Skip to main content

Why Shopify’s AI is failing the accessibility test (and how to fix it)

A campaign graphic for #AltAtSource with a pale yellow background. The top half shows a screenshot of the Shopify image editor featuring a photo of female athletes at the 2017 World Athletics Championships. A large blue hand-drawn arrow points to a purple box titled "Suggestion," which displays the mediocre AI-generated alt text: "Athletes on a track with Toyota branding." At the bottom, large black serif text reads: "Shopify’s AI generated ALT text is meh!" The graphic highlights the contrast between high-stakes human achievement and overly simplistic automated descriptions.

Digital accessibility is often treated as a manual tax on content creators. The #AltAtSource campaign is here to change that by leveraging the metadata that already exists within our images.

I recently conducted a technical stress test. I asked a colleague to upload an iconic photograph of the 2017 World Athletics Championships to a Shopify account to see how their native AI tool handled the automated generation of alt text.

A screenshot of the Shopify image editing interface. On the left is a photograph of several female athletes in the middle of a race on an outdoor track. The athletes wear competitive running kits and bibs that include "Toyota" and "London 2017." On the right, an "Information" sidebar is visible. Under the "Alt text" field, a purple "Suggestion" box is highlighted, containing the AI-generated text: "Athletes on a track with Toyota branding." This comparison illustrates the difference between the rich detail of the actual event and the generic summary provided by the automated tool.
Orignal image: Erik van Leeuwen, attribution: Erik van Leeuwen (bron: Wikipedia)

An experiment in capability

I recognise that a professional merchant is unlikely to be selling a high-resolution press shot of the Women's 1500m final on their storefront. However, this image was chosen deliberately as an experiment in AI capability. It represents a "worst-case scenario" for automation: a complex, high-emotion scene where the true value lies in the context (who is winning, and where) rather than just the objects in the frame.

The result was a perfect illustration of the current gap in digital inclusion. Shopify’s AI offered the following suggestion:

"Athletes on a track with Toyota branding."

Technically, the AI isn't wrong. It identified the pixels correctly. But for a screen reader user, this is a massive loss of context. It prioritises a corporate logo over the human achievement of the athletes, the significance of the 1500m final, and the historical weight of the moment.

The stark comparison: pixels vs. context

To see what was possible, I ran the same image through an advanced AI (Gemini). Without any specific prompting other than a request for a description, it produced this:

"Six professional female athletes sprint toward the finish line during the Women's 1500m final… Faith Kipyegon (Kenya) celebrates with a raised fist as she crosses the line first, followed closely by Jenny Simpson (USA) and Sifan Hassan (Netherlands)"

I asked the AI how it knew these names. It confirmed that it had no access to the file’s internal metadata; instead, it identified the athletes and the event by cross-referencing the image against public records, such as the official Wikipedia entry for the London 2017 Women's 1500 metres final. When I asked if having embedded metadata would have helped, its response was the smoking gun for our campaign:

"Yes. Embedded metadata acts as a contextual anchor. It allows me to stop 'guessing' who people are and instead focus on describing the scene with 100% accuracy. It turns a pixel-hunt into a verified description."

Why AI needs a source of truth

This test perfectly demonstrates the gap in AI-generated alt text. Some systems do it well, others don’t — but they are all essentially guessing. Because current AI often prioritises recognisable patterns over human context, it sees a logo and assumes it is the subject, whilst ignoring the people in the frame.

The photographer who captured that finish line in 2017 already knew exactly who those women were. That information was recorded in the original metadata. If Shopify natively extracted this IPTC AltTextAccessibility or XMP dc:description data, merchants wouldn't have to choose between a generic guess and a manual typing marathon.

We could provide an authoritative, portable description that travels with the image from the photographer's camera to the end-user's screen reader.

The #AltAtSource mission: action, not invention

The #AltAtSource campaign isn't asking for a new invention. I am asking for the industry to use the tools it already has.

I’m seeing fantastic momentum from other leaders in the CMS space:

  1. Craft CMS is launching native metadata mapping functionality.
  2. WordPress 7.0 has this on the roadmap.
  3. Drupal is exploring a model that prioritises embedded metadata over AI fallbacks.

A call to lead

I have reached out to the Shopify community to propose this shift. While my recent campaign graphics have been deliberately abrasive to gain traction, the goal is entirely collaborative. I want to see world-class platforms like Shopify lead the way in world-class inclusion.

We need to stop treating accessibility as a manual tax and start treating it as a native property of digital media.

It is time to make alt text portable. It is time for #AltAtSource.

Article by Simon Leadbetter

The Accessibility Guy at Kindera

Simon Leadbetter
Next