How AI image scanning for trademarks works, and its limits

Image scanners for trademarks are useful, and they are often misunderstood. People tend to treat the result as a verdict, either "clean" or "infringing". It is neither. It is a guess about what a picture contains. Knowing how that guess is made helps you decide how much weight to give it. This is general information, not legal advice.

What an image model actually does

A scanner looks at a product photo and compares what it sees with patterns it has learned from many labelled images. Typically that means three things:

• Recognising shapes and marks. A swoosh, a crest, a cartoon mouse's silhouette, a repeating monogram pattern.
• Reading text. Brand names printed on a shirt, a slogan on a mug, a team name on a jersey.
• Judging similarity. How closely a design resembles something it knows, and with what confidence.

The output is a likelihood, such as "this looks like a licensed character," not a finding of fact. The model has no idea who made the product, who owns the design, or what your supplier agreed to.

What it cannot know

The questions that decide whether a listing is a problem are mostly outside the photo:

• Whether you have permission. A genuine licensed product and a knockoff can look identical in an image.
• Whether the mark is registered, and where. Trademark rights are tied to countries and to categories of goods. A name can be protected for clothing in one place and not at all in another.
• Whether your use would cause confusion. Courts look at context, such as how the mark is used, what is being sold and what buyers would assume. A picture can't capture that.
• Which kind of right is involved. Trademarks, copyright, design rights and trade dress are different things with different rules. A character can raise several at once. Copyright expiring on an old work, for example, doesn't automatically end trademark rights in the character.
• Exceptions such as parody or nominative use. These vary by country and depend heavily on the facts.

A scanner can tell you a design resembles something protected. Whether that matters legally is a different question, and one for a lawyer.

Why it gets things wrong in both directions

Misses (false negatives) happen when the model doesn't recognise what it's looking at. Common causes:

• Small or partial logos, such as a tag at the edge of the frame or a print cropped by the photo.
• Unusual angles, heavy shadows, low-resolution supplier images.
• Stylised or redrawn versions of a character that a human would still spot immediately.
• Newer, regional or niche brands that weren't well represented in the training data.
• Marks that appear only in the title or description, or on the back of the item and not in the photo.

False alarms (false positives) happen when something merely resembles a known mark. Stripes, stars, check-mark shapes, common words and generic animal outlines are everywhere. A model tuned to be cautious will flag some of them. It may also flag a genuine, authorised product, or a plain item that happens to share a colour scheme with a famous brand.

Tuning a model is always a trade-off. Make it stricter and you get more false alarms. Make it looser and you get more misses. No setting removes both.

Using a scanner without trusting it blindly

The sensible way to use any scanner, including ours, is as triage. It sorts a large catalogue into "look at this first" and "probably fine," which is useful if you import hundreds of products. It is not a substitute for judgement. A few habits help:

1. Review flags yourself. Look at the image, the title and the description. Ask what a buyer would think they were getting.
2. Spot-check the unflagged. Skim a sample of "clear" products, especially anything themed around films, sports, games or fashion.
3. Search the official databases. The USPTO, EUIPO and WIPO's Global Brand Database are free to search. They show what is registered and in which jurisdictions.
4. Ask suppliers for proof. If a product claims to be licensed, request the licence or authorisation in writing. "It's what everyone sells" isn't evidence.
5. Unpublish rather than delete. Setting a questionable product to draft keeps your records and lets you revisit it once you have an answer.
6. Keep notes. A short record of what you checked and why you decided as you did is useful if a complaint ever arrives.

This matters because the consequences land on you, not on the tool. Shopify acts on complaints from rights-holders. It can remove listings and, in more serious or repeated cases, restrict or close a store, and funds held through Shopify Payments can be affected. The details change, so read Shopify's current Legal pages and Help Center rather than relying on a blog post, this one included.

When to stop and ask a professional

If a product is central to your business, if you've received a complaint, or if you're unsure whether a design is a parody, a tribute or an infringement, talk to an intellectual-property lawyer in the relevant country. Law differs between countries, and most of the examples here are from the US. Your sales market, your supplier's location and the rights-holder's registrations can each change the answer.

A scanner such as MarkSentry can reduce how often a risky product slips through unnoticed. It can't tell you what is lawful. Treat its flags as prompts to look closer and keep making the final call yourself.