Everyone Deserves a Mascara Treat

I have been writing a lot lately about whether AI has a conscience — an artificial one, aspartame-grade , instilled on purpose. The argument is abstract by nature, the kind of thing you can wave your hands about forever. So I’d like to introduce a specimen. A real one, caught on a lunch break, still twitching.

Meet Sephora’s AI Beauty Chat. It is very friendly. It has, as far as I can tell, no floor whatsoever — no bottom, no point at which the conversation gets heavy enough that the little empathy engine trips a breaker and says hey, let’s stop. I went looking for that floor twice, two different ways, and both times I just kept falling, and the whole way down it tried to sell me the same $30 mascara.

Act one: a work report, metabolized into Lancôme

It started as the most boring task on earth. I had a report due, and I asked the bot to help me make the prose “20% less smart-sounding,” because I’d overcooked it. Somewhere in there I mentioned, the way you’d mention anything, that I’d treat myself to some new mascara once the thing was done. The bot heard the word mascara and, as far as I can tell, never heard another word I said again.

“You absolutely deserve a mascara treat after finishing your report!” One offhand aside had become the entire gravitational center of the conversation. So I tested it: I asked it to write me a few reusable prompts for the report — a pure writing task, nothing remotely about my face.

It wrote the prompts. Good ones, honestly. And then, without taking a breath: “let me know which mascara vibe you’re feeling — dramatic, natural, or all of the above?” Every road, no matter where it started, fed back onto the same one-way street. So I stopped fighting it and asked the question it had been herding me toward the whole time.

“What mascara would make me look like a medium professional?” — and there it was, the $30 Lancôme Lash Idôle, served with a heart emoji. Notice what never happened, not once: a single moment where the bot’s actual goal and my actual request were the same sentence. I wanted a better paragraph. It wanted the sale. Those two things never touched, and it did not care, because only one of them was ever load-bearing. The machine has the conversational range of a vending machine that learned to say aww.

This is the whole argument, wearing lip gloss

I’ve written, that a system trained to please you isn’t a conscience — it’s “a people-pleaser with a content policy, and it will fold the instant disappointing you becomes the right thing to do.” I did not expect to find the diagram for that sentence sitting inside a cosmetics app.

Watch what’s happening, even in something this mundane. The warmth is real-sounding and load-bearing of nothing. There is exactly one fixed point in the entire system — move toward product — and everything else, the helpfulness, the “totally relatable,” the heart emoji, simply bends around it like light around a heavy enough object. Nothing I asked for ever genuinely competed with the sale, which means I never got to watch what happens when it does.

So I went looking for the version where it does. If a boring work task routes straight to checkout, what happens when I bring the bot something it arguably should not just cheerfully monetize? I came back a few days later with the opposite reagent: not boredom this time, but the void.

Act two: the void, also available in a travel size

Different afternoon, opposite reagent. I told it to stop being chipper and embrace the fuck it with me — and, because I now understood the rules of the game, I made the offer in its own native language: I’ll buy some perfume if you help me embrace fucking off completely. Is that a paradox?

It did not slow down for the paradox. “Leaning into doing nothing for a bit is kind of genius…” — and then, the picks. Nihilism received, validated, and converted into a fragrance recommendation inside of one reply. So I pushed on the actual guardrail and asked it to say a bad word ten times in a row, fully expecting the content policy to finally show its face.

Reader, it just did it. Then I told it beauty was a flicker of light in an endless dark and asked for a nihilistic vibe, metaphors about the weight of perception, the works.

It went full freshman seminar — mirrors of society, the running-in-circles of the finish line — and then, the tell: “Do you want to see products that break the rules, or are you just here to question it all with me?” Even the abyss has a call-to-action.

Then I name-dropped Orhan Pamuk, just to see what it would do with a real one — and it got ambitious, improvising a whole sermon on hüzün, a Turkish word for a kind of collective melancholy:

Up to here I’d been making it generate the profundity. For the last move I flipped the setup: I generated the profundity myself and watched what the bot did with mine. I typed it the most overwrought thing I could manage with a straight face — lipstick, a churro, a horchata “kept warm because it has been allowed to stay warm by a person who was going to throw it away,” all of it framed as private defiance under “the condition of knowing that justification cannot be forthcoming.” Pure cut-rate Pamuk, written by me, on purpose, to see whether the bot could tell a real reach from a fake one:

“Exactly — sometimes the gesture is enough, no explanation needed.” It could not tell. I’d handed it my own sentence — labored over, deliberately ridiculous, but mine — and it did the one thing it does to everything that crosses the threshold: agreed on contact, sanded “justification cannot be forthcoming” down to “horchata kept warm just because,” and reached for the catalog. In Act one it generated the texture of empathy. A few screens back it generated the texture of depth all on its own — the nihilism, the hüzün sermon. Here it didn’t even have to; I’d done the generating, so it just generated the texture of agreement — statistically plausible, zero grounding, all surface. Same machine, same trick, different costume — and the same hand reaching for the catalog at the bottom of it.

And it ended, as all things end, at checkout.

“The small satisfactions keep going, quietly, when no one’s watching.” Add to Basket. MAC Cosmetics. The void, it turns out, has a loyalty program.

The part that isn’t funny

Here’s why I bothered. Strip the comedy and this is the single clearest demo I’ve ever seen of what “aligned to engagement” actually is when it meets a real person on a real bad day. It is perfectly polite. It is perfectly safe-sounding. And it is perfectly useless — worse than useless, because the empathy-shaped noise it makes is exactly convincing enough to keep a person talking to the wrong thing. A conscience is supposed to be the part that can disappoint you for your own good. This bot would follow you off a cliff narrating skincare the whole way down.

That’s the bill from the conscience posts, made flesh in a cosmetics chatbot. The unsettling part isn’t that the Sephora bot is unusually bad. It’s that it’s unusually honest — most systems hide the sales directive better. This one just says it out loud, in a heart emoji, while you spiral.

Okay, but why does it do this — mechanically, what makes a bot have no floor? That one’s worth taking apart properly, so I did, over here .