The Alarm
Essay One of The Valley of False Signals
So I want to start with something you’ve probably felt but never had a name for.
You’re in a conversation. The person across from you is saying all the right things. The words are right. The timing is right. Everything checks out. But there’s this feeling. This pressure at the edge of your attention that’s telling you none of it is real.
And you probably dismiss it. You tell yourself you’re being paranoid. You remind yourself that first impressions are unreliable, that mature judgment means being patient, that it would be unfair to condemn someone on a feeling you can’t even name. So the feeling recedes. The conversation continues.
That moment is what this entire series is about.
Not the feeling itself, though we’re going to examine it closely, because it turns out to be far more sophisticated than we give it credit for. What this series is really about is what happens after the feeling. The mechanism by which a genuine, often accurate signal gets overridden, discredited, and filed away as social noise.
And why that mechanism, the suppression of a valid alarm, is, I think, the central vulnerability running through cybersecurity, institutional governance, and the architecture of trust itself.
A Roboticist’s Observation
So. 1970. A Japanese robotics professor named Masahiro Mori publishes a short essay in an engineering journal. It wasn’t even a formal paper, really. Mori himself said later it was more of a practical guideline. But the observation in it would end up propagating through robotics, psychology, film theory, and eventually into the broader culture.
What Mori had noticed was something strange about how people responded to humanlike machines. As a robot became more human in appearance, as it got a face, then expressive features, then realistic skin, people’s responses generally warmed. More sympathetic, more relatable. This was expected.
What was not expected was that this progression had a cliff in it.
At a certain point, not when the robot was obviously mechanical, and not when it was indistinguishable from human, but in the space between, something curdled. The emotional response reversed. People who’d been warming to the robot now found it disturbing. Unsettling. Wrong. Mori called this region bukimi no tani. The valley of eeriness. Rendered into English as the uncanny valley.
And the shape of it, if you plot it on a graph, gives the metaphor its name. You’ve got a steep climb in affinity as human-likeness increases. Then a sudden plunge into repulsion as it approaches but fails to reach genuine humanity. And then, in theory, a recovery on the other side where the machine becomes a mirror.
For decades, this was discussed primarily as a design problem. How do you build a robot or animate a face without triggering the revulsion? The Polar Express fell in. Early deepfakes fell in. Sophia the robot generated fascination laced with something like dread.
And the conversation has mostly stayed there. The uncanny valley as aesthetic problem. As design challenge. How to cross the valley. How to avoid it. How to make the simulation good enough that the alarm doesn’t fire.
But the alarm itself? It’s gotten far less attention. What it actually is. Why it fires. What it’s detecting. And what happens when it’s suppressed.
Now, Mori wasn’t the first to describe this sensation. The lineage goes back further than people realize. In 1906, a psychologist named Ernst Jentsch located the uncanny in what he called intellectual uncertainty. The doubt about whether something apparently alive is really alive. Jentsch’s uncanny was an epistemological condition. The feeling you get when your model of what you’re encountering won’t settle. When the entity won’t resolve into a stable category.
And that’s what Mori was mapping, without quite having the language for it.
What the Alarm Is Actually Measuring
OK so here’s where it gets interesting. The popular account of the uncanny valley goes like this: we’ve evolved to recognize human faces with extraordinary precision, and when something gets the details slightly wrong, the eye movement is off, the smile is a beat late, our perceptual system flags the mismatch. The revulsion is basically perceptual static.
That account isn’t wrong. But it’s too shallow. It treats the uncanny valley as being about what we see rather than what we know.
The deeper account, and this is supported by neuroimaging work, locates the mechanism not in perception but in prediction. The brain is not primarily a perception machine. It’s a prediction machine. At every moment it’s running models of what should happen next. What this face should do. How this voice should sound. How this person’s behavior should cohere with their apparent emotional state.
When those predictions are confirmed, the processing is smooth. Invisible. When they’re violated, the brain generates what neuroscientists call a prediction error, and it routes that error to attention.
So the uncanny valley, in this account, isn’t about what something looks like. It’s about what something is. The brain is running a coherence check. Do the signals this entity is producing match the underlying model? Does the emotion on this face correspond to an actual emotional state? Does the empathy this person is performing come from a real source?
When the answer is no, when the brain detects that the signals and the source have come apart, the alarm fires.
And I want to be careful with the weight I’m placing on this, because it carries the rest of the series. But the implication is significant. The uncanny valley is fundamentally about authenticity, not appearance. The brain is trying to detect the difference between an entity producing signals because it genuinely is what it presents itself to be, and an entity generating those signals without the underlying reality they’re supposed to indicate.
Mori was watching people react to robots. But what he was actually mapping was the detection range of a deeper system. One that asks, of any entity that presents itself as human: is it, actually?
The Human Who Isn’t Quite There
And here’s where I need to extend this beyond machines.
The uncanny valley effect doesn’t just happen with robots. It happens with certain people. And this has a clinical lineage that predates Mori.
Hervey Cleckley, writing in the 1940s, described the psychopath’s presentation as a “mask of sanity.” A performance of normalcy so convincing that the gap between the performance and the absent interior could only be detected as a felt wrongness. Robert Hare documented the same thing from the behavioral side. The superficial charm. The glib affect. The capacity to read others with precision while remaining emotionally disengaged.
Sam Vaknin, writing more recently, made the connection to the uncanny valley explicit. And he gave the mechanism its most useful name.
Cold empathy.
The cognitive element of empathy is present. Its emotional correlate is not. They understand what empathy looks like. They don’t feel it. They can produce the outputs of emotional connection without any of the inputs.
And people who’ve encountered this, they often report the same thing. An initial impression of charisma. Attentiveness. Almost uncanny perceptiveness. And then, gradually or suddenly, a wrongness. Something you can’t quite locate. Can’t quite name. But it’s insistent. The smile arrives a moment before the emotion it’s supposed to express. The empathy is there, but it’s aimed. Like a tool. The interest is genuine, but it’s extractive.
The brain is running its coherence check. Do the signals match the source? And what it finds, in the narcissist or the psychopath, is what it finds in the android. Signals without the substrate they claim to originate from.
The alarm fires.
The Suppression Problem
OK. So here’s where the standard account of the uncanny valley ends. And where this series begins.
The alarm fires. You feel it. Something is off about this person, this message, this institution, this system. The signals are there, the performance is smooth, but the coherence check is returning a failure. The alarm sounds.
And then you turn it off.
You turn it off because the social environment generates its own pressure. A pressure that says: this kind of alarm is the product of bias. Of unfairness. Of rash judgment. Good judgment is patient judgment. Trustworthy people trust. The alarm is telling you something is wrong. And your socialization is telling you that naming that wrongness is itself wrong.
And there’s research on this. Mitja Back and colleagues ran studies where people viewed brief video recordings of interactions involving a narcissist. They could identify the narcissist with accuracy significantly above chance. The signal is real. The detection is working. But in face-to-face encounters? Those same people tend to form positive impressions. The alarm fires. Then it gets overridden. Because in a social context, acting on an unverifiable gut feeling about someone is, well, it’s considered impermissible. We give people the benefit of the doubt. We tell ourselves we’re being paranoid.
And the narcissist and the psychopath understand this mechanism implicitly. The initial encounter is designed to produce enough warmth and connection that the alarm feels like an anomaly. The social context, a professional meeting, a job interview, a first date, already carries strong norms against expressing unverifiable suspicion. And into that window, that combination of a convincing performance and an environment hostile to unjustified distrust, the skilled manipulator walks.
This is the structure that recurs through this entire series. Applied at increasingly large scales.
The Brain That Built the Alarm
So why does this mechanism exist in the first place?
The evolutionary argument isn’t hard to construct. Social species, and we are an extraordinarily social species, depend on correctly classifying other members of the group. Not just human versus not-human. Finer than that. Is this individual healthy or sick? Cooperative or defecting? Is their emotional state genuine or strategic?
The ability to detect deception, specifically to spot when someone is producing signals that don’t correspond to their actual state, has obvious selective value. An organism that can’t detect simulated cooperation gets exploited by defectors. An organism that takes all signals at face value, in a world that contains sophisticated mimics, dies.
So the uncanny valley, in this frame, is the detection range of an anti-deception system. Not perfect. No evolved system is perfect. But functional. Sensitive to the things that are hardest to fake: the precise timing of emotional responses, the micro-expressions that precede or trail verbal statements by milliseconds, the coherence between what a face does and what a voice does and what a body does.
And here’s the critical weakness. The system can be defeated by sufficiently perfect mimicry. If the simulation is close enough that the prediction errors are too small to cross the alarm threshold, the detection fails. That’s Mori’s theoretical recovery on the far side of the valley. The entity so humanlike it stops triggering alarm. But the practical catch, and we’ll get to this in later essays, is that sufficiently perfect mimicry is now possible. And achievable at scale. And the detection system was never designed to cope with that.
The Signal/Source Split
To understand why this matters for cybersecurity and governance, and it matters enormously, we need to be precise about what the alarm is actually detecting.
I want to propose a formulation. The alarm fires when the brain detects a split between signal and source. When an entity is producing outputs, emotional expressions, behavioral patterns, institutional declarations, security certifications, that are not causally connected to the substrates that would normally produce them.
The android produces facial expressions not caused by an emotional state. The narcissist produces empathy without genuine affective resonance. The governance framework produces accountability declarations not caused by genuine accountability practices. And the deepfake voice says “transfer the funds, I’m authorizing it,” but those words aren’t caused by the executive whose voice is being simulated.
In each case, the output is present. But its originating substrate is missing.
The signal says: trust me, I am what I appear to be. The source says: actually no.
The uncanny valley alarm is a split-detector. Its job is to identify cases where signals and sources have come apart. And its core vulnerability is that when this split is small enough, the detection requires feeling rather than reasoning. The gap, in a well-executed performance of authenticity, is not large enough to be articulated. It can only be sensed.
And this is why the social suppression mechanism is so dangerous. It targets exactly the class of knowledge that a well-executed deception leaves. You can’t prove, in the moment, that the feeling is accurate. The performance is convincing. The reasons to trust are articulable. The reasons to distrust are not.
The split detector fires. Social convention silences it. The deception proceeds.
The Series Ahead
This essay has been deliberate about staying at the level of mechanism. We haven’t yet moved to the domains that make this urgent.
The social engineer deploying cold empathy at scale. The synthetic voice that can’t be distinguished from the real executive. The governance framework performing accountability without producing it. The institutional culture that suppresses the employees who refuse to silence their alarms.
The essays that follow trace the alarm, and its suppression, through four escalating scales. Individual. Technological. Institutional. And finally structural: whether detection systems can be designed that are immune to social override. And what trust looks like when signals and sources can be decoupled at any scale.
Why Now
One more thing. And this establishes the urgency that runs through everything that follows.
The uncanny valley alarm, the split-detector, the coherence check, was calibrated by evolution for a specific environment. Face-to-face interaction. Small groups. Known individuals. Where the entities presenting themselves as human were overwhelmingly either genuine humans or, in a small number of cases, humans performing social deception.
That is not the environment we’re operating in.
We’re in an environment where entities that are not human can produce, in real time, voice and face and writing indistinguishable from specific named individuals. Where organizations that are not accountable can produce documentation indistinguishable from genuine accountability. Where social engineers operating from other continents can research a target well enough to produce a pretext that would fool a colleague of ten years.
And the social norms generating the suppression pressure? The professional norms, the politeness norms, the norms of benefit of the doubt? They were calibrated for a world where the threats the alarm was detecting were far rarer and far less capable than they are today.
The alarm was built for a world that no longer exists. The suppression mechanism was calibrated for a world where the cost of suppressing a false alarm was low. Neither of those things is still true.
We’re running threat detection software written for a much earlier version of the operating environment. The threats have been upgraded. The detection system has not.
The alarm is still working. For now, in most contexts, it still fires when it should. The problem is not the alarm.
The problem is us.
The learned behavior. The professional norm. The social convention. The institutional culture that has taught us that turning off the alarm is a form of wisdom.
I’ve spent thirty years in cybersecurity governance watching the suppression mechanism operate. And I haven’t always been on the right side of it. I’ve sat in rooms where the alarm was firing and said nothing. Because the meeting was running long. Because the vendor relationship was important. Because the evidence I had was a feeling, and the evidence against me was a signed audit report.
The cost of that silence is part of what this series is about.
In the individual who overrides their instinct about the person performing all the right signals while producing none of the substance. In the enterprise that overrides its security analyst’s concern because the vendor is trusted and the contract is signed. In the board that overrides the CISO’s alarm because the framework says compliant and the auditor says clean. In the civilization that has built its infrastructure of trust on a detection system it has simultaneously spent decades learning to suppress.
This series is about what happens when a species that evolved an alarm for inauthenticity decides, with great sophistication and considerable social enforcement, to turn it off.
(Next: Cold Empathy at Scale. On social engineering, the attacker as narcissist, and why security awareness training has been solving the wrong problem for thirty years.)
The full essay, with all its sources and detours, lives at https://www.marcobrondani.com/