Artificial intelligence (AI)-powered smart eyeglasses, used as an assistive tool for individuals with low vision or blindness, work well for identifying common objects but fall short in tasks involving colours and coin money, as shown in a pilot feasibility study.
The smart eyeglasses had an accuracy of 99 percent for identifying single common objects and 83 percent for telling whether an object was positioned vertically (ie, top, middle, or bottom shelf) or horizontally (ie, left, right, or front), reported principal investigator Dr Carol Shields and her co-investigators from Thomas Jefferson University, Philadelphia, Pennsylvania, US. [JAMA Ophthalmol 2026;doi:10.1001/jamaophthalmol.2026.2185]
For object counting, the accuracy was 50 percent. The device performed well with larger objects but less reliably with smaller objects and an increasing number of objects. Accuracy was 73 percent for counting pencils (100 percent, 98 percent, 60 percent, and 32 percent for counting one, two, five, and 10 pencils, respectively) and 27 percent for counting rice grains (78 percent, 24 percent, 6 percent, and 0 percent for counting one, 20, 50, and 100 grains, respectively).
For colour discrimination, the device had reduced accuracy at 64 percent. Colour discrimination accuracy was highest for neutrals, including black (92 percent), white (96 percent), and brown (74 percent), but markedly dropped for green (50 percent), blue (6 percent), and violet (0 percent).
In terms of reading performance, accuracy was 59 percent for standard text, 88 percent for neat cursive handwriting, and 93 percent for children’s books. The device performed less accurately for reading medication labels (36 percent) than it did for food labels (82 percent), as noted by Medina and colleagues.
“In the medication trials, the AI eyeglasses had a propensity to state that they could not provide medical advice rather than read what was on the label,” they said.
With regard to money, the device performed reliably for identifying paper money but poorly for identifying coins, with accuracy rates of 91 percent and 2 percent, respectively.
“This study can provide a basis for the applicability of AI eyeglasses to the field of ophthalmology and improve on the understanding of how this technology can assist patients who experience low vision and blindness,” according to Shields and colleagues.
“Users of AI smart eyeglasses should be aware of these devices’ current limitations, which might improve as technology evolves in this field,” they added.
Compelling with caveats
The pilot feasibility study is a compelling demonstration of how smart glasses integrated with a large language AI model can assist patients with visual impairment, noted Dr Benjamin Young from Oregon Health & Science University in Portland and Dr Peter Zhao from the University of Michigan in Ann Arbor in an accompanying editorial. [JAMA Ophthalmol 2026;doi:10.1001/jamaophthalmol.2026.2409]
“The findings represent a potentially transformative advance in assistive technology for patients with low vision or blindness,” they added.
However, Young and Zhao highlighted several limitations to the study. First, the investigators acted as study participants, and the device was evaluated based on task-specific accuracy rather than real-world function.
“[The investigators] did not test whether patients with true low or no vision may perceive enough to know when to prompt the glasses or to orient the camera appropriately. Further, there was no detailed description of the types of errors the AI made, which could range from inconsequential to dangerous,” according to Young and Zhao.
“One unexpected finding was poor colour discrimination with green, blue, and violet. It was unclear whether this reflected hardware or software limitations,” they added.
Young and Zhao also raised privacy concerns regarding the use of AI smart eyeglasses, saying that individuals interacting with users of the device may be unaware they are being recorded and have no control over how their images are being processed and used.
“Put another way, physicians need to be aware that patients wearing smart eyeglasses at their appointments may be recording or transmitting audio and video of the entire visit. Commercial platforms may not fully protect patient privacy or the privacy of bystanders, and clinicians must clearly understand this evolving legal framework before supporting the use of any such system,” they said.
On a hopeful note, Young and Zhao concluded that patients living with low vision or blindness stand to benefit even in the absence of a perfect device. “These patients are navigating a world designed for the sighted, sometimes without the tools they need to do so safely and independently. If the promise of wearable AI can be validated, refined, and made equitably accessible in a manner with strong ethical guardrails, it has the potential to be among the most meaningful advances for those patients for whom ophthalmologists have fallen short.”
Study overview
The AI smart eyeglasses evaluated in the current study is a commercially available device that can capture images and video via a 12-megapixel ultrawide single camera, record audio via a 5-microphone array (two each in the right and left arms, and one near the nasal bridge), and respond to commands through a discreet open ear audio device with adaptive volume. Users can also listen to music, talk and text via handsfree communication, and store photos and videos. This device can be fitted with single and progressive prescription lenses in either clear, blue-violet light filtered, tinted, or transition options.
Six investigators served as participants (mean age 30 years, 50 percent female, mean height 67 inches, men age at English acquisition 3 years). The participants used voice commands to prompt the smart eyeglasses to execute tasks.
Tasks included single object identification, colour discrimination, directionality, big and small object counting, reading (medication labels, food labels, handwriting, children’s books), and paper and coin money identification and counting tasks. Study materials included pencils, pen, laptop, jacket, shirt, coffee mug, fork, knife, spoon, banana, apple, orange, coloured pencils, rice grains, over-the-counter medications, foods with nutrition label, handwriting samples, children’s books, and US currency.
The participants sat at a white tabletop with a white background for the tasks, except for vertical directionality tasks, and completed at least five trials of each task. The primary outcome measure was accuracy defined as percentage of correct responses for each task.