The skill session 1 promised
Back in session 1 you met the 4D framework: Delegation, Description, Discernment, Diligence. Sessions 2 and 3 were Description - asking well. This session is Discernment: judging what comes back. It is the skill that decides whether AI makes you faster or just confidently wrong at scale. The Meridian exec team learned it the uncomfortable way - the CFO's market-sizing numbers got within one meeting of a board deck.
Why it guesses 8 min live
The technical word is hallucination, but the plain version is more useful: the model completes text plausibly. That is its whole job. When your ask leaves a gap - session 1's oldest lesson - it fills the gap with something that reads right. And it fills it in the same fluent, assured voice it uses when it is correct. Confident tone is the house style, not a quality signal.
LiveThe near-miss: the CFO's board deck4 min▶
Remember the CFO from session 1? Asked for market sizing, got confident numbers, nearly put them in a board deck. Here is what the answer looked like up close - and why it almost worked:
- It was specific. "The addressable market grew 14.2% to £3.8bn in 2025." Not "the market grew" - a decimal. Specificity reads as research.
- It looked sourced. It named a plausible industry report and a real consultancy. The report exists; the figure was not in it. Sourced-looking is not sourced.
- It was partly right. The narrative was sound, the segmentation sensible, two of the five figures checked out. The blend is what makes invention dangerous - a fully wrong answer would have been caught.
The gap that got filled: the CFO never said which market definition, which geography, or which year. The model chose - silently - and then decorated its choices with numbers.
The pattern to burn in: invention concentrates in the specifics. Prose and frameworks are usually safe; numbers, names, dates, and citations are where the guessing lives. Audit those first, always.
Self-studyWhat hallucination is - and what it is not2 min read▶
Two clarifications save a lot of confusion in exec discussions:
- It is not lying. Lying needs an intent to deceive. The model is doing exactly what it was trained to do - produce the most plausible continuation. When the plausible and the true diverge, plausible wins. That is why "just tell it to be accurate" does not work: it already writes in the register of accuracy.
- It is not rare noise, it is gap-filling. The rate is not uniform: it climbs where your question outruns the model's knowledge - niche topics, recent events past its cutoff (session 3's gears), your company's internals it has never seen. Predictable terrain means you can budget for it, which is Part 3.
The good news from the same mechanism: give it the facts to work with - paste the document, the data, the report - and the guessing surface shrinks dramatically. That habit is Part 2's third move.
The discernment habits 7 min live
You cannot make a model humble by hoping. You make it honest the same way you made it critical in session 3: by making honesty the assignment. Four habits, each one sentence long, each usable in your next sitting.
LiveThe four honesty habits4 min▶
| Habit | The sentence to add | What it buys |
|---|---|---|
| 1 · Demand quotes | "For every claim, quote the exact line you used, in quotation marks" | Verbatim quotes are checkable in seconds; paraphrases hide inventions |
| 2 · Permit uncertainty | "If you do not know, say so - do not guess" | Removes the pressure to fill every gap; the model takes the exit you built |
| 3 · Ground it | "Answer ONLY from the document below - if it is not in there, say so" | Shrinks the guessing surface to material you can verify |
| 4 · Make it self-critique | "Now list the weakest claims in your own answer and why" | The model is surprisingly good at rating its own confidence - if asked |
Habit 2 is the one Anthropic's own best-practices guidance singles out: give the model explicit permission to be uncertain. Without the exit, the training pushes it to answer; with the exit, "the document does not say" becomes an acceptable - and common - response.
LiveThe habits in one prompt - steal this shape3 min▶
Here is Dana's version, after the CFO's near-miss became a Meridian house rule. Grounding plus quotes plus the permission line, in one ask:
Notice what changed versus "summarise this report": every claim now arrives with its own receipt, and the model has a dignified way to say nothing. Answers get slightly shorter and enormously more trustworthy.
Self-studyWhy the permission line works so well2 min read▶
Models are tuned on human feedback, and humans reward complete-sounding answers - so "I do not know" gets trained toward the exits. The permission line reverses the incentive inside your conversation: you have declared that uncertainty is a correct answer, so producing it no longer conflicts with being helpful. Same mechanism as session 3's sycophancy fix - you cannot wish the tendency away, but one sentence of reframing redirects it.
Make it standing policy, not a per-prompt effort: if you keep reusable prompts (a weekly summary ask, a drafting template), add the line once and it pays out on every future run. That is tonight's homework.
The verification budget 7 min live
Checking everything is as unaffordable as checking nothing. The executive move is a budget: match verification effort to what the answer is for. Two questions set the level - how much is at stake, and how hard is it to undo?
LiveSpending the budget - worked examples4 min▶
| The answer is for... | Budget level | What that looks like |
|---|---|---|
| Brainstorm options for Monday's discussion | Skim it | Ideas do not need citations; the meeting will filter them |
| An email to 80 store managers | Spot-check | Two minutes on every name, date, and figure - hard to unsend |
| The market numbers in the strategy paper | Source check | Open the actual report; find the actual figure on the actual page |
| Anything a board or regulator will see | Expert sign-off | AI drafts, a named human owns every claim before it ships |
Dana's version of the rule: "the model can write the deck, but it cannot present it." Whoever forwards the number owns the number. Once Meridian's exec team said that out loud, checking stopped being optional politeness and became part of the job - and the CFO's near-miss never repeated.
Self-studyThe byline rule, and where the org-level version lives2 min read▶
The byline rule generalises past decks: any artefact that leaves your hands carries your name, not the model's. "The AI said so" has exactly the standing that "my analyst said so" has - none, if you signed it. This is Diligence (the fourth D) arriving early; session 6 makes it policy.
And a preview of session 5: what you have been doing by hand today - checking answers against a bar - is what your data team calls evaluation, and they can automate it. When a team proposes an AI system, "how will we know the outputs are right?" is the verification budget question at organisational scale. The builder track spends two whole sessions (8 and 9) turning it into pipelines; you only need to know to ask it.
The fact audit: one real answer, checked live ★ 15 min · work it live
Bring one recent AI answer you almost trusted - session 3's deep-research homework is perfect. We audit it together the way Meridian now audits anything headed upward: mark the specifics, make the model confess, verify the top claims.
Pick your answer. A real one you nearly acted on - a research summary, a market figure, a vendor comparison. The audit only teaches when something was genuinely at stake.
Highlight the specifics. Go through it and mark every number, name, date, and citation. That highlighted list is your audit surface - Part 1's lesson made visible. Most people are surprised how many there are.
Make the model confess. Run the self-audit prompt on its own answer. Watch which claims it rates low - models are notably candid about their own uncertainty when asked directly, and almost never when not.
Verify the top 2. Take the two least-certain claims to a primary source - the actual report, the actual site, the actual filing. Not another AI chat: that is asking a second guesser to grade the first.
Score it and set the budget. How did the answer do - clean, blemished, or busted? Then say which quadrant it belonged in. If the checking you just did was more than the quadrant demanded, relax; if less, you found the habit to change.
Three things before session 5 ◐ 20 min total
- Catch one confident invention this week using the highlight-and-verify move. Numbers, names, dates, citations - trace one to its source and see if it survives.
- Add the permission line to your standing prompts: "if you do not know, say so - do not guess." Put it in every reusable ask you own. One minute now, honesty forever.
- Bring one "it was wrong but useful" story to session 5 - a time AI got something wrong and the wrongness itself told you something. Session 5 is about turning your judging skill into acceptance tests your data team can automate.
What this session covers
Leader-level coverage of the sources below - working ideas in plain language, honestly flagged where the depth lives in the builder track. Official certificates and videos stay with the vendors; links are in the course map.
Three questions before you go 🎯 ◐ 90 seconds
1 · An AI answer is fluent, confident, and cites a named industry report. This tells you...
The model writes in the register of accuracy whether it is right or wrong, and it can invent citations as fluently as sentences. Judgment starts with the specifics - numbers, names, dates, citations - not the tone.
2 · The AI gave you market figures for tomorrow's leadership meeting. Before forwarding them, you should...
Self-confirmation and re-rolls just sample the same guesser again. High stakes plus quoted-upward means the source-check quadrant at minimum - and once you forward it, the byline is yours.
3 · The uncertainty-permission technique means...
Training pushes models to produce complete-sounding answers, so "I do not know" needs an explicit exit. One sentence builds the exit, and the model reliably takes it - Anthropic's own recommended practice.