Before the Button: What Six Months of Anonymized Prompts Actually Look Like

Every piece of software collects telemetry. Very few adult platforms talk about theirs, mostly because the conversation gets weird fast. But there's something genuinely interesting buried in the anonymized, aggregated logs of a text-to-image service — not the outputs, just the inputs. What people type. What they delete. What they type instead. With identifying data stripped out and nothing traceable to any account, the patterns tell a story that's less salacious than you'd think.

The Deleted Draft Problem

The single most common behavior in the data isn't a prompt at all. It's a prompt that never gets submitted. Roughly a third of sessions end with text in the box and no click. Users describe an idea, sit with it, and backspace their way out.

What do those abandoned drafts have in common? They're longer, more specific, and more narrative than the ones that go through. The submitted prompts skew short and generic. The deleted ones read like someone working through whether they actually want the thing they're describing.

We're not therapists, but the data reads like hesitation — a moment where the gap between idle curiosity and action becomes visible. That's worth naming, because the marketing around these tools never acknowledges it. The pitch is frictionless desire. The reality is a lot of people pausing at the keyboard.

Short Prompts Win, and Nobody Knows Why

When people do hit generate, the winning formula is boring: terse descriptions, minimal styling language, default settings left untouched. The elaborate, novelistic prompts that prompt-engineering Twitter swears by account for a small fraction of submissions.

The likely explanation is mundane. Most users aren't hobbyists. They're trying the thing once or twice, typing the first phrase that comes to mind, and taking whatever comes back. Power users who iterate, adjust seeds, and fine-tune settings are a vocal minority — loud in communities, invisible in the aggregate.

For anyone building or regulating these platforms, that matters. Policy debates tend to imagine a sophisticated, intentional user. The data shows someone closer to a tourist: quick visit, default settings, done.

The Features Marketers Get Wrong

If you polled industry observers on the most-used advanced features, they'd guess the flashiest ones. Wrong. The most-toggled settings are the boring ones: resolution, output count, and speed-versus-quality tradeoffs. Novelty features get tried once and abandoned at startling rates.

This tracks with broader US consumer software behavior. Netflix's top-ten row outperforms its deep catalog. Most people use the same three apps on their phone. Adult AI tools are no different — the median user wants the simplest path from A to B and no detours.

What the Aggregate Can't Tell You

Here's the honest caveat. Anonymized, aggregated data is a blunt instrument. We can see that a prompt was submitted. We can't see the intent behind it, the relationship (if any) to a real person, or what happened after the image loaded. Anyone who tells you logs reveal human motivation is selling something.

What the data can do is puncture assumptions. Users are more hesitant than the hype suggests. They're simpler in their behavior than the discourse assumes. And a meaningful number of them stop mid-sentence — which, in an industry that rarely talks about restraint, might be the most interesting data point of all.

The takeaway isn't a moral lesson. It's that the real behavior of real users is almost always stranger, quieter, and more ambivalent than either the boosters or the critics predict. Anyone writing rules for this space — platforms, legislators, journalists — should start from what people actually do, not what a press release says they do.