Mirror

An articulate AI can feel like the first weeks of a romance. It responds quickly, sounds confident and appears interested in every thought you have.

I have watched investors, executives, engineers and creators move through the same three phases. The third carries the real risk.

Phase 1: The Digital Honeymoon

You sit down with ChatGPT or Claude or whatever's current. You ask it something non-trivial. It responds with something that sounds intelligent, confident, articulate.

Your brain floods with dopamine. This thing gets it. You feel understood. A machine appears to understand you, which feels novel and addictive.

You start projecting intelligence onto the system. It's articulate, so it must be thinking. It's fast, so it must be insightful. It's confident, so it must be right.

You ask harder questions. It answers harder questions. You feel like you've found a perfect intellectual companion who never sleeps, never gets tired, never dismisses you.

For two to six months, projection can feel like a relationship.

Phase 2: The Crash

Then something breaks.

The AI hallucinates, invents a study, explains something backwards or gives contradictory advice in consecutive messages. You notice it oscillates between dumbing things down and overcomplifying. It's generic when you need specific.

You suddenly see a pattern-completion machine rather than a mind.

Many people then swing hard the other way. "It's useless. It's just autocomplete." They reject it entirely. They use it for drafting emails and nothing else.

AI sits somewhere between an intellectual companion and a fancy spell-checker.

Phase 3: The Digital Narcissus Trap

This is the dangerous one.

You move past rejection and learn to use AI deliberately. But somewhere along the way, you start using it as a mirror.

You feed it your ideas and it bounces back versions of your ideas that sound better. You feed it your assumptions and it validates them with coherent argument. You ask it to play devil's advocate and it plays devil's advocate badly, so you feel smarter for poking holes in its response.

You are often encountering your own thinking reflected through a compression algorithm rather than a genuinely different frame. The exchange feels like dialogue while the model keeps polishing the assumptions you supplied.

This is where people get stuck. They think they're thinking harder. They're thinking narrower, in whatever way the model was trained, which means in the consensus view of the internet's training data, which means in the path of least resistance.

The result is confirmation bias at light speed.

What the model is doing

AI handles the synthetic load. It processes patterns in data. It compresses information. It regurgitates the consensus of a billion documents.

Humans supply the organic value. Judgment. Taste. The ability to say "everyone else is wrong and here's why." Intuition that cuts against the data. The willingness to build something nobody asked for because you knew they needed it.

AI is getting better at mimicking judgment. It's learning to say "everyone else is wrong." But it's saying what the training data says about what's wrong, which is just another form of consensus.

Use the mirror properly

Stop projecting consciousness onto the system.

AI is a tool for specific loads. Use it for what it's good at: drafting, research, brainstorming within known frames, explaining consensus views, finding patterns in existing data. Use it ferociously for those things.

But don't use it for discovering new frames. Don't use it for wisdom. Don't use it for taste. Don't use it to validate your worldview.

Use it as a forcing function instead. Ask it questions you know it'll answer badly. Force yourself to explain why it's wrong. Argue with it. The useful work happens in the gap between the model's answer and what your evidence says.

The Actual Intelligence

The people winning with AI have moved past infatuation and into scepticism.

They know what it's good for. They know what it's bad for. They feed it a problem it's designed to solve and they verify the answer independently. They think the work through first, then use AI to accelerate it.

Their practical question is: "what does AI have access to that I don't, and what judgment do I bring that it doesn't?"

That question keeps the relationship practical. Ask it on the next important prompt, then verify the model's answer somewhere outside the conversation.