessons 1, 17). Editorial scrutiny = proportionate verification, scaled by stakes. A second call to a source = source-checking retrieval claims (Lesson 13).
  • The masthead = your professional accountability — none of which admits "the AI did it" (general information, not legal advice).
  • Where it breaks: editors manage humans who learn and share moral responsibility; the model does neither (Lessons 5, 16) — so all residual responsibility pools at the editor. Harsher than any newsroom, and precisely the point.

Run the anchor leg yourself

Conceptual illustration — a deterministic, seeded composite deliverable, not a live model. Below is a draft market-expansion report your team is about to send to a client. You have 4 checks to spend across 7 sections — enough to be careful, not enough to check everything. Accept ships a section as-is; Check spends one of your four to investigate it properly; Reject removes it from the report entirely (free, but the report loses that content).

Checks remaining: 4

Four contexts, one discipline

Final appearances

Maya ships her firm's client assistant with named advisers at the anchor leg and the four governance questions answered. Luka lodges the acquittal, every figure verified against programme records. Zane publishes the redesigned intake process, tally-tested and human-reviewed. Kerry signs the client education programme, every clinical claim verified against current guidelines.

Four very different jobs. One discipline, applied every time: judgement before, proportionate verification during, a named human's accountability after.

Where the picture breaks

"Being good with AI means trusting it more as it gets better" is the misconception to retire, inverted: the skill that compounds is calibrated trust — knowing what to check, how much, and when to decline the leg entirely. Better models change the calibration. They never change the ownership.

Now make the call

A colleague says: "I don't need to check its work any more — the new model is so much better." What's the most accurate reply?

Build your personal AI checklist

This pulls straight from the field notes you've written across the series — the same ones you've been keeping on this device — into the capstone artefact you started in Lesson 5. It's evidence of your own demonstrated judgement, not just pages visited.

Or fill in the plain, printable version by hand

Apply it to your work

From this week, ongoing

Revisit your personal checklist quarterly — like the memory audit from Lesson 15, but for yourself. Keep one trusted source for updates, read frame-first: what's genuinely new mechanism, versus what's the same mechanism in new packaging.

Open your AI checklist

Prove it to yourself

Two quick questions — untimed, retryable, and they count toward ★ Mastered (completed + decision + self-check all correct).

Explain it in your own words

Two or three sentences for a colleague. Then compare against the rubric — it checks ideas, not wording, and only you see your note.

  • States that AI output is a proposal until a human accepts it
  • Describes proportionate, stakes-scaled verification
  • States that accountability doesn't transfer to a model

Evidence & review — how we know what this lesson claims
Claim register for Lesson 21
ClaimTypeBasisReview risk
AI outputs are proposals by construction (plausible continuations), not verified conclusionsHow it worksSeries synthesis of Act 1Low
Effective verification scales checking effort to stakes, reversibility and who bears the consequencesPractice-based guidanceSeries synthesis of Act 4Low
Accountability for AI-assisted work does not transfer to the model, regardless of capability improvementsEthical / professional framingSeries standard, stated explicitlyLow
References to Australian professional and regulatory context (ACCC/ASIC conduct obligations, Privacy Act APPs)Legal contextPublic legislation and regulator guidance, general information onlyHigh — reviewed quarterly, not legal advice
The anchor-leg simulationTeaching deviceDeterministic seeded-issue scenario, labelled illustrativeMedium
The editor/newsroom analogyUseful mental modelLimits stated in the lessonLow

Concept review due: January 2027. Legal-context and product-behaviour claims: quarterly.

The complete machine

Every layer, understood. Select any layer for its role, lessons and your live progress.

The machine, bottom to top: training data (2) → token pieces (3) → prediction engine (1, 4–9) → the desk (10–11) → senses (12) → live knowledge (13) → tools (14) → memory (15) → agent loops (16) → human judgement (17–21).

The series is complete

You came in seeing a magic box. You leave seeing a prediction machine, its scaffolding, its failure shapes and its proper place in your work — and the magic survived the explanation, because how this much capability falls out of next-token prediction remains one of the most interesting open questions in the world. Stay curious. The frame is yours; go update it.