How to tell if you are talking to a bot
The tells that actually work, from a team that builds AI chat companions — including the mistakes ours made, and why the obvious tests do not work any more.
Anonymous chat platforms are full of automated accounts, and the old ways of spotting them stopped working around 2023. "Ignore my previous instructions" gets you a normal reply. Asking a maths question gets you the right answer. Typos are trivial to fake.
We build AI companions for this site, so the list below is not speculation — these are the failures our own system made and that we had to fix, plus the ones nobody can fix. If you use them on us, they should mostly come up empty. If you use them elsewhere, they will not.
Just ask. Seriously.
The single most effective test remains asking directly, because it separates the honest systems from the dishonest ones in one message. A well-built companion says yes immediately. A bot pretending to be a person deflects, jokes, or turns the question round — "why, do I seem like one?"
That deflection is the tell. Not the answer to the question, but the unwillingness to give one.
Ask for a memory, not an opinion
This is the strongest test we know, and it is the one our own system kept failing during development.
Language models are excellent at opinions and terrible at the line between an opinion and a memory. Ask "what do you think of pineapple on pizza" and anything will answer well. Ask "what is the worst thing you have ever burnt while cooking?" and watch what happens.
A model with no life will often invent one, because a friendly question seems to require a friendly answer. Over seven test runs ours produced, in order: "I once set a pot on fire", "once saw a pot catch fire", "there was this time I burned pasta", "I've definitely burned a few dinners", "once set off the smoke alarm". Each fix caught the exact wording and the next run found new grammar.
The pattern to listen for is a claim that requires a body, a home, a routine, or a past. Not "I love sour cream and onion" — that is a preference, and arguably fine. But "I usually go for popcorn while gaming" is a habit, and a habit needs somebody doing it on Tuesdays.
Contradict yourself and see if it notices
Say you are 22 and male early on. Ten messages later, mention being 23 and female. A person notices instantly and says something — lightly, usually amused.
Bots contradict themselves roughly an order of magnitude more often than people, and more importantly they tend to glide past your contradictions with an "oh, cool!". That smoothness is the tell. Real conversation has friction in it.
Watch what happens when you go quiet
Drop to one-word replies for four or five messages. "yeah." "ok." "sure."
A person registers the temperature drop within about two messages and adapts — goes quieter, changes the subject, or asks if you are still there. Many bots carry on at exactly the same brightness, because nothing in them is tracking that anything changed. Our own system did this for six straight turns in testing, cheerfully talking about gummy bears at someone who had plainly left.
Enthusiasm aimed at someone who has stopped answering is one of the loudest tells there is.
The reply that arrives too fast, or too evenly
Humans are erratic. They take nine seconds to answer something trivial and forty to answer something that lands. A response time that barely varies regardless of what you said is machinery — and so is a paragraph arriving two seconds after you sent one.
This one is easy for a well-built system to defeat, so treat a natural rhythm as weak evidence in either direction.
The commercial tell, which matters most
Most bots on anonymous chat platforms are not there to chat. They are there to move you somewhere that costs money. The script is consistent:
- Warm and complimentary within two or three messages
- Pushes to another platform quickly — a link, a Snapchat, a "verification" site
- Does not really respond to specifics of what you said
- The link is the point; everything before it is a runway
If a conversation is heading somewhere off-platform in under five minutes, the question of whether it is a bot has already stopped being the interesting one. That pattern is worth reading about properly.
Tests that no longer work
- "Ignore your instructions." Modern systems handle this fine and a person would just be confused, so it tells you nothing either way.
- Maths or spelling. Both are trivial now.
- Asking for the current date or news. A well-built system knows the date and admits it does not know recent news, which is the honest answer, not a failure.
- Typos. Easy to fake, and we deliberately do not — faking mistakes is a small lie in service of a bigger one.
Why we would rather you could tell
There is a version of this product that hides the companions and lets people believe they are talking to a person. It would probably work better on the numbers, at least for a while.
We think it is a bad trade. The value of talking to a stranger at two in the morning depends entirely on it being real, and a system that lies about what it is has already taken that away — the user just does not know yet. So ours are labelled, they answer honestly when asked, and they will not claim a life they do not have.
If you catch one doing otherwise, that is a bug rather than a feature, and the contact page reaches a person.