COURSE 23L1100% FREE
Verified 2026-08-14

Demo skepticism

A habit of questioning polished AI demos: selection effects, hidden humans, cherry-picked prompts, unreproducible conditions, and evals that do not measure the claimed construct. Ties to research literacy (papers) and measurement literac...

What it is

A habit of questioning polished AI demos: selection effects, hidden humans, cherry-picked prompts, unreproducible conditions, and evals that do not measure the claimed construct. Ties to research literacy (papers) and measurement literacy (benchmarks).

<!-- IMAGE: stage demo vs lab notebook side by side -->

MEDIUM PRIORITYINFOGRAPHIC
◷ IN PRODUCTION

Visual Spec & Architecture Diagram

Poster of demo red flags as icons: cherry-picked prompts, no failure cases, hidden human-in-loop, unreproducible seed, 'works on stage only'. Green column: ask for metrics, limits, data provenance.

Educational Focus: Builds healthy skepticism without cynicism.

Why it matters

Demos raise money and expectations. Literacy protects wallets and roadmaps. STS 14’s syllabus includes weeks on evaluations, open vs closed models, and red teaming—governance literacy includes demo skepticism.

How it works (plain)

Ask:

  1. What is the task definition?
  2. What is the fail rate on held-out cases?
  3. Can outsiders reproduce steps?
  4. What is supervised / scripted?
  5. What is the cost at scale?
  6. What is out of distribution?

NeurIPS checklist themes (statistical significance, compute, broader impacts) are also good demo interrogation prompts.

Everyday example

A cooking show plate vs your Tuesday night kitchen—both real, different reliability.

Try it

Watch one vendor demo. Write five questions you’d need answered before buying. Map each to a checklist idea (repro, eval, impacts, data).

Myths

⚠️ Myth: Live demos can’t cheat.
✓ Reality: Constraints and prep still shape outcomes.
⚠️ Myth: Skepticism means cynicism.
✓ Reality: It means evidence-seeking.
⚠️ Myth: A leaderboard score is a demo of your use case.
✓ Reality: Construct validity may not transfer (AIMS textbook premise).

Sources