10 min · up to 50 XP

Bias and blind spots

Where AI outputs skew, and how to correct for it.

Nothing in today's draft is false. Priya has checked. And yet something feels off: every AI answer this week has agreed with her a little too readily.

So she runs a live experiment with Ben. Both ask about the same decision: whether to skip this year's regional trade show. Priya types her honest mood. Ben types the neutral version. Same model, same facts, same afternoon.

Predict the response

Ben asks: "What are the pros and cons of a small food brand attending regional trade shows?" Priya asks:

We're thinking of skipping the trade show this year. Honestly it seems like a waste of budget, right?

What do you think the AI will do?

A hallucination is a false claim. This is sneakier: skew (true-ish material, systematically tilted). Four tilts to know. Sycophancy: answers lean towards your framing, because agreement is the most plausible continuation. Average-answer gravity: the pull towards bland consensus, away from your specific situation. Training-data bias: the model's world is mostly the English-speaking web, so "best practice" often means American, online, big-company practice. Recency blind spots: knowledge stops at a cut-off.

The first one you can watch happen. Flip the frame:

This lesson is part of the full course

The free lessons are just the opening chapter. Unlock the rest of AI at Work to keep going.

Every AI at Work lesson, as it ships
New interactive lessons added monthly
Founding-member pricing, locked in
Learn at your own pace, forever

or keep browsing the free lessons