AI improving – Facial recognition is spreading across venues, airports, and public services as deep learning models push accuracy beyond 99% in controlled settings. But real-world conditions and biased training data still fuel false positives and false negatives—failures tha
When your face is scanned at an event entrance, it can feel like a convenience—until you think about what happens if the system gets it wrong.
At Madison Square Garden in New York. a facial recognition system scans visitors as they enter. including for major events such as NBA Finals games. The approach is becoming more common across the country: some big venues. including Citizens Bank Park in Philadelphia and Oracle Park in San Francisco. use facial recognition to offer visitors optional ticketless admission. Outside entertainment, the technology is also moving into daily life, from public buses to public buildings.
In aviation, the Transportation Security Administration has deployed the latest facial recognition technology at security checkpoints at numerous airports. The agency says the new system will be used in cities across the U.S. that are hosting FIFA World Cup 2026 soccer matches.
The push for broader use comes as the technology improves quickly—yet it carries a hard edge. Concerns about accuracy and bias have grown alongside adoption, and the stakes are not hypothetical.
Deep learning can be more than “better”—in controlled environments. advanced models have reached accuracy levels of more than 99% in settings such as cellphones. airports. and border checkpoints. That improvement is largely driven by advanced deep learning models trained on hundreds of millions of face images. and it is showing up in real deployments.
The core idea behind facial recognition is straightforward, even if the results can be unforgiving. The systems locate a face in an image or video frame. create a faceprint that catalogs salient features such as the shape of the face and landmark points including the eyes. nose. and mouth. and record the texture of the skin. They then compare that faceprint to records in a database—whether inside a smartphone or at a bank or hospital—to verify identity or grant access.
In the physical world, systems are faster and simpler than requiring people to show IDs. In the online world, they are easier than entering a login name and password. The technology also reduces the possibility of forgery or fraud compared with ID cards or passwords.
Several research projects have helped lift performance. FaceNet, a deep learning model developed by Google, upgraded recognition for faces that are partly covered or hidden in images. DeepFace. a landmark AI-powered facial recognition system developed by Facebook AI Research. aims for the same high level of verification shown by humans. NEC’s NeoFace, an AI-powered algorithm built into Mobile Fortify, a mobile facial recognition system used by U.S. Immigration and Customs Enforcement to identify people.
The hard part is that the real world rarely looks like a controlled test.
Poor lighting. difficult viewing angles. extreme facial expressions. concealment by face masks or sunglasses. and poor image quality can still hamper performance. These failures generally show up as two kinds of mistakes: false positives and false negatives. A false positive happens when a person is incorrectly matched to someone else in a database. A false negative occurs when an individual is not found in a database even though their image exists there.
In security and safety applications, false positives are the most dangerous. They can lead to wrongful accusations, discrimination, or detention. In 2025. a 50-year-old woman in Tennessee was arrested and put in jail for six months after an AI-powered facial recognition system incorrectly tied her to a North Dakota bank fraud investigation.
False negatives can be devastating too, just in a different direction. They may prompt authorities to deny services to people who qualify for them.
Some errors are driven by what the system learned in the first place. Accuracy can suffer when models are trained on data that does not reflect real-world demographics. A 2025 study showed that systems trained on public databases where people with darker skin tones are lacking lead to lower recognition accuracy. That kind of unintentional bias in training data may drive misidentification of women. people of color. and young and old people. One report found that facial recognition systems used by 42 U.S. government agencies falsely identified African American and Asian faces 10 to 100 times as often as white faces. in some cases leading to wrongful arrests.
Even without demographic gaps, performance can shift with the way people look day to day. Accuracy deteriorates when people wear heavy makeup. and for young children and old people. because their landmark features tend to change more quickly than adults of other ages. Efforts to balance datasets—by collecting more representative images across age. gender. and ethnicity—and frequently updating databases are aimed at improving accuracy and producing fairer results.
There are also technical tweaks before images are sent for matching. Changing brightness levels, for example, can improve accuracy. People squint in dark or very bright light. and advanced processing software can mimic this human trait to help facial recognition systems extract facial features from images more reliably.
The best systems also learn to cope with partial information.
Humans can identify someone even if part of their face is covered by sunglasses or a face mask. The brain assigns more significance to exposed details. and researchers are working on ways for facial recognition programs to do the same—reducing false positives and false negatives when cameras capture only part of a face.
Facial dynamics can matter as well. Recognizing a middle school friend you haven’t seen in years can be hard at first. but a smile can quickly improve recall. Researchers are developing a facial recognition method known as volumetric directional patterning. which captures subtle movements of facial muscles—including eyelid blinks—in consecutive frames of video. It tracks how facial landmarks shift over time and the context in which a face is being observed. with the aim of improving recognition accuracy.
Other researchers are creating more accurate AI-powered three-dimensional systems designed to capture precise geometry of a face. including contours of the eye socket. nose. and chin. Work like this could enable antispoofing techniques that prevent facial recognition systems from being fooled by fake faces generated by computers and their human operators.
Set aside the privacy and cybersecurity debates and lingering issues of bias, and one conclusion remains hard to escape: facial recognition technology is improving. And that improvement points toward fewer errors—and fewer of the serious consequences tied to those mistakes.
For now, though, accuracy gains are not uniform across conditions, demographics, or situations. The same systems that can identify more reliably in controlled environments are still exposed to the messy reality of cameras. crowds. lighting. and human variation—where an error is more than a technical glitch.
Vijayan Asari is a professor of electrical and computer engineering at the University of Dayton.
This article is republished from The Conversation under a Creative Commons license. Read the original article.
facial recognition deep learning bias accuracy false positives false negatives TSA FIFA World Cup 2026 Madison Square Garden ticketless admission University of Dayton Vijayan Asari ICE Mobile Fortify