AI limitations
What HiArt cannot tell you.
A HiArt report is a structured opinion produced by a vision model. Knowing where it stops being useful is part of using it well.
Last updated 4 August 2026
It is a model, not an audience
HiArt estimates how a visual may be perceived. It has never met your audience, has no access to your analytics, and cannot run a test. It will not tell you that a visual will get more engagement, more clicks, more conversions or more reach — and any report that appeared to promise that would be wrong.
Treat the score as a structured second opinion that makes your own judgement sharper, not as a verdict that replaces it.
It does not judge people
When a person appears in a visual, HiArt is instructed to comment only on framing, lighting, expression as visibly presented, styling, composition, hierarchy and brand context.
It will not identify anyone, rate anyone's appearance, infer race, religion, sexuality, health, disability or political affiliation, diagnose personality or mental health, or describe a person as high or low value. If you ask it to, it should refuse.
It can be wrong
Vision models misread small text, unusual layouts, cultural references and context that sits outside the frame. Every report lists the assumptions the model had to make and a confidence figure — read those before acting on the recommendations.
Two runs on the same image can differ. That is a property of the model, not a bug in your work.
It has taste, and taste is a bias
Models trained on a lot of design work tend to reward convention. HiArt is explicitly instructed not to recommend minimalism by default and not to strip personality in order to raise a professionalism score — but a distinctive, deliberately awkward or culturally specific choice may still be scored down simply for being uncommon.
If the model tells you to remove the thing that makes the work yours, that is a case where you should probably disagree with it.
Accessibility notes are a starting point
HiArt flags likely contrast and legibility problems by eye. It does not compute exact contrast ratios from your source file. Use a proper contrast checker before shipping anything where accessibility compliance matters.
Which model is used
Reviews are produced by Google's Gemini vision models. The specific model is configurable, so the exact version behind a report can change as better models ship.