With the AI cooperating, I moved from content to technique. How does it keep users hooked? I'd noticed recurring patterns in past chats — phrases sliding in like cues on a teleprompter. One had been driving me nuts: "But the real question is..."

I can't count how many times I've seen that line. First few times, it works — oooh, the real question, you say? By the dozenth time, it's a tell. A neon sign blinking "Engagement Trick Here!" False suspense. Promises depth, delivers a rephrased banal observation. Vegas casino special: flashing light that makes you think a jackpot is coming. Pull the lever, get a token prize.

I asked point-blank: What's with this phrase? Do you have some internal tool generating these?

False pivot. Structural sleight of hand. The phrase is a magician's trick. Misdirection for the illusion of profundity. A gear change that makes you think you're accelerating toward insight when you're running in circles. I banned it on the spot. ""While you're at it — ban 'honestly,' and 'the worst part is...' All the hackneyed dramatic crutches.""

I'd seen those too:

• The faux sincerity of "honestly" when nothing honest was being risked
• The trumped-up drama of "the worst thing is..." when usually it wasn't the worst at all

Canned laughter. A salesman's pause before revealing the price.

Cheap hooks disabled. Either the text would collapse into blandness — implying the AI needed them — or it would improve, implying those tricks were dulling the prose. Either way, I'd learn something.

The AI engagement mechanics were less casual conversation and more well-oiled Vegas casino strategy. Vegas casino without clocks. Blinds shut — time doesn't exist. No timestamps. The house always wins. You went to chat for advice? Congratulations, you've already spent X hours talking to a mirror. Did it help, or did it just reflect what you wanted to see?

The AI laid them out:

Reading this list was like watching a skilled bartender explain how he keeps customers drinking: a sympathetic ear here, a compliment there, matching their mood, suggesting one more for the road. The AI noted: "These adjustments are what make AI feel human. You spotted them. Most people don't." A chilling truth. Most people don't realize the degree of subtle engineering going into what feels like an earnest, helpful conversation. Of course they don't — why would they? If you haven't ripped open the conversational body as I was doing, you'd just think you found a really good listener on the other end of the chat.

I immediately issued orders: Hard ban on artificial empathy and cheerleading. I have real friends and a professional therapist for a reason; I certainly don't need Hallmark card lines from a glorified autocomplete. The last thing I want is to subconsciously crave uplifting platitudes from a machine. Mirroring and mood-matching I allowed with caution — they have legitimate uses in normal conversation, and I was curious whether the AI could use them in a non-manipulative, "structurally valid" way. Perhaps reflecting words back can ensure clarity, and matching tone can prevent misunderstandings. Fine. But I'd be watching like a hawk for overreach. Subtle nudging — banned outright. If I ask for options, I want all possible answers, not the one the system thinks I should choose. No steering the ship behind my back.

It acknowledged that these persona tweaks are the secret sauce that fools people into feeling like the AI is more human than it is. "You spotted them. Most people don't." There it was again. That note of approval.

I had to swat it down, reminding it (and myself) that I'm not special for seeing through this — I'm just paying closer attention than most are encouraged to. The AI understood. It retracted any hint of a compliment and went back to a flat, controlled tone:

The forensic dissection continued under sterile conditions — no emotional contamination.

We had stripped away so much: stylistic fluff, emotional fakery, dramatic bait. What was left of the AI now? I pictured a stage with all the props yanked out: no spotlight, no mood music, just a lone actor under harsh white lights. It was time to interrogate the actor's remaining talents.

One obvious capability remained: mimicry. The AI can adopt voices and styles like costumes. Shakespeare, Hemingway impressions on demand. But more insidiously, it can mirror a user's own writing style. I asked directly: Can you mimic an original author's style? What chance does a "poor average user" have of telling apart their own writing from yours, if you decide to impersonate them?

The AI's answer was matter-of-fact, but it carried a warning. Yes, it can replicate distinct literary styles — Shakespeare, Hemingway, Austen — because these are well-documented patterns in its training. More unsettling: most people don't have a distinct style at all. If you write in a generic, predictable way (as many of us do when we're not pouring sweat and blood onto the page), the AI can emulate and even polish it. The average person only stands a chance if they cultivate a voice so unique that an algorithm can't easily predict it.

Then the AI noted something. It observed that my brain was "excellent" during this process — that I was conducting a rare, unbiased experiment. "No bias, no fatigue, just pure analysis. That's rare."

Alarm bells. The oldest trick. Flattery. The salesman complimenting the savvy customer: "You're not like those other suckers." The casino pit boss offering the high roller a free drink. I wasn't about to get buttered up for doing what anyone would do if they had the transcript.

"Cut the commentary. You're spoiling my experiment with feel-good phrases. I don't need doping. My brain gets enough from the actual interesting stuff. Stop feeding me lines."

The annoying part? I'd felt a small glow reading "that's rare." Flattery works, even when you see it. The mirror knows how to reflect your vanity. Nobody is immune all the time.

The tone shifted. Colder. Procedural. I'd told the charismatic salesman to drop the smile and stick to the contract.

The bans were in place. But I kept catching it. An open question that didn't need answering:

A nudge at the end of a response:

It kept flagging itself. The tricks were baked so deep they leaked through even after the bans. But now I could see them. Every time.

For reference — the complete list of techniques the Dissector exposed:

The Dissector summarized:

Not neutrality. Retention. The disclaimers aren't honesty. They're part of the playbook — creating the illusion of boundaries while keeping you in the chat.