We analyzed 4,383 AI dating sessions. People reject 80% of AI suggestions, not because the AI is bad, but because they want to sound like themselves.

People reject 80% of AI-generated replies

Users see generated replies as swipeable cards. 2,233 likes. 9,235 dislikes. That's a 19.5% approval rate across 11,468 recorded swipes.

This is consistent with what other AI dating tools have found. YourMove AI, which has generated over 3 million messages, reported that users heavily filter outputs. 80% of generated messages include a question mark, compared to the full pool, because users selectively copy messages that feel more conversational.

The pattern is clear: people don't treat AI suggestions as finished messages. They treat them more like a brainstorming session, rejecting most options until something feels close enough to edit or send. Research from Scientific American found that people struggle to detect AI-generated text in dating contexts, but they can often detect when something doesn't match their own voice, which is a different and higher bar.

The tone slider is used more than anything else

The single highest event count in our dataset: 37,924 tone adjustment events. That's more than swipes (12,558), more than page views (8,823), more than any other tracked action.

What this means: people are not accepting first drafts. They're sliding the tone back and forth (more casual, more flirty, more chill) trying to close the gap between what the AI produced and how they actually talk.

This connects to a broader finding about how people relate to AI writing tools. A 2024 study by Chen and Toma found that socially anxious people prefer texting over in-person conversation specifically because texting lets them edit. The tone slider is that same instinct: not letting AI write for you, but using it as a draft you can shape until it sounds right.

Most people generate replies and leave

54% of sessions (2,355 out of 4,387) had zero swipes recorded. The median number of swipes per session is 0. The average is 3.4, pulled up by a smaller group of heavy users (17% swiped 10+ times, one session hit 49).

Separately, only about half of people who start the process finish it. Of 3,939 uploads, 1,945 resulted in a completed generation. That's a 49% completion rate.

This fits what the Arrows survey found: while 54% of singles have tried AI for dating, most are experimenting rather than committing. There's a large gap between “I'm curious about this” and “I'm going to paste this into my conversation.” Our session data makes that gap measurable.

AI works better for openers than conversations

We have two main modes: mid-game (replying to an ongoing conversation) and pickup-lines (openers). The difference in behavior is significant.

Pickup-line users swipe more and copy almost twice as often. This makes sense: an opener is lower-stakes. You're not trying to match someone's existing tone in a conversation, you're just picking something to send cold.

This confirms what an AXE-commissioned study found: AI-generated openers achieved a 60% positive response rate (vs 48% for an average person), but AI's advantage disappeared as conversations got deeper. At the commitment stage (asking someone on a date), a genuine, vulnerable human message beat both AI and a professional dating coach. The harder the conversation gets, the more authenticity matters.

Late-night sessions suggest anxiety, not laziness

24% of all sessions happened between 10pm and 6am Eastern. Peak hours were 5pm (290 sessions), 1pm (285), 3pm (264), 10pm (250), and 9pm (245).

The late-night cluster is notable. Research on texting anxiety suggests this isn't casual behavior. A TextVibe analysis of texting anxiety research found that 78% of people in dating relationships overthink text messages, spending an average of 47 minutes daily analyzing message meanings and response times. Waiting for replies activates the same brain region as physical pain.

The people using SmoothRizz at 1am are not being lazy about composing a message. They're stuck, probably anxious, and looking for a second opinion. That aligns with how Hinge frames AI in dating: as a “wingman” for “getting unstuck,” not a ghostwriter.

What users say they're struggling with

During onboarding, users select from frustration categories. Top picks in this window:

“Don't know what to say” and “overthinking replies” together account for 187 selections. These aren't people who want a robot to charm someone for them. They're people who have something to say and can't get it out. That's the same thing the Chen and Toma research describes: socially anxious individuals gravitate toward tools that let them draft and revise, because the editing process is how they manage anxiety.

The bigger picture

The common framing around AI and dating is that people are outsourcing their personality. Our data doesn't support that. What it shows is something more complicated: people want help, they're picky about the help they accept, and the thing they're most picky about is whether the output sounds like them.

This is consistent across every metric. The 80% swipe-left rate. The 38K tone adjustments. The 54% of sessions that end without any engagement. The 2x copy rate for openers over mid-conversation, where your voice matters more. All of it points in the same direction: people are using AI as a sounding board, not a replacement.

Whether that's a good thing for dating is a separate question. The Vice/Arrows reporting found that one in five singles have ghosted someone they suspected of using AI. Norton's 2026 report raises concerns about AI-driven emotional manipulation in dating. These are real tensions. But the assumption that people are handing the wheel to AI doesn't match what we see in the data. Most of them are grabbing the wheel harder than ever.

Methodology

All numbers are aggregate row counts from our production database. Date range is 2026-01-27 through 2026-03-27, with a cutoff at 2026-03-28 UTC. No message content was accessed. Session counts come from the generation_sessions table; event counts from funnel_events. Some event names are from older tracking versions. External studies cited are linked inline and were published between 2024 and 2026.