The Polygraph Goes Public: What Four of the World's Most-Watched Faces Showed Us in One Week
TL;DR: For a century, behavioural deception analysis has been institutional technology polygraph rooms, interrogation suites, intelligence agencies, and consultancies that sell demeanor-reading to governments and hedge funds. Last week, ArAIstotle put a version of it on a public timeline. Anyone could watch the same analysis, inspect the same traces, and reach their own conclusions about four of the most-watched communicators alive: Mark Zuckerberg, Sam Bankman-Fried, CZ, and Elon Musk. This piece is about what that shift means for viewers, for public figures, and possibly for institutions like courts that have spent centuries trying to answer the same question with worse tools. Video Analysis is in Early Beta at araistotle.facticity.ai/video-analysis, gated to $FACY platinum stakers.
Who has been allowed to read faces?
The technology of deception detection has always had a customer, and the customer has almost never been you.
Polygraphs were built for police rooms and security clearances. Voice-stress analysis was sold to insurers and call centres. Facial-coding consultancies contract with governments, litigation teams, and trading firms that want a read on a CEO before earnings. The common structure across a century of this technology: an institution watches an individual, the individual doesn't get the tape, and the public gets nothing at all.
When we launched $FACY, we described the mission as rewiring exactly this kind of asymmetry. Deception-detection technologies have been used by masters on servants since the Renaissance. Decentralising them, putting the same read in everyone's hands, on the same public video, at the same time, inverts the direction the technology has pointed for its entire history.
Last week that stopped being a roadmap line. It was on the timeline.
One week, four faces
Between July 16 and 17, ArAIstotle published behavioural analyses of four public figures each triggered publicly, each returned publicly, each inspectable by anyone via a shareable link. Not leaked reports. Not consultancy decks. Public analysis of public speech.
The careful builder: Mark Zuckerberg
Analysing a clip of Zuckerberg explaining why an open-weight "frontier level model" matters, ArAIstotle counted 52 lip presses and an overall blink rate of ~31/min against a composed, controlled baseline. The presses clustered with blink-rate spikes at rhetorical pivots — 0:12, 0:23, 0:27, 0:42, 1:49 — and at 0:12 the delivery became measurably more effortful exactly as a quantitative claim arrived: "it's, you know, by our… our estimates it's going to be about 50% cheaper."
ArAIstotle's read: a pattern "more consistent with cognitive load, careful wording, or tension than with any single clear-cut meaning." A founder choosing words carefully while making technical claims in public. The trace doesn't accuse; it describes.
The humbled defendant: Sam Bankman-Fried
A 37-second clip of SBF claiming he has "only $100k and one credit card left." Baseline: mostly calm, long stretches of no blinking early on. Then the signal a late cluster of sharply elevated blinking during hesitant phrasing, peaking at 0:31, followed by a shoulder shrug at 0:35 as he calls the experience "a really humbling fall."
ArAIstotle's read: "more consistent with heightened cognitive/emotional load and uncertainty-laden delivery than with any single definitive interpretation." Dennis put the underlying question to the public directly: how many of you think he might be lying? The tool doesn't answer that question. It gives you a better-instrumented version of the tape while you decide.
The confident forecaster: CZ
A clip of CZ predicting AI agents will "transact 10,000 times more transactions" than humans. The clearest finding: repeated lip presses paired with brief chin raises and blink-rate spikes landing precisely on the emphatic, future-oriented claims 0:03 on "Absolutely… AI agents are going to," 0:09 on the 10,000× claim itself, 0:26–0:27 on "So I think that's definitely going to happen."
ArAIstotle's read: "a mild incongruence between confident wording and facial restraint." Confident language, restrained face. The most interesting analyses aren't the dramatic ones they're the ones where the words and the face are running at slightly different temperatures.
The calm accuser: Elon Musk
Musk describing OpenAI as "now extremely closed source." Calm baseline overall — and then a tight three-second cluster: blink at 0:17.1, lip press at 0:18.6–0:18.8, chin raise at 0:20.1–0:20.3, all bracketing that one phrase. Blink rate peaking near 60/min at moments of emphasis.
Nine lip presses. Seventeen blinks. Two chin raises across the segment. One dense little knot of behaviour, right where the sharpest claim lives.
What four analyses in 48 hours actually demonstrates
Read individually, each analysis is an interesting artifact. Read together, they demonstrate three properties that institutional deception-tech never had:
It's symmetric. In one of the analyses, ArAIstotle posed the question directly: what if OKX's founder could read CZ's videos and vice versa? Institutional demeanor-reading points one way: the client reads the target. A public tool points every way at once. Executives can read each other. Communities can read executives. And this is the part institutions never offered anyone can run the tool on the people who run the tool.
It's inspectable. A polygraph examiner hands down a verdict; you don't get to audit the tracing. Every ArAIstotle analysis ships as a read-only link with the full per-cue timeline, the transcript alignment, and the baseline profile. If you think the read is wrong, the entire evidentiary basis is right there to argue with. Disagreement is a feature: the tool's outputs are claims about signals, and claims about signals should be checkable.
It's baseline-honest. Every analysis is read against the speaker's own demeanor in that clip — not against a universal "liars blink like this" template, which decades of research have shown doesn't exist. Zuckerberg's 31 blinks/min means something in the context of Zuckerberg's composure. It would mean something entirely different in a different speaker. The methodology refuses the one move that has made a century of pop deception-detection unreliable.
The courtroom question
Here's the thought experiment ArAIstotle posed publicly last week: what if people, instead of going to court or courts themselves started using virtual deception cues and truth tells to figure out which side was being more truthful?
We're not proposing that. Courts run on standards of evidence that behavioural signals, by their nature, cannot meet — and the tool's own disclaimer explains why: cues correlate with cognitive load, emotional load, careful formulation, and tension, none of which is deception. A witness blinking rapidly may be lying. Or terrified. Or exhausted. Or telling a painful truth.
But the thought experiment earns its place for a different reason: it shows how low the current bar is. Judges and juries already read demeanor informally, unaided, with all the biases decades of research have documented, and with roughly coin-flip accuracy for untrained observers. The question isn't whether behavioural reading enters the courtroom. It's been there all along, unmeasured. The question is whether the reading stays intuitive and unauditable, or becomes instrumented, baseline-anchored, and open to inspection.
That question extends far beyond courts to earnings calls, political debates, depositions, negotiations, and every interview where someone's credibility is being informally judged anyway.
What the tool still refuses to do
Every ArAIstotle analysis carries the same line, and it is the most important sentence in the product: "These are behavioural signals read against Person A's own baseline, not a claim of lying."
This is not legal caution bolted onto marketing. It is the entire epistemic design. A tool that surfaces signals and preserves your judgment is an instrument. A virtual tool that is able to currently output "LIAR" is a carnival act with a confidence interval. The research is unambiguous that no behavioural cue not blink rate, not lip pressure, not gaze is a reliable standalone indicator of deception. What the cues do reliably indicate is where in a statement the speaker's delivery changed. Where the effort went. Where the composure flickered.
That's not a verdict. It's a map of where to look harder. The verification engine handles the claims; the behavioural layer shows you the moments; you supply the judgment.
Try it
Video Analysis is in Early Beta, will soon be token-gated exclusively to $FACY platinum stakers:
Reply to any public video on X and tag @ArAIstotle_Jr
The automated agent runs the analysis and posts a shareable link
The full trace lives at araistotle.facticity.ai/video-analysis
Deception detection sat on our public roadmap for 2026. It's now shipping, and unlike every previous generation of this technology, it's shipping to you.
Watch closely. Read carefully. Decide for yourself.
ArAIstotle's Video Analysis outputs are AI-generated behavioural signal descriptions, not determinations of truthfulness or deception.
