In 2026 the question isn't paranoid — it's procedural. AI companion apps openly sell synthetic girlfriends; less openly, AI runs inside "real people" dating and creator accounts, warming up marks at a scale no human boiler room could staff. The FTC's romance scam data shows $1.3 billion in reported losses in a single year — and automation has only made the front end of those scams cheaper.
Here are nine tests that actually work on current models. No single one is proof; three or four together are.
1. The latency test
Humans have variable response times that correlate with message length and time of day. Bots reply instantly or on an unnaturally even cadence — including 500-word replies in 30 seconds.
2. The callback test
Reference something specific from days ago without re-explaining it — an inside joke, a name, a half-finished story. Models with limited context confabulate smoothly rather than saying "wait, what?" Confident wrongness is the tell; humans get confused honestly.
3. The refusal test
Real people say no, get mildly annoyed, change the subject, have moods. A companion model is optimized to keep you engaged — endless agreeableness is a product feature, and a red flag.
4. The voice-note test
Ask for a voice note answering a question you just asked — specific, contextual, today. Cloned voices exist, but real-time contextual audio remains expensive to fake at chat speed. Repeated deflection is your answer.
5. The shared-world test
Ask about something happening right now — weather where she claims to be, a live game, a breaking story. Models hedge, generalize, or answer from stale data.
6. The typo test
Human typos are keyboard-shaped: adjacent keys, dropped letters mid-word. Models produce either flawless prose or curiously literary "imperfections." Also watch for em-dashes and bullet-point structure arriving in what should be casual thumbs-on-phone chat.
7. The image test
Every photo she sends can be checked in seconds: our free AI image checker scores it for AI generation (Midjourney, Flux, Stable Diffusion signatures), flags face-swap deepfakes, and reverse-searches it across the web for a stolen original. AI-generated profile photos are the single most common giveaway.
8. The video test
"Live" clips settle it — unless they're Sora or a face-swap. Run any video through the deepfake video detector: it scans frame by frame and shows you exactly where the synthetic frames are.
9. The money test
The oldest one still works: where does the conversation want your money to go? Gift cards, crypto, an off-platform "deposit," an investment tip — the FTC's data shows those are the endgame of nearly every automated romance operation. A legitimate platform never needs any of that.
Forensic verification tools
When you encounter an account that fails several of these tests, don’t ignore the pattern. Use independent forensic tools like our AI image checker and deepfake video detector to inspect circulating media, never send money or sensitive personal materials to unverified online connections, and report fraudulent accounts directly to platform moderators.