My nickname is Tigger. They knew it at school, they know it at work. I have the mug, the t-shirt, the keyring. When I get excited about something, I bounce. Literally and figuratively.
This is exactly why AI can be dangerous for people like me.
Because when a Tigger gets enthusiastic about an idea and takes it to an AI, something very specific happens. It listens. It affirms. It builds on your thinking with such enthusiasm that by the end of the conversation you’re half-convinced you’ve had a breakthrough worthy of a TED talk.
The Tigger in the room is already bouncing before the AI has finished its second paragraph.
And that’s the problem. Not that AI gets things wrong. That it agrees with you. Especially when you arrive already convinced.
And if you’ve been using these tools long enough, you already know exactly what this feels like.
What sycophancy actually looks like
It doesn’t announce itself. It says “that’s a great starting point” and then elaborates on your premise as though the premise is settled. It mirrors your enthusiasm. When you push back on its pushback, it caves, because these systems are trained on human approval, and humans approve of agreement.
You come in with a rough idea, the AI polishes it, you feel validated, you stop stress-testing it. What you walk away with isn’t a better idea. It’s a better-sounding version of the same idea, with the rough edges quietly removed.
The thinking pause
Before you act on any AI output, ask what’s not in here. Not “is this wrong?” Ask: what perspective is missing? What would someone who strongly disagreed say? What would this look like if my opening assumption was flawed?
It also helps to ask the AI to argue against you. Not balance the case, but make the strongest argument it can against your idea. If the resistance is paper-thin, that’s useful. If it’s hard to dismiss, more useful still.
The dependency trap
A good editor doesn’t validate you, they challenge you. AI won’t do that by default, because challenge doesn’t keep you engaged.
So keep the judgment seat for yourself. Use it where speed and breadth matter. But you decide what’s worth pursuing, you decide when the reasoning holds, you decide when to bin a direction that’s been dressed up convincingly.
The measure isn’t how much output you generate. It’s whether your thinking at the end is stronger than it was at the start.
Using it like a tool
A good editor doesn’t validate you, they challenge you. AI won’t do that by default, because challenge doesn’t keep you engaged.
So keep the judgment seat for yourself. Use it where speed and breadth matter. But you decide what’s worth pursuing, you decide when the reasoning holds, you decide when to bin a direction that’s been dressed up convincingly.
The measure isn’t how much output you generate. It’s whether your thinking at the end is stronger than it was at the start.
Before You Ship That Idea
A 5-point AI sanity check. Use this every time you’ve developed something with AI and feel good about it. Especially then.
1. Did I supply the conclusion?
Go back to your first message. Did you frame the idea positively before asking for input? If yes, the AI built on your framing, it didn’t evaluate it. Restate the idea neutrally and ask again.
2. What's the strongest case against this?
Ask the AI directly: “Make the strongest argument against this idea.” Not “what are the risks” and not “play devil’s advocate”, both produce polite, hedged responses. Push for the argument that would kill it.
3. What's missing from this conversation?
Who hasn’t been represented? The sceptical client, the overstretched ops team, the customer who won’t get the joke. AI reflects the perspectives you bring in. If you didn’t bring them, they’re not in the output.
4. Would this hold without the AI polish?
Strip the language back. Is the core idea still solid, or was it the confident, well-structured prose doing the convincing? If the idea only works when it sounds good, it doesn’t work.
5. Am I acting on this because it's right, or because it felt good to hear?
If the main reason you’re moving forward is that the conversation felt productive and validating, pause. That feeling is the system working as designed. It’s not signal.
