Start by asking the agent a few questions

A detector is no use on one chat message, so what works instead is asking better questions, since bots fail in predictable ways at things a person does with no effort at all. These are the ones we would try, in roughly this order.

  • The memory test. Mention some detail early on, like your order number or the city you are in, and then ask about it three messages later. A scripted bot, and a good many of the AI agents as well, will lose the thread and ask you again for the thing you already gave them.
  • The curveball. Ask something a little off the script, like “what is the weather like where you are” or “could you say that again as if I were 12”. A person improvises, and usually does it a bit clumsily, while a bot deflects, or loops back to the script, or does what you asked with an instant polish that no tired person on a shift would manage.
  • Speed and shape. If you get 3 flawless paragraphs back in 2 seconds, at 3 a.m., in the same tone as every previous reply, you are almost certainly not talking to a person, since no support agent on a shift writes like that.
  • Just ask. Say “are you an AI, and can you transfer me to a person”. A legitimate operator will tell you the truth, and we would treat a reply that dodges the question as a yes.

Checking a whole transcript

A scan like ours needs length, and there is no way round that. One chat reply is far too short for it, and a tool that sounds confident about 15 words is guessing. A whole conversation is a different thing. Once you have got past roughly 150 words you can paste it into the Cobalynx checker, and if you can pull the two sides apart, paste the agent’s side on its own. What comes back is a probability with a published error rate behind it, or an “inconclusive” when it could have gone either way. A transcript carries something one message cannot, which is repetition, the same tone and the same structure reply after reply, and that is the kind of thing the scan picks up.

In the EU you now have a right to know

Since August 2, 2026, Article 50(1) of the EU AI Act has been enforceable. A company that deploys an AI system to interact with people has to tell you that you are dealing with AI, at first contact, unless that would already be obvious to a reasonably observant person. The penalties go up to €15 million or 3% of worldwide turnover. So in the EU, “am I talking to a bot” is a question you are entitled to have answered, and asking directly is exercising a legal right. A company that dodges it takes on regulatory risk on top of the trust it loses with you. Who is bound by this, what the exemptions are, and what the deadlines look like in other places, is all on our AI content law page. That page and this one are a plain language summary and not legal advice.

Why companies blur the line anyway

Support teams put AI in since it answers straight away and it never gets tired, and there is nothing wrong with that on its own. It goes wrong when the bot gets a human name and a stock photo of a smiling woman with a headset, and then people find out, which they usually do, and feel they were lied to. If the conversation matters to you, say you were arguing about a refund or cancelling something or agreeing to anything you might have to rely on later, then it matters whether a person or a machine made you the promise. Save the transcript either way.

If you run a support bot

Tell people at first contact, clearly, in the conversation itself, and not only in a paragraph of the terms of service, since people do not read those. Keep the label right when the chat gets handed over to a person, so the customer can tell which replies came from whom. The deadlines and the penalties, with the primary sources, are linked from the law page.

What Cobalynx can and cannot do here

The questions up the page work one message at a time, and the statistics only start once you have the whole transcript, so we set a minimum length rather than guess under it. And nothing a scan says about the text settles what the company owes you. In the EU you have a right to be told, whatever any tool says. If it was an email that worried you, the email guide covers that, and if it was a dating chat, where what you stand to lose is your money, that is covered in the AI dating scam check, and the rest is in the FAQ.

Common questions

What is the fastest way to test a chat agent?

Ask something the script did not plan for. We would go back to a detail you mentioned three messages ago and ask about it, or ask them to step off the standard flow for a moment. A person handles that imperfectly but they handle it, whereas a bot will loop, or deflect, or answer in an instant with a polish that gives it away.

Can I check a single chat reply with a detector?

No. One chat message is far too short for any statistical detector, ours included, and a tool that gives you a confident verdict on 15 words is guessing. What we would do is wait until you have the whole conversation and it has passed roughly 150 words, and paste that in. Below that length the behavioural tests on this page are the right tool.

I run a support team. What are my obligations?

If your bot serves people in the EU, tell them it is an AI at first contact, clearly and in the conversation and not buried in the terms of service, and keep that disclosure going through any handoff so it is clear which side of the chat was a human. We should say that this is not legal advice. The primary sources and the deadlines are linked from the law page.