Why behavioral surveillance cannot predict intentions

Behavioral surveillance and the promise of predicting intentions

The concept of behavioral surveillance is not new, but it gained significant momentum starting in 2008 when the Department of Homeland Security began testing systems aimed at behavioral prediction in sensitive environments such as airports and borders. The proposal was simple in its form yet ambitious in its rhetoric: to use remote sensors and artificial intelligence to identify, in real-time, individuals with potential hostile intentions.

This was not traditional surveillance involving human-operated cameras, but a more complex technological arrangement. Physiological sensors, eye tracking, thermal cameras, facial expression analysis, body movement, and vocal variations would feed computational models capable of classifying emotional states and patterns deemed “atypical.” The most well-known system in this suite became associated with the FAST program—Future Attribute Screening Technology.

Institutional discourse presented preliminary results showing around 70% accuracy. In isolation, this number sounds promising. For managers pressured by efficiency and prevention, it suggests an operational advantage. The problem begins when one asks a basic question: 70% accuracy in relation to what, exactly?

Here, the first critical confusion arises. These systems do not predict behavior in a causal sense. They perform probabilistic screening based on non-specific physiological signals. Nevertheless, public and political discourse began treating the system as if it were capable of “reading intentions.” This transposition is not a mere semantic detail; it completely alters the type of decisions the State begins to justify.

---

Behavioral prediction and the risk of operational false positives

For those working in public safety, the central point is not whether the technology is sophisticated. The point is the operational false positive. In high-flow environments like airports, major events, or borders, even a modest error rate carries grave consequences.

Anxiety, fear, grief, stress, fatigue, and personal experiences activate the Autonomous Nervous System in a manner similar to the activation caused by a real threat. Behavioral surveillance detects this activation. It does not distinguish the cause. This limitation is not technical; it is biological and cognitive.

When a system flags a risk based on elevated heart rates, facial micro-tensions, or respiratory changes, it is identifying a state, not an intention. The practical result is the multiplication of false positives: ordinary citizens treated as suspects due to normal physiological responses in stressful contexts.

Researchers critical of FAST warned of exactly this. In a busy airport, a system with 70% accuracy, applied to thousands of people per hour, produces more unjustified interruptions than useful interventions. The cost is not just reflected in lines or delays; it manifests as the erosion of institutional trust, expanded discretionary power, and the normalization of suspicion.

For the manager, this requires an uncomfortable but necessary question: what decision am I willing to authorize based on signals that cannot differentiate threat from distress?

If the answer is not clear, the technology stops being a support tool and becomes a liability.

---

Body language, physiological activation, and the limits of state inference

Behavioral surveillance usually relies on an implicit assumption: the idea that body signals, facial expressions, and physiological responses contain direct clues about future intentions. This assumption is fragile—not for a lack of sensors or computational capacity, but because it ignores a structural limit of human behavior.

Body language was not designed to communicate criminal intent. It precedes verbal language and is deeply associated with the body’s automatic regulation. Sweating, respiratory changes, facial micro-tensions, pupil dilation, and shifts in speech rhythm indicate activation, not meaning. The body responds before it interprets.

This point is decisive for law enforcement and managers. A system can identify that someone is under heavy stress, but it cannot determine if that stress stems from a fear of flying, a personal conflict, a recent loss, or a violent plan. All these situations produce similar physiological patterns. The system does not fail due to imprecision; it fails due to causal indistinction.

When one attempts to convert physiological activation into a threat indicator, an inferential leap occurs. This leap is not scientific; it is normative. It moves ambiguous data into a concrete state decision, such as a stop, detention, coercive interview, or expanded surveillance.

This framework is fundamental to dismantling the belief that more data produces more truth. In practice, it produces more decisional noise. The risk is not in observing behavior—something police work has always done. The risk is in automating inferences that an experienced human observer already knows cannot be made reliably.

When behavioral surveillance begins to guide operational decisions, it does not expand control; it redistributes uncertainty. And, frequently, it shifts that uncertainty onto the wrong citizen.

---

Operational decision-making guided by amplified uncertainty

In police practice, every decision is born under uncertainty. Stopping, detaining, interviewing, or releasing someone involves situated judgment, reading the context, and risk assessment. The promise of behavioral surveillance is to reduce this uncertainty through data. The actual effect is usually the opposite.

When physiological signals guide operational decisions, uncertainty does not disappear; it simply moves. Instead of being explicit in the officer’s judgment, it hides behind an algorithmic classification. The data creates an appearance of objectivity but does not solve the central problem: the causal ambiguity of human behavior.

For the manager, this creates a clear institutional risk. Decisions start being justified by systems whose logic cannot distinguish common emotional activation from a concrete threat. The frontline officer tends to trust the alert because it comes “from the system.” The system, in turn, is not held accountable for the decision’s impact. Responsibility becomes diluted.

This type of arrangement favors excessive stops, unnecessary detentions, and unjustified escalations. In high-flow environments, operational costs explode. Institutionally, the damage is deeper: it normalizes the idea that suspicion can precede any concrete evidence of wrongdoing.

Here, the operational false positive stops being a statistical error and becomes a governance problem.

---

Behavioral surveillance, governance, and institutional risk

The debate over behavioral surveillance is often shifted toward privacy. This is a legitimate concern, but insufficient. Even if data is anonymized and deleted after screening, the decisional effect has already occurred. The citizen has already been classified, stopped, or pulled from the flow.

The central problem is not data storage. It is the state use of fragile inferences to guide coercive action. This directly affects the governance of public safety.

Such systems broaden discretionary power without broadening the quality of the decision. They create technological dependency, reduce the agent’s critical autonomy, and make subsequent accountability difficult. When something goes wrong, the justification falls on the protocol or the system, never on the underlying premise.

For managers, the right question is not whether the technology is modern, but whether it improves the quality of decision-making under risk. If the answer is no, the system adds no strategic value. It merely shifts the problem to a less transparent technical layer.

Behavioral surveillance, when presented as a predictive solution, tends to produce more noise than control. Not out of bad faith, but because it attempts to answer a question that does not allow for an objective answer: what someone intends to do.

---

Operational Conclusion

What managers need to understand before authorizing use

Behavioral surveillance does not fail because it is imprecise. It fails because it is applied to a type of decision that does not permit reliable physiological inference. The error lies in the institutional expectations created around the system.

For public safety managers, certain guidelines must be made explicit:

1. Do not delegate decisions to ambiguous signals
   

   Physiological signals indicate activation, not intention. Using them as a basis for coercive action shifts uncertainty to a point that is less visible and harder to audit.

2. Do not confuse screening with evidence
   

   These systems can, at most, support broad screening. They do not sustain a stop, detention, interrogation, or any restrictive measure on their own.

3. Treat false positives as a structural cost
   

   A false positive is not an exception; it is an expected consequence. If the institutional model cannot absorb this cost, the system becomes operationally unviable.

4. Preserve the agent’s critical judgment
   

   The greater the blind trust in a technical alert, the lower the agent’s ability to contextualize, question, and halt wrong decisions.

5. Evaluate institutional impact, not just technological impact
   

   The criterion should not be innovation, but the actual effect on flow, legitimacy, accountability, and public trust.

Authorizing behavioral surveillance systems without these clear boundaries does not increase control. It increases decisional risk.

---

What this type of system can and cannot offer

What it can guide:

• Broad prioritization of attention in high-flow environments
• Exploratory support in studies and simulations
• Analytical training on the limits of behavioral reading
• Strategic discussion on risk management and screening

What it cannot guide:

• Inference of criminal intent
• Individual threat classification
• Isolated police stops
• Detention or interrogation
• Immediate coercive decisions
• Labeling individuals as suspects

Physiological activation is not intention.

Correlation is not causality.

Screening is not a decision.

This distinction is not academic. It is what separates the responsible use of technology from repeated institutional error.

---

For those who wish to go deeper…

This post is also available in pt_BR and pt_PT.