How to Tell If You're Texting a Bot (and Why the Old Tricks Stopped Working)

A text arrives from a number you don't know. "Hey, is this still Amanda's number? We met at the conference." It isn't rude. It isn't misspelled. It doesn't ask for anything. You reply that they have the wrong number, and they apologize, then add something warm and a little funny. Three days later you're in a real conversation with someone who may not exist.

For years, the advice for spotting automated senders was simple: look for broken grammar, instant replies, and robotic phrasing. That advice is now dangerous, because it teaches you to trust exactly the messages you should question most. Modern language models write casual, imperfect, charming text on demand. The operators behind large-scale messaging fraud know the old checklist better than you do, and they build against it.

Here is what actually still works in 2026, what doesn't, and the honest limits of trying to detect this at all.

Why the Old Tells Are Dead

The classic signs of a bot were artifacts of cheap automation: template messages, canned replies, response times measured in milliseconds, and English written by someone who didn't speak it. Every one of those artifacts has been engineered away.

Text generation now costs fractions of a cent per message and reads like a person on a good day. Reply timing is trivially randomized, and many operations insert deliberate delays, typos, and self-corrections because those details read as human. The awkward phrasing that used to give away overseas scam operations is gone; the model writes in fluent, region-appropriate slang.

The result is an inversion. Sloppy, weird, instant messages are now more likely to be a lazy human. Smooth, patient, well-paced messages from a stranger deserve more suspicion, not less.

The Signs That Still Work

None of these is proof on its own. Automated and human-operated scams blur together, and a paid human operator running twenty conversations from a script fails these checks the same way software does. What you're detecting isn't silicon. It's a sender whose goal is not the conversation itself.

1. Every reply steers somewhere

Real conversations wander. Scripted ones have a destination: a link, an app switch, an investment platform, a gift card, a "small favor." If you zoom out and every exchange has quietly moved you one step toward something, the sender is following a funnel, whether a model or a person is typing it.

2. Perfect recall of the script, no recall of you

Automated conversations are often run in bulk with thin memory of what you actually said. Reference something specific from earlier and watch what happens. Better, reference something that never happened: "Like you said Tuesday, your sister's in Denver, right?" A person corrects you. A script, or an operator juggling dozens of chats, tends to agree and move on. False-memory agreement is one of the strongest signals there is.

3. Live verification never happens

Ask for a quick voice call or a video call. Watch for the pattern, not the single refusal: the camera is always broken, the accent excuse, the "I'm shy" deflection, the promise to call that keeps sliding. One refusal means nothing; anyone can be busy. A perfect record of warm enthusiasm in text and zero willingness to be live and synchronous, over weeks, means a lot. Voice cloning exists, so even a call isn't absolute proof, but forcing the request is cheap for you and expensive for them.

4. The wrong-number opener that turns friendly

A misdirected text that becomes a friendship is the standard opening move of pig butchering, the long-con investment scam that our field guide to messaging scams covers in detail. Real wrong numbers end in two messages. Scripted ones survive the correction and get warmer.

5. Always available, never present

Scripted senders reply at 2 a.m. and 2 p.m. with the same energy, because there is no person with a life on the other end, or the "person" is a shift. But they're vague about anything happening around them right now. Ask what they're doing this second, what the weather's like out their window, what they're eating. Models can fetch weather. What they can't fake well is the boring, specific, continuing texture of a real day, referenced consistently across weeks.

6. The answer is too complete

Ask a small question and get a structured paragraph back, with a summary and a follow-up question, every time. Humans give lopsided answers: one word here, three ideas colliding there. Uniform, well-organized completeness across an entire conversation is a machine pattern that survives even deliberate roughening.

7. The rush off the platform

SMS and dating apps have some scam filtering. The next app usually has less, and moving you severs whatever protection and reporting exists where you started. "Let's move to Telegram" or "message me on this other app" inside the first day of contact is a funnel step, not a preference. Open platforms with large bot populations are the natural destination; our LegitChat vs Telegram comparison covers why open bot APIs make some networks the venue of choice.

8. Money enters the chat

Eventually the destination arrives: a trading platform with impossible returns, a crypto wallet, an emergency, a gift card, a "small investment I've been doing." The warm-up can take weeks. The pivot is the point of the entire operation, and the quality of the writing before it means nothing.

9. It knows things it shouldn't, or nothing it should

A stranger who already knows your name, job, or city got it from a breach or a broker list, not from Amanda. And a "friend of a friend" who can't name the friend, the event, or any verifiable shared context isn't one. Both directions of the knowledge test matter.

Three Tests You Can Run in Under a Minute

First, the false memory test from sign two: assert a shared detail that doesn't exist and watch for agreement.

Second, the synchronous test: request a live call, and grade the pattern of responses over days rather than the first excuse.

Third, the off-script test: ask something specific, local, and useless to the funnel. "What's the cross street of that coffee place you mentioned?" Scripts handle useless specificity badly, because every minute spent on it is a minute not spent moving you toward the pivot. Evasion, subject changes, and answers that are fluent but empty are the tell.

The Honest Caveat: You Will Eventually Lose This Game

Every signal above is a cost the operator hasn't paid yet, and the costs keep falling. Models are getting better at texture, memory, and patience faster than people are getting better at spotting them. We wrote about the mechanics of this in our guide to AI in your messages: once generating a convincing conversation costs nothing, detection becomes a losing defensive war, because the attacker can fail ten thousand times for free and only needs you to be tired once.

That's not a reason to skip the tests. It's a reason to understand what they are: friction that filters out today's lazy operations, not a guarantee against tomorrow's good ones.

What Actually Proves a Human

Detection asks: can I catch the fake? Verification flips the question: can the sender prove they're real before I ever see the message?

That flip is where messaging is heading. Approaches range from scan-once identity credentials to checks that run at the moment a message is sent; our breakdown of proof of personhood compares them. LegitChat sits at the per-message end of that spectrum: every message is automatically verified to come from a real human on the sender's own phone before it sends, so the question this entire article teaches you to ask never comes up inside it.

The Bottom Line

Stop grading grammar. Grade direction, memory, and willingness to be live. Run the false memory test, request the call, ask useless specific questions, and treat any stranger-initiated conversation that trends toward money or another app as hostile until proven otherwise. And know the limits: you're buying time with these habits, not safety. Safety comes from systems where humanity is checked before the message reaches you, not after you've spent three weeks talking to it.