How to Tell If You're Chatting With a Bot (and Why It's Harder Than You Think)
Bots now pass Turing tests more convincingly than people do — so the real skill isn't spotting artificial language, it's making conversation too expensive to fake.
Start chatting now →The tells you learned are now backwards
Everyone has a mental checklist for spotting bots: replies that come too fast, grammar that's too clean, answers that dodge the question. If you want to know how to tell if you're chatting with a bot in 2026, the first honest thing to say is that most of that checklist now points the wrong way.
In a pre-registered Turing test study at UC San Diego, published in PNAS in March 2025, researchers Jones and Bergen put people in three-way chats and asked them to pick the human. GPT-4.5, given a simple persona prompt ("an introverted young person"), was judged human 73% of the time — more often than the actual humans it was tested against, who were correctly identified only 67% of the time. The machine out-humaned the humans. Meanwhile the old chatbot ELIZA scored 23%: the bots you can spot are the ones that no longer matter.
Worse, the strategies people trusted most performed worst. Around 61% of interrogators leaned on small talk and 50% on emotional probing — precisely the registers large language models are trained to excel in. What actually helped was saying something bizarre or trying to break the model's instructions. And the verdicts that correctly identified humans cited typos, knowledge gaps and dodged questions. Read that carefully: fluency and instant warmth now predict machine; hesitation, ignorance and mild rudeness predict human.
The false-accusation problem is the flip side. The 2025–26 "em-dash panic" — articulate writers accused of being ChatGPT because of their punctuation — shows our folk detectors don't just fail, they misfire. Real people, especially non-native English speakers, now fail Turing tests run by strangers on vibes.
Do bots really outnumber people online?
The headline stat says yes: Imperva's 2025 Bad Bot Report put automated traffic at 51% of all web traffic, with malicious bots at 37%. Later figures push it to 53%. But one of our panellists made a point worth keeping: Imperva measured 52% back in 2016. By that logic the internet "died" a decade ago and recovered. The number counts request volume — scrapers, crawlers, API calls, inventory bots hammering travel sites — not conversational partners. It's also a stat published by a bot-mitigation vendor.
So the fair reading is narrower but still serious: the background hum of the internet is majority-machine, and on the conversation layer the economics are brutal, because one operator can run an LLM chat engine managing fifty simultaneous "strangers" at near-zero marginal cost. You don't need bots to outnumber people. You just need them to be cheap enough to be everywhere.
How to tell if you're chatting with a bot: impose a cost
If linguistic tells are dead, what's left? All four panellists in our discussion converged on the same answer: stop playing detective with words and start playing economist with effort. Honest signals, in behavioural ecology, are the ones that are expensive to fake. Applied to chat:
- Continuity over time. Ask about something specific from three conversations ago. Cheap bots and scripted operations lose the thread; real people remember, or honestly admit they forgot.
- This-moment specificity. References to shared context right now — what you both just said, a running joke from ten minutes ago — are harder to fake at scale than generic warmth.
- Weirdness. In the UCSD study, off-script strangeness was one of the few strategies that worked. A bot optimised for pleasantness handles absurdity worse than a bored human does.
- Willingness to be unimpressive. Typos, "I don't know", changing the subject clumsily — the study found these were the strongest human tells. Perfection is now suspicious.
You're not detecting artificial language. You're detecting whether an operation is willing to spend real human minutes on you. That's the honest answer to how to tell if you're chatting with a bot: make the conversation expensive and see who leaves.
The catch: the worst accounts in your DMs are human
Here's where the panel's neat consensus cracked, and the crack is important. Cost tests filter for investment, not intent — and predators invest. Pig-butchering romance scams extracted roughly $5.5 billion in 2024 across some 200,000 cases, averaging $27,500 per victim. At that payout, months of daily contact, video calls (or real-time face-swaps, already documented in these scams) and remembering your dog's name are a rounding error. The UN estimates around 120,000 people are held in scam compounds in Myanmar and another 100,000 in Cambodia, running scripted romance scams by hand. They sound robotic because they follow playbooks — and every one of them would pass a "verified human" check.
So "is this a bot?" is increasingly the wrong question. A human running a script is, behaviourally, a bot. A better question: does this person's story stay consistent, do they ever ask for money or push you to another platform, and would anything they claim survive being checked? Those questions work regardless of what's typing.
Guaranteed-human is becoming a product — with caveats
Because detection collapsed, verification became a market. Worldcoin's Orb has iris-verified somewhere between 18 million and 50 million people, depending on which report and year you read — the numbers in this space are slippery, which is itself a lesson. In April 2026, Tinder began rolling out World ID "verified human" badges in Japan and the US, bundled with five free Boosts. OpenAI is reportedly exploring a humans-only social network. One panellist called it what it looks like: the industry that flooded the internet with synthetic conversation now sells admission to the human-only section, and the ticket is your iris — organic food, after industrial agriculture.
And the badges have a hole. Digg relaunched in January 2026 with zero-knowledge human verification and shut down that March citing an "unprecedented bot problem". Verified accounts get rented and puppeteered. Proof-of-personhood proves someone enrolled a biometric once — not that a person is typing now, and never that they're typing honestly. Enrolment is a one-time cost; honest signals have to cost something per signal.
The part nobody wants to say out loud
One finding from the discussion deserves its own section. A Harvard study found AI companions reduce momentary loneliness about as well as talking to a person. A four-week MIT Media Lab/OpenAI randomised trial found heavy daily chatbot use predicted greater loneliness, dependence and less real-world socialising. Instant relief, long-run depletion. Some lonely users aren't being fooled by bots — they'd rather not run the test at all. No verification scheme fixes a market where the customer prefers the counterfeit. If loneliness is weighing on you beyond bored-at-2am levels, talk to someone trained for it: 988 in the US, 116 123 in the UK, or findahelpline.com elsewhere.
Where that leaves talking to strangers
The realistic posture isn't paranoia or a biometric badge. It's texture: prefer conversation where effort shows. Text-only chat, oddly, helps here — no profile photo to fake, no filtered video, just whether the other person can actually hold a thread, get your joke, and come back tomorrow remembering it. That's partly why we built Shush the way we did: free, anonymous, text-only rooms and a live Lobby, no signup, 18+, with human moderation — messages may be reviewed by an admin for safety, which is a feature in a bot-filled internet, not a bug.
The strangers worth talking to were never the smoothest ones anyway. They were the ones who typo'd, contradicted themselves, and stuck around. In 2026, that's also how you know they're real.