The cultural bias problem that changed how we build SuperScout

When you upload a photo to SuperScout, something happens automatically: every image gets tagged with a set of keywords that make it instantly searchable. Upload a Victorian kitchen and you can search for “Belfast sink”, “original tiles”, or “exposed brick” the moment it lands. Upload a clifftop fortification and it comes back for “coastal”, “medieval”, “dramatic horizon”.

That automatic keywording is one of the things our customers mention most. It’s why a location that used to take 40 minutes to upload now takes under 5.

But getting there wasn’t straightforward. One specific problem forced us to rebuild something we thought was already working.

The bias we found

Early in development, we were using a well-known commercial vision API to generate the keywords. On standard European architecture it performed well. On interiors it was accurate. But when we started testing it across a wider range of locations (the kind of global, diverse portfolio that a working location scout actually builds), the cracks appeared.

The clearest example: an Indian living room (with Hindi characters visibly on the wall) was labelled as a Victorian English interior.

That isn’t a trivial error in a location context. A scout who uploaded that image expecting it to be findable as Indian décor, South Asian design, or simply a non-Western interior would simply not find it. It would surface in entirely the wrong searches, or not at all. For a platform that exists to help people find the right location for a specific brief, a keyword engine that can’t read what’s literally written on the wall isn’t just inaccurate. It’s a tool that works better for some locations than others.

Photo above: similar to the image that triggered the issue. An Indian living room a commercial vision model insisted on calling a Victorian English interior.

What we did about it

We stopped using the model, even though we had free credits that made it cost nothing to run. These were not the droids we were looking for.

That’s worth saying plainly: this wasn’t a cost decision. The commercial API was free for us at the time. We walked away from it purely because the output wasn’t good enough.

We spent several weeks testing alternatives: working through the extraction workflow, the search embedding generation, and the structured metadata output that sits behind every tagged location. We moved to an open-weights model via HuggingFace, which gave us better control over how the model was evaluated and allowed us to test across a far more diverse range of architectural styles, geographies, and visual contexts.

The goal was accuracy across the full range of locations our customers actually scout, not just the locations that look like they belong in a British period drama.

Why this matters for location work specifically

The house that triggered the issue wasn’t in India. It was in London, designed by an Indian architect, with an interior that looked and felt exactly like a classic Indian living room. It just happened to be in South West London.

That distinction matters. A production looking for an Indian-style interior doesn’t want to be told they need to go to India. They want to know that the right location might be twenty minutes from their studio. And a keyword engine that assumes Indian design means Indian geography will never surface that house for the right brief.

This is the broader problem with models that map style to country: people design and decorate in ways that cross every border. A scout’s database might include a Moroccan-styled riad conversion in Manchester, a mid-century American interior in Berlin, or a Japanese-influenced minimalist space in Edinburgh. Style and location are independent variables, and the keywording has to treat them that way.

This is also why we think it’s worth being transparent about it. The film industry is rightly cautious about where AI gets applied. Showing our working (including the mistakes we found and fixed) matters more to us than presenting a polished version of the story.

The CoSTAR recognition

Earlier this year, the CoSTAR Foresight Lab (part of an AHRC-funded initiative mapping AI adoption in the UK screen sector) published a case study on SuperScout. Their write-up specifically highlighted the responsible model selection process:

“SuperScout’s decision to test and reject models producing culturally biased outputs demonstrates practical responsible model selection, prioritising accuracy over convenience.”

That framing (accuracy over convenience) is exactly what we were trying to achieve. It’s good to see it recognised externally, and it’s the kind of external scrutiny we’re happy to be held to.

The full case study is worth reading if you’re interested in how location tech fits into the broader picture of AI adoption across the UK screen industry.

What this means in practice

If you’re a SuperScout customer, this work is invisible to you, which is the point. Every location you upload gets accurate, useful keywords regardless of where in the world it is or what architectural tradition it belongs to.

If you’re evaluating SuperScout, it’s worth knowing that the keywording engine has been tested across a genuinely diverse range of locations, not just the ones that are easiest to get right.