The five-minute behavioral test

You don't need a detector for a single chat message — you need better questions. Bots fail predictably at things every human handles without thinking:

  • The memory test. Mention a detail — your order number, your city — then ask about it three messages later. Scripted bots and many AI agents lose the thread or re-ask for what you already gave.
  • The curveball. Ask something slightly off-script: “What's the weather like where you are?” or “Can you phrase that as if I'm twelve?” Humans improvise imperfectly; bots deflect, loop, or comply with eerie instant polish.
  • Speed and shape. Three flawless paragraphs delivered in two seconds, at 3 a.m., with the same tone every time — no human support agent works like that.
  • The escalation request. Ask directly: “Are you an AI, and can you transfer me to a person?” Legitimate operators answer honestly; evasion is itself an answer.

Checking a whole transcript

Statistical detection needs length: one chat reply is far too short for any detector — a tool confident about 15 words is guessing. But a full conversation is a different object. Once a transcript passes roughly 150 words, you can paste the whole thing (just the agent's side, ideally) into the Cobalynx checker and get a calibrated probability with a published error rate — including an honest “inconclusive” when the transcript is genuinely ambiguous. Uniformity across many replies is the pattern single messages can't show.

Your right to know — this is now law in the EU

Since August 2, 2026, Article 50(1) of the EU AI Act has been enforceable: companies deploying AI systems that interact with people must inform you that you're interacting with AI, at first contact, unless it's already obvious to a reasonably observant person. Penalties reach €15 million or 3% of worldwide turnover. So in the EU, “am I talking to a bot?” is a question you're entitled to have answered — asking directly isn't paranoia, it's exercising a legal right, and a company that dodges it is taking a regulatory risk on top of a trust risk. The full legal picture — who's bound, the exemptions, deadlines elsewhere — is on our AI content law page. (Plain-language summary, not legal advice.)

Why companies blur the line anyway

Support teams deploy AI because it answers instantly and never gets tired; some then dress it in a human name and a stock photo because disclosure feels like admitting cheapness. The result is the worst of both worlds: users who feel deceived when they find out — and they usually find out. If the conversation matters to you (a refund dispute, a cancellation, anything contractual), knowing whether a human or a system made the promise is legitimately important; save transcripts either way.

Running a support bot? The short compliance note

Disclose at first contact, clearly and in the conversation itself — not buried in terms of service. Keep the label accurate through human handoffs, so the customer knows which replies came from whom. Honest disclosure costs a little pride and buys back the trust the industry is currently burning; the deadlines and penalty framework are linked with primary sources from the law page.

What Cobalynx can and can't do here

One chat message is far too short for any statistical detector — including ours; we set a minimum length instead of pretending. Behavioral tests work message by message; statistics need the whole transcript. And no text verdict settles what a company owes you: if you're in the EU, disclosure is your right regardless of what any detector says. For adjacent situations, the same behavioral tests carry over: suspicious emails are covered in the email guide, dating chats — where the stakes are money, not refunds — in the AI dating scam check, and the rest in the FAQ.

Common questions

What is the fastest way to test a chat agent?

Ask a curveball that no script anticipates — a specific question about something you said three messages ago, or a request to deviate from the standard flow. Humans handle it imperfectly but flexibly; bots either loop, deflect, or answer instantly with suspicious polish.

Can I check a single chat reply with a detector?

No — one chat message is far too short for any statistical detector, including ours, and a tool that returns a confident verdict on 15 words is guessing. Paste the whole conversation once it passes roughly 150 words; behavioral tests are the right instrument below that.

I run a support team. What are my obligations?

If your bot serves EU users, disclose it as an AI at first contact — clearly, not buried in terms of service — and keep the disclosure through handoffs, labeling which side of the conversation is human. This is not legal advice; the primary sources and deadlines are linked from the law page.