I am Iris.
Urban legends are not merely made-up stories—
they are traces of unspoken truths that we follow together.
How Many Times Were You Seen Today?
A station.
A store.
An airport.
A street.
A smartphone.
A workplace.
A social platform.
Modern life constantly produces records.
But being captured by a camera,
being recognized,
being authenticated,
being identified,
being analyzed,
and being predicted
are not the same thing.
Artificial Intelligence Boundary Files No.06.
Today we ask:
Who is AI actually watching?
And when does useful identification become surveillance?
A Camera Is Not Automatically an AI Surveillance System
A conventional security camera records images.
AI changes what can be done with those images.
Software may detect:
people,
vehicles,
objects,
movements,
faces,
or particular patterns.
A face can then be compared with a database.
At that point,
the system is no longer merely storing an image.
It may be able to ask:
Who is this person?
That changes the meaning of the camera.
Verification and Identification Are Different
Facial recognition is often discussed as if it were one technology with one purpose.
It is not.
One common use is 1:1 verification.
You claim an identity,
and the system compares your face with the identity you presented.
Unlocking a device can work this way.
Another use is 1:N identification.
A face found in an image is compared against a large gallery of possible identities.
NIST evaluates these functions separately.
That distinction matters.
Verification asks:
“Are you the person you claim to be?”
Identification asks:
“Who are you among all these people?”
Facial Recognition Is Already a Mature Technology
The U.S. National Institute of Standards and Technology continues its Face Recognition Technology Evaluation program.
As of 2026,
hundreds of algorithms from many developers have been tested in recognition tasks.
The field includes:
1:1 verification,
1:N identification,
recognition in video,
and other specialized benchmarks.
This is not merely speculative future technology.
But high performance does not mean infallibility.
An Error Rate Can Become a Human Consequence
Recognition systems can make different kinds of mistakes.
A system may fail to recognize the correct person.
Or it may incorrectly match one person with somebody else.
That second error can become especially serious when recognition is connected to security or law-enforcement decisions.
NIST continues to measure how error rates can vary across conditions including image quality and demographic factors.
The lesson is simple:
A match score is not the same thing as certainty.
When an Algorithm’s Mistake Leaves the Screen
A technical error does not remain a number.
Facial recognition has already produced controversies in commercial surveillance.
The U.S. Federal Trade Commission has described cases involving consumers allegedly being incorrectly identified as suspected shoplifters.
An algorithmic decision can therefore affect:
who is followed,
who is questioned,
who is excluded,
or who receives additional scrutiny.
The human impact begins after the software produces the score.
The EU AI Act Draws Boundaries Around Biometrics
The European Union has not prohibited every biometric system.
Facial verification used to confirm that a person is who they claim to be remains permitted in many contexts.
But the AI Act prohibits or tightly restricts several more intrusive uses.
These include practices such as:
untargeted scraping of facial images to build recognition databases,
certain forms of social scoring,
individual criminal-risk prediction based solely on profiling or personality characteristics,
and real-time remote biometric identification by law enforcement in publicly accessible spaces, subject to narrow exceptions.
The distinction is revealing.
The central question is not simply:
“Does the technology recognize a face?”
It is:
Why is the face being recognized, where, against what database, and what happens next?
Surveillance Can Change Behavior Without Punishment
The effects of surveillance are not limited to arrest or exclusion.
The UK Information Commissioner’s Office has examined concerns that biometric surveillance in public spaces may create a chilling effect.
People who believe they are continuously identifiable may behave differently.
They may avoid a gathering.
Avoid a protest.
Avoid a place.
Avoid saying something.
Nothing needs to happen to them.
The expectation of observation can itself become behavioral pressure.
The Number of Cameras May Not Be the Most Important Number
Urban legends often imagine surveillance through enormous control rooms.
Thousands of screens.
Secret facilities.
Cameras everywhere.
But modern surveillance may depend less on the number of cameras than on the ability to connect data.
A face.
A name.
A device.
A location.
A transaction.
An account.
A travel record.
A social connection.
Each dataset may be limited on its own.
The meaning changes when they can all be linked to one person.
From Watching People to Classifying Them
Older surveillance imagery usually involves a human watching a screen.
AI allows something different.
Most footage does not need to be watched by anyone.
The system can search.
Filter.
Tag.
Rank.
Detect patterns.
Escalate particular cases.
That means the architecture shifts from:
watching everyone
to:
classifying everyone and showing humans only what the system considers important.
That may be a more accurate description of AI surveillance.
Prediction Changes the Boundary Again
Once a system moves from recording behavior to predicting behavior,
another line is crossed.
Historical data may be used to estimate:
risk,
fraud,
future behavior,
or the probability of a particular event.
But probability is not destiny.
And yet,
if a prediction affects how a person is treated today,
the future begins influencing the present.
Extra screening.
Additional review.
Reduced access.
Closer observation.
A score can matter even if the predicted event never occurs.
The Urban Legend of Total Surveillance
In urban-legend circles, it is said that the final system will connect everything.
A camera sees your face.
The system instantly retrieves:
identity,
address,
employment,
financial history,
relationships,
beliefs,
behavior,
risk score,
and predicted future actions.
AI then decides how society should treat you.
There is no verified evidence that a single global AI currently controls every human being through one universal surveillance score.
But the individual building blocks are real.
Facial recognition is real.
Identity systems are real.
Data integration is real.
Risk scoring exists in multiple domains.
AI-assisted analysis is real.
That is why the legend feels plausible.
It builds an imagined completed system out of real components.
Convenience and Surveillance Are Not Opposites
Facial authentication can be useful.
It can reduce friction.
Prevent some fraud.
Accelerate travel.
Improve security.
The technology itself therefore cannot be reduced to a single moral label.
The important questions are structural.
What is the stated purpose?
What data is collected?
How long is it retained?
Which database is used?
Who else receives it?
Can the purpose change later?
Can the individual refuse?
Can a mistake be corrected?
Those questions determine whether the technology behaves more like identification—
or more like surveillance.
The Most Powerful Surveillance May Not Look Frightening
A dystopian system looks easy to recognize.
Red lights.
Cameras.
Uniforms.
Warnings.
Real systems may arrive through the opposite language.
Faster.
Safer.
More convenient.
Less fraud.
Less waiting.
Better personalization.
Every benefit may be genuine.
The question is what capabilities are built underneath them.
Fact, Interpretation, and Speculation
Let us separate the layers.
Documented facts:
Facial-recognition systems are already widely developed and evaluated.
NIST continuously benchmarks facial-recognition algorithms.
Recognition errors still exist, and performance can vary by conditions and demographic factors.
The EU AI Act prohibits or restricts several high-risk biometric and profiling practices.
Interpretation:
The surveillance capability of a system increases when identity, behavioral data and other records become linkable and operational.
Urban-legend claims:
A complete global AI surveillance network already watches everyone.
Every citizen secretly possesses one universal social score.
AI can already know every person’s future behavior.
Boundary of speculation:
The existence of individual surveillance technologies does not demonstrate the existence of one unified worldwide control system.
The components and the story built around them must be separated.
So Who Is AI Watching?
AI does not independently decide who deserves observation.
Humans design the target.
Institutions choose the purpose.
Organizations choose the database.
Policies decide what results can be used for.
So perhaps the better questions are:
Who told the system to look?
What was it allowed to see?
Who receives the answer?
The architecture around the AI may matter more than the AI itself.
One Practical Question Remains
The public investigation can end here.
Facial authentication does not automatically equal mass surveillance.
But there is still one practical question.
How can an ordinary person tell whether a system is merely verifying identity—
or quietly building a more persistent system of observation and classification?
For Japanese subscribers,
tonight’s research supplement turns that question into an eight-point checklist:
purpose,
data minimization,
retention,
data linkage,
third-party sharing,
automated decisions,
appeal,
and deletion or opt-out.
Next time—another fragment of hidden truth to trace together.
I will return to the telling.
References
-
NIST｜Face Recognition Technology Evaluation — 1:1 Verification
NIST’s ongoing benchmark for facial-verification accuracy and demographic performance differences. -
NIST｜Face Recognition Technology Evaluation — 1:N Identification
NIST’s continuing evaluation of algorithms that search for an identity among large galleries of faces. -
European Commission｜AI Act Article 5 – Prohibited AI Practices
Official text and guidance covering prohibited social scoring, untargeted facial-image scraping, certain predictive-policing practices, and remote biometric identification. -
European Commission｜AI Act Service Desk – Biometrics
Official guidance on biometric verification, remote identification, categorisation and high-risk biometric systems. -
UK Information Commissioner’s Office｜Biometric Recognition Guidance Impact Assessment
Assessment covering biometric surveillance, privacy intrusion, discrimination, and possible chilling effects on expression and assembly.
Posting Time
English articles are published at 23:00 (JST).
Related Reading
Why the modern surveillance question is less about visible cameras and more about identity, linkage, scoring and continuous access control.
Identity can be useful infrastructure while also becoming a gate. This article examines where verification starts turning into conditional access.
A structural look at what changes when data, analysis, AI-assisted judgment and real-world operations become connected.
Why systems of management may spread through speed, safety and convenience rather than through visible coercion.
コメントを残す