Steering & Calibration
Recognizing when AI is drifting off course and correcting it — without losing the thread of what you were trying to do.
What this skill is
Steering and calibration is the real-time skill of guiding AI toward what you actually want. It's the ability to read an AI output, notice when it's drifted, hallucinated, oversimplified, or missed the point — and then issue a correction that actually fixes the problem rather than starting over.
This sounds mundane but it's genuinely difficult. AI outputs that are wrong or off-target often *feel* right. They're coherent, well-structured, confident in tone. The skill of recognizing the gap between "this sounds good" and "this is what I needed" — and then bridging it efficiently — is something most people, kids and adults, are still developing.
Why it matters in an AI world
As AI gets better, its drift becomes harder to detect. Early versions of AI gave obviously wrong answers. Current versions give subtly wrong answers — answers that are directionally correct but miss a key nuance, that are accurate in general but wrong for your specific context, that are well-written but arguing a slightly different point than the one you intended.
A child who can't detect and correct this drift will produce AI-assisted work that looks polished but is subtly off. A child who can steer effectively is producing work where they remain the intelligent director — using AI as a capable instrument rather than being carried along by its defaults. This is one of the clearest and most consequential differences between AI-directed and AI-directing.
The calibration part is equally important. Over-correcting — becoming so skeptical of AI output that you second-guess everything — is as unproductive as under-correcting. The skill is knowing when AI is right enough to trust, when it needs a nudge, and when it needs to be discarded and restarted. That calibration comes from experience and from developing a clear enough sense of the target to evaluate how close you are.
What it looks like in your child
- When AI gives an off-target response, they can articulate specifically what's wrong (not just "that's not right") before issuing a correction
- They vary their correction style based on the type of error: they know a factual correction needs different language than a scope correction or a tone correction
- They maintain awareness of the original goal throughout a long AI conversation and can tell when the conversation has drifted away from it
- They don't restart from scratch every time AI misses — they've developed a repertoire of correction approaches
Challenge: Try this this week
The Drift Detection Game. Take an AI conversation your child already had (or start a new one together). Ask AI a question, then deliberately steer it slightly off-target. Can your child notice when AI has drifted from the original intent? Take turns: one person steers the AI subtly wrong, the other has to (a) name exactly what drifted and (b) issue a single correction that brings it back. No starting over allowed.
What to watch for
- Accepting drift silently: They notice the output isn't quite right but use it anyway, either out of laziness or because they can't formulate a correction
- Blaming themselves for AI failure: When AI misses, they assume they asked wrong — and keep apologizing in their prompts — rather than diagnosing what the AI did wrong
- Starting over instead of correcting: Every time AI misses, they start a new conversation rather than steering the existing one. This is inefficient and means they're not developing the skill of live correction.
Games that develop this skill
Codenames Drifting Spymaster — The spymaster gives clues, but halfway through each round, the spymaster's clue must be slightly modified (not retracted, just calibrated). Players must identify which cards the modified clue still applies to and which it no longer covers. Builds the skill of tracking drift in real time.
Monopoly Oracle Landlord — One player acts as an "oracle" who gives property-valuing guidance that's usually right but occasionally wrong. Other players must decide when to follow the oracle and when to override it — and explain why when they deviate. After the game, review: when was the oracle right? When were the overrides right?
Back-Seat Driver — In any video game, one player plays while the other gives steering instructions in real time. The player must judge which instructions to follow, which to modify, and which to ignore — and narrate their reasoning out loud. Switch roles every 10 minutes.
See all nine skills
Each skill builds something distinct. Browse the full set to see where your child might have gaps.
All Skill Builders →