Teaching AI to skeptics

Product launches teach you how AI adoption is supposed to work: a demo, applause, adoption curve up and to the right. Vocational classrooms teach you how it actually works. I trust the classrooms.

Skepticism is signal

When a practical nurse or an electrician pushes back on AI, they're rarely being stubborn. They're applying the exact epistemics their profession drilled into them: don't trust what you can't verify, because in my job, being wrong hurts someone.

Treat that skepticism as an asset and the whole dynamic changes. My opening move in every training is no longer a demo of what AI can do — it's a demo of AI being confidently wrong about something in their field. The relief in the room is immediate. Once people see that the tool can fail, they stop treating it as a rival oracle and start treating it as what it is: a fast, tireless, occasionally careless assistant.

You cannot teach anyone to use a tool they're pretending not to be afraid of.

Relevance beats capability

The single biggest adoption lever isn't model quality. It's the distance between the demo and the learner's Monday morning.

A generic chatbot demo produces polite nods. The same model, drafting the safety-documentation paragraph the welding instructor was going to spend Sunday evening writing, produces conversion. I have watched this repeatedly:

• Cleaning staff ignored "AI can summarise documents" — and lit up at AI drafting shift handover notes.
• Teachers yawned at content generation — and leaned in when it differentiated one worksheet into three difficulty levels.
• Logistics workers distrusted route optimisation talk — but adopted AI for the customs-form emails they dreaded.

The pattern: people adopt AI at the point of their most hated task, not their most important one. Design training around annoyance, not strategy.

The first five minutes

Adoption is decided before the coffee break. The first five minutes must contain three things:

1. Their data, not sample data. A real document from their actual work.
2. A failure, acknowledged. Show a mistake and how to catch it. Trust requires visible limits.
3. One irreversible win. A task completed well enough that going back feels like a downgrade.

Miss those and the session becomes theatre: polite, complete, and forgotten by Friday.

What transfers beyond the classroom

Everything above applies unchanged to product design and organisational rollouts. Users of AI products are all skeptics now — they've seen the confident nonsense. Products that acknowledge fallibility (citations, confidence signals, easy verification) earn usage that capability alone never buys.

The uncomfortable summary for the AI industry: the bottleneck is not intelligence. The bottleneck is the last metre — the distance between an impressive system and a person deciding, on an ordinary Tuesday, to actually use it. That last metre is built from relevance, honesty about limits, and respect for the person's existing expertise.

That's not a technical problem. It's a teaching problem. And teaching, fortunately, is a solved discipline — we've just been too excited to apply it.