# What AI glasses cannot do, from the people who make them > An honest list of the limits. What it gets wrong, where it should never be trusted, and the situations where a person is still the right answer. Worth reading before you buy anything in this category. Published: 2026-08-19 Category: Everyday help Canonical: https://glimpseai.jp/blog/what-ai-glasses-cannot-do Every product page in this category is written by someone with a reason to be optimistic. This one is written by the same people and is about the other half, because the fastest way to lose trust in a tool is to be surprised by a limit nobody mentioned. We make [Glimpse](https://apps.apple.com/jp/app/glimpse-vision-ai/id6764219526?l=en-US), so treat this as interested rather than neutral. It is still the article I would want somebody to read before deciding whether this belongs in their life or their parent's. ## It can be confidently wrong This is the limit that matters most, because it does not look like a limit. A vision model rarely says it is unsure. It produces a fluent, specific answer at the same volume whether it is right or guessing, and the tone carries no information about which one is happening. People calibrate trust from confidence, which is a reasonable habit with humans and a bad one here. If you take one thing from this page, take that. ## Where it should never be the final word - **Medication.** Reading a label is useful. Deciding which pill to take is not a job for a camera. Use it as a second pair of eyes and keep a system that does not depend on it. - **Anything you are going to eat that came out of the ground.** Mushrooms above all. The margin for error is death and no model is close to good enough. - **Medical questions.** It can describe what a rash looks like. It cannot tell you what it is, and it has no way of knowing what it does not know about you. - **Legal and financial documents.** It will read a contract aloud and summarise it. A summary is not advice, and the clause that matters is often the one a summary drops. - **Money and prices at speed.** It can misread a digit. Check anything where a misplaced zero costs you. ## Physical limits that surprise people **Bad light.** A camera in a dim restaurant sees roughly what you see in a dim restaurant, which is not much. Answers get worse and the failure is quiet. **Movement.** Reading small print while walking rarely works. Stop for a second, the way you would to read anything. **Glare and glass.** Museum cases, laminated menus, phone screens in sunlight. Reflection defeats it fairly easily and the fix is usually to change angle. **Handwriting.** Printed text is largely solved. Handwriting, particularly cursive or a doctor's note, is a different problem and it is not solved. **Noise.** It has to hear the question. A crowded station platform is hard for the same reason it is hard for a person. ## Things it will not do by design It does not run all day watching everything. Sessions start when you start them. That is a deliberate decision about the people around you rather than a technical limit, and it means it is not a memory of your day that you can search later. It also does not know your life. It sees this moment. It does not know which of the two identical jackets is the one you wore yesterday, because it was not there yesterday. ## When a person is the right answer Worth saying plainly, because assistive technology is often sold as independence and sometimes gets used as a substitute for contact. If a decision matters and is reversible only at cost, ask a person. If the reason you are asking a camera is that asking a person feels like an imposition, that is worth noticing, and the honest version of this product is one that reduces small daily friction rather than one that replaces the pharmacist. ## Why we would rather you knew A tool you trust correctly is more useful than one you trust completely, because you keep using it in the places where it is strong instead of abandoning it after one confident error. Everything above is the honest shape of what this is. Within that shape it does a genuinely useful job many times a day, and outside it we would rather you did not rely on it. ## Common questions ### How often does it get things wrong? For clear printed text in decent light, rarely. For identification of objects, plants or anything ambiguous, often enough that you should check anything that matters. The problem is that the wrong answers sound exactly like the right ones. ### Can I rely on it for medication? As a second check on a label, yes. As the thing that decides which pill you take, no. Keep a system that works without it, such as a weekly pill organiser filled by a person. ### Does it work in the dark? Poorly, for the same reason your own eyes do. If you cannot read something in that light, it usually cannot either. ### Is it always recording? No. It looks during a session you started. It is not a background camera collecting your day, which also means there is nothing to search afterwards.