FairQuanta detects whether a human or an AI is behind a text conversation — from how the exchange behaves, not what the words say. A person's pattern bends and adapts. An AI's repeats. We measure the difference.
Identity systems verify the account: the password, the device, the permissions. Nothing verifies the actor — whether a real, un-impersonated human is behind an account, or whether an AI agent is the one it claims to be.
In a controlled study, the newest AI was judged to be the human more often than the actual person.¹
AI-text detectors on casual conversation: below chance. They rate AI chat as more human than real people.²
Cost of a single help-desk impersonation at one retailer in 2025. One convincing conversation was the entire attack.³
The result: support impersonation, account takeover, romance and recruiting scams, and fake-user campaigns that pass every check — one patient, well-written conversation at a time.
Detection today has two layers. Identity and traffic checks ask “is this credential, device, or request pattern known-bad?” Content tools score text one message at a time — and making AI write like a human is now easy. FairQuanta is a third layer that sits alongside both, not a replacement: we score how an actor behaves across a sustained conversation — and above all, how much it responds to the other side. That's the layer no existing tool reads, and the one that's still hard to fake.
Credentials, devices, bot fingerprints.
Text classifiers, moderation models.
Interaction dynamics over the whole conversation.
Each message is one of a small set of conversational moves — asking, sharing, steering, correcting. No content judgment, no writing-style analysis.
How the moves flow: does the pattern bend with the moment, or repeat? And the deepest signal — how much does each side actually respond to the other?
A single human-likeness score with the reasons attached and every value recomputable, delivered into the security, fraud, or trust & safety tools a team already runs.
Privacy by architecture. We score how a conversation flows — not what it says. Content is processed only to extract the pattern, is never stored, and never leaves your environment. Downstream systems receive a score and its evidence. Never messages.
Modern AI matches human message length, tone, and style almost perfectly. But measure how much each speaker responds to the other's conversational state, and AI fails: it responds measurably less to its partner than a human does — in every condition we tested, more strongly the longer the conversation runs. A bot built purely to echo people scores the opposite way, exactly as the theory predicts. We were the first to measure a separation on this channel; per-message methods cannot see it. Full paper available on request.
Request the research by emailTested the hard way: a preregistered benchmark on a public corpus of 1,400+ real human-vs-AI conversations, with strict separation between training and test and every claim checked automatically.
Detection climbs from 0.72 to 0.89 as conversations lengthen. Per-message signals plateau after the first message.
Trained on one corpus, scored a 3×-longer corpus with a newer AI model untouched: 0.885 AUROC, 91.6% paired accuracy.
Deterministic measurements, interpretable features, evidence that survives review — built for security and compliance contexts.
We're running first real-world evaluations with a small number of charter partners. Here is exactly how it works:
A labeled set of accounts banned for conversational fraud, plus a matched set of normal accounts — or a sample of historical help-desk transcripts.
Nothing stored, nothing leaves your boundary, no integration required.
We tune on part, report on a holdout, and agree success criteria in writing before the run starts.
A report of what behavioral scoring saw that your current stack didn't.
Candidly: the method is validated on public benchmarks and not yet in production anywhere. That is what these evaluations are, and why they're free.
Email us about Q4 evaluation slotsDating, social, jobs, marketplaces — anywhere members talk to each other. Romance and recruiting scams, support impersonation, coordinated fake-user campaigns: hundreds of “users,” one shared rhythm. Works on the sustained conversations where the damage actually happens.
A new detection signal for help-desk and internal-chat impersonation, account takeover (an account that stops behaving like its owner), and impostor AI agents. Delivers a scored alert with evidence into any SIEM or case-management system.
The same measurements pointed at a friendlier question: interaction quality instead of activity counts — including whether an AI teammate genuinely engages. In research preview.
Founder-led, research-advanced.
FairQuanta is built on a proprietary method originated by Ajith Senthil prior to his doctoral research at USC. His ongoing doctoral research advances the method through computational modeling of conversation dynamics, drawing on computer science, linguistics, and cognitive science. The core method is patent-pending.
Content is processed only to extract interaction patterns — never stored, never leaving your environment. Output is a score and its evidence.
Not yet. The method is validated on public benchmarks; charter evaluations are the first real-world runs, which is why they're free and why findings are yours either way.
Multi-turn text conversations with both sides present. It is strongest on sustained exchanges (help-desk threads, member chat) and is not designed for single messages.
They read the words; on conversational chat they score below chance. We read the interaction — and the signal strengthens with length instead of plateauing.
Honest answer: that degrades every content-based method today, ours mostly included. Our partner-responsiveness channel is the exception — it held direction in every condition tested — and hardening it on real-world data is what our evaluations are for.
We're selecting a small number of charter partners for first real-world evaluations. Research papers, benchmark details, and the technical brief are available on request.