Anti-patterns & failure modes

The concept

Knowing when not to trust the output is as valuable as knowing how to prompt. Models fail in specific, predictable ways, and the danger is that a wrong answer is delivered with exactly the same confidence as a right one — there is no tone-of-voice tell. Your defence is not suspicion of everything (that destroys the value); it's knowing which categories of output are load-bearing and verifying those specifically.

The failure modes, and the countermeasure for each

1. Hallucination — invented facts, numbers, citations, laws, quotes. The model produces the most plausible-sounding answer, and plausible-sounding is not the same as true. Highest-risk categories: statistics, prices, dates, legal provisions, named sources, URLs, industry benchmarks, anything about very recent events, and anything about your business it wasn't told. Countermeasure: add to any factual prompt — If you are not certain of a fact, mark it [VERIFY] rather than stating it. Do not produce a citation, statistic, or URL you cannot be confident of. Then treat every [VERIFY] as a task. And never let a number you didn't supply reach a client.

2. Sycophancy — agreeing with you. Push back on a correct answer and the model will often fold. Imply the direction you want and it will find reasons to support it. This makes it a terrible sounding board unless you force adversarial framing. Countermeasure: ask for the case against. "Argue the opposite position as strongly as you can." "You are the sceptical CMO — tear this apart." Never phrase a strategic question in a way that reveals your preference.

3. Confident mediocrity — the plausible-average answer. Ask for "a campaign idea" and you get the idea that the average agency would have. It's not wrong; it's just the centre of the distribution. Countermeasure: constrain hard (Module 1), demand a spread including a deliberately risky option (Module 6), and use your own best work as few-shot examples (Module 5).

4. Instruction drift over long chats. By message 30, early constraints have faded. Word limits slip, forbidden words creep back, format degrades. Countermeasure: restate critical constraints in the message that matters, or start fresh with the good draft (Module 7).

5. Fabricated completion — "done" when it isn't. Agents especially will report success on an action they didn't verify. Countermeasure: build verification into the instruction. "After writing the file, confirm it exists and report its size." "Check the live URL returns 200 before reporting the deploy succeeded." It should work is not a status.

6. False precision in translation and legal wording. Arabic legal and contractual phrasing has conventions that a fluent-sounding output can miss entirely. Countermeasure: AI drafts, a human with the relevant licence signs. Always.

When NOT to use AI

Be honest about the boundary — this is the mark of someone who actually understands the tool:

• Anything legally binding without human review. Contracts, statutory notices, regulatory filings. Draft yes, sign never.
• Final numbers. Pricing, budgets, financial commitments. Use it to structure the model, not to produce the figure.
• Genuinely sensitive relationship moments. Firing someone, apologising for a real failure, a condolence message, a hard conversation with a long-standing client. People can feel synthetic warmth, and being caught using AI here costs more than the time it saved.
• Anything requiring current facts you can't verify. Competitor pricing, who's just moved agency, what a regulation says today.
• The strategic core of your own thinking. Use it to pressure-test and to explore, not to outsource the judgement that is literally what your clients pay you for.
• Confidential material where the exposure isn't worth the gain. If redacting it costs more than the task saves, do the task yourself.

Before → After

What's the average CPM for Instagram advertising in Egypt in 2027 and
what percentage of Egyptian consumers watch video ads to completion?

I need media benchmarks for an Egyptian client deck.

Rules:
- State only figures you are confident about, and for each give the
  source and the year it's from.
- If you're not confident, write [VERIFY] and tell me exactly what to
  search for and which source would be authoritative.
- Do not estimate, interpolate, or produce a "typical range" unless
  you label it clearly as your estimate with the reasoning shown.
- If you have no reliable data at all, say "no reliable data" — that
  is an acceptable and useful answer.

Questions: [list them]

Then: for anything marked [VERIFY], give me a short research plan —
who publishes this, and what I should ask our media buyer for.