Can AI Identify Criminals by Their Face?

Yes, AI can be trained to pick out faces that human raters label as criminal, but it cannot detect criminality itself—it only learns statistical patterns from the labels it is given, and those patterns are heavily influenced by bias. In short, the technology exists, but it is scientifically controversial and legally and ethically fraught.

Where the Idea Came From

The idea that facial features might reveal character is old—physiognomy, the pseudoscience of judging personality from faces, was popular in the 19th century and has been repeatedly debunked. Modern AI research revived it in 2016 when a paper claimed that machine learning could distinguish “criminal” from “non-criminal” faces with about 90% accuracy. The study used mugshots and ordinary photos, and it was widely criticized for methodological flaws: the photos came from different sources, lighting, and angles, and the “criminal” group was not matched to the control group by age, ethnicity, or expression. Later replications found far lower accuracy, often close to chance.

How the AI Actually Works

Most face-classification systems use deep neural networks, specifically convolutional neural networks (CNNs). These networks are trained on thousands of labeled images. During training, the network adjusts internal numerical weights to minimize the difference between its output and the human-provided label. For a “criminal” classifier, the labels come from human decisions—usually whether a person was convicted or arrested. The network does not learn a universal “criminal look”; it learns the statistical patterns that separate the two groups in that particular dataset. That could include skin tone, facial hair, age, or even the background of the photo.

When a new face is input, the network outputs a probability score between 0 and 1. A threshold (often 0.5) decides the label. The entire process is a black box: even the engineers usually cannot say exactly which features drove the decision.

Why It Fails in the Real World

The most serious problem is bias. Training data for crime often comes from mugshots or arrest records, which in many countries disproportionately include minority groups and lower-income neighborhoods. The AI then learns those correlations as if they were facial markers of criminality. A 2019 study by the National Institute of Standards and Technology (NIST) tested dozens of face-recognition algorithms and found that many had higher false-positive rates for Black and Asian faces compared to white faces. That means a system used to flag suspects would more often wrongly single out people of color.

Another issue is the fundamental assumption that criminality is visible. Crime is a legal and social category, not a biological trait. Two people can commit the same act, but only one is arrested, charged, and convicted—so the label depends on policing practices, not just behavior. Even the most accurate algorithm can only reproduce the biases in its training data.

How to Evaluate Such a System

If you are considering a face-analysis tool, you should ask for more than a single accuracy number. Key factors include:

• Dataset composition – Were the training images matched by age, ethnicity, and photo conditions?
• False-positive rate – How often does it label an innocent person as criminal? This is more important than overall accuracy.
• Test on independent data – Did the system perform well on faces it never saw during training?
• Explainability – Can the developers say why a particular face was flagged?
• Human oversight – Is there a process for a trained person to review the AI’s output?

In practice, no such system has been validated for real-world criminal identification. Several researchers have called for a ban on using facial analysis for this purpose until it can be proven fair and accurate.

Legal and Ethical Concerns

Using AI to flag potential criminals raises serious civil liberties issues. A false positive could lead to a wrongful stop, arrest, or worse. In many jurisdictions, there are no clear laws governing police use of facial recognition, and some cities have banned it outright. The European Union’s proposed AI Act would classify remote biometric identification in public spaces as high-risk, requiring strict oversight. Even where legal, the use of such tools can erode public trust and reinforce discriminatory policing.

There is also the problem of “automation bias”—people tend to trust computer outputs more than they should. If an officer sees an AI score of 0.9, they may treat that as strong evidence, even though the score is just a statistical correlation, not a fact.

Alternatives and the Future

Rather than trying to read faces, law enforcement uses other AI tools with clearer value: matching a suspect’s photo to a database of known individuals (face recognition), analyzing patterns in crime data to allocate patrols, or scanning for specific objects in footage. These still have bias risks, but they do not assume guilt from appearance.

Some researchers are exploring whether certain physiological markers (like thermal patterns or micro-expressions) correlate with deception, but results are mixed and not ready for court. The consensus among computer vision scientists is that predicting criminality from a static face is not scientifically sound. As one 2020 review put it, “there is no evidence that facial appearance can reliably predict criminal behavior.”

Frequently Asked Questions

Can AI really tell if someone is a criminal?

No. AI can only label a face based on patterns in its training data. If the data includes mugshots, the AI learns to associate certain visual features with the label “criminal,” but that is not the same as detecting criminality. Studies claiming high accuracy have not been replicated, and the technology is widely considered unreliable for this purpose.

Why do some studies report 90% accuracy?

Those studies often use small, biased datasets and fail to control for variables like age, ethnicity, and photo quality. When independent researchers test the same methods on more diverse data, accuracy drops significantly, sometimes to near chance. High numbers in a lab do not translate to real-world performance.

Is face recognition the same as criminal detection?

No. Face recognition matches a face to a known identity (e.g., “this is John Smith”). Criminal detection claims to infer a trait (e.g., “this person is likely a criminal”). Face recognition is widely used and has legitimate uses; criminal detection is not scientifically validated and is generally discouraged.

What are the main risks of using AI to pick criminal faces?

The biggest risks are bias and false positives. Because training data often over-represents certain groups, the AI may disproportionately flag innocent people from those groups. A false positive can lead to wrongful stops or arrests. There is also the risk of over-reliance on a flawed tool.

Could this technology become reliable in the future?

It is unlikely, because the core assumption—that criminality has a visible facial signature—is not supported by science. Even with better data, the label “criminal” depends on social and legal factors that are not biological. Most experts believe the approach is a dead end.

Related

• Embodied AI at Work: Robots, Uses, and Limits
• Where Winds Meet: Seasonal Challenges and Tracking
• Secure AI Agents That Use Third-Party Tools