We Sent 400 Visitors on a Date With a Van Gogh. Here’s What We Learned

The Van Gogh Museum built an AI feature that matched people to a painting and told them a personal story about why they belonged together. Visitors loved it. Then the team had a decision to make.

At a recent Museum AI Summit, two people from the Van Gogh Museum in Amsterdam — curator of education Jolein van Kregten and digital education product manager Fenne de Beer — walked through a pilot they had clearly poured themselves into. It’s a good story about building with AI inside a museum. It doesn’t end where you’d expect.

Matching a visitor to a Van Gogh painting

The premise started with a frustration any gallery educator will recognise. People wander through rooms. They glance, they photograph, they move on. They don’t really look. And looking — slowly, with some context — is the thing that turns a painting from wallpaper into something you remember.

Van Kregten describes herself as a bridge builder, someone whose job is to connect people to art. Van Gogh makes that job easier than most. He has, as she put it, the status of a pop star. People want to get close to him, and they tend to see themselves in his story: the strained relationship with his parents, the long search for meaningful work, the messy love life. Add famously colourful, recognisable paintings on top of that, and there’s a lot to work with.

So the question became: what if you could match a visitor to a specific artwork based on who they are — and then tell them a personal story about why the two of you belong together?

The idea had sat in a drawer for years because it was simply impractical. Doing it by hand meant paper questionnaires, a guide interpreting one visitor’s answers, and a bespoke mini-tour assembled on the spot, all in a reasonable amount of time. Not doable. Generative AI is what finally made it look possible. With a sponsor (Samsung), a software partner (Newz), and the museum’s existing audio guide to build on, the team had what they needed.

How the “date” actually worked

The team started with content, not technology. They pulled out themes that ran through Van Gogh’s life and work and that would still land with a modern visitor — nature, religion and spirituality, leisure, mental health — and built a questionnaire on top of them.

In the final version, a visitor answered nine questions drawn from three randomly selected themes. How do you relax? Where do you go on holiday? How do you react to a new challenge? Ordinary questions, designed to harvest something real about the person standing in the gallery.

The recipe that fed the AI had three ingredients: personal details from the visitor, verified facts about the artwork and the artist, and a quote pulled from one of Van Gogh’s letters to add emotional weight. That mix was enough for the model to propose a match and write a short, personal story tying the visitor’s answers to the work in front of them.

De Beer, who tested it herself, was matched with Church in Nuenen — a painting she says she wouldn’t have picked. She had talked in her answers about new ways of seeing, about endings leading to beginnings, about processing emotions on long walks by the sea. The story drew those threads into the painting, and she started noticing transition and quiet change in a work she’d otherwise have walked past. It did, she admitted, feel like going on a date.

The hard part: staying truthful without being boring

This is where it got interesting for anyone building with AI, inside a museum or not.

The Van Gogh Museum is a knowledge institution. A great deal has been written about Van Gogh online, and plenty of it is wrong, especially when it comes to his inner life. Letting a model roam the open web and improvise about his feelings was a non-starter.

So the team locked the model down, restricting it to sources they trusted, such as their own catalogue entries. It worked — the output was accurate. It was also, in their own honest words, a bit boring and straightforward. For this pilot, the team found it difficult to combine the source restrictions they needed with the tone and empathy they wanted. They spent the project working through that tension.

What the two-week pilot found

Rather than guess, they tested. Over two weeks, with small tweaks between week one and week two, they offered the feature to people who’d already taken the normal audio tour. Around 400 visitors went on a date and then filled in a short survey with both open and closed questions, giving the team numbers and narrative.

They analysed the cold metrics first — drop-off rates, how long the questionnaire took, general usage — before looking at the softer stuff about whether people enjoyed it and felt closer to the art.

People did enjoy it. They especially liked getting a better sense of what Vincent was feeling and thinking, woven together with their own answers. Around 70% called it fun and engaging.

One result did not meet the team’s hopes: about 55% of respondents said they felt a deeper connection to the painting. That shows some visitors reported the intended connection, but the team judged the overall value insufficient to continue the service.

The survey gave the team two different signals. Enjoyment showed that the encounter could appeal to visitors; the connection question spoke more directly to the educational ambition behind it. Both mattered, but neither settled the decision on its own. The museum also had to consider the work required to keep producing trustworthy stories and the questions visitors raised about their data. Seen together, those factors explain why an appealing experience could still fall short of the case needed for a continuing service.

Assessing desirability, feasibility, viability and responsibility

To decide whether the pilot deserved a future, the team used a product-management staple: the DFV model. Its three classic dimensions are desirability, feasibility and viability, with product innovation sitting where they overlap. The Van Gogh Museum added a fourth: responsibility.

Here’s roughly how the pilot scored.

Desirability was decent. About 70% of visitors called it fun, but only around 55% felt a deeper connection to the artwork afterwards — fine, not transformative.

Feasibility exposed the tension the team encountered: it needed trustworthy material but found its more tightly controlled stories less surprising.

Viability was the quiet problem. The team also had to account for the pilot’s prospective running costs as well as its build costs. Generating content, maintaining systems, monitoring output and refreshing the experience are all ongoing costs, which means the value to visitors has to be high enough to justify them indefinitely.

A separate case offers reported figures for the running costs of London Transport Museum’s AI guide, although its scope differs from this pilot.

Responsibility raised flags too. Visitors had questions about their data, and several said in open feedback that they’d expected something more educational and less like a gimmick.

Why the museum chose not to continue

Put it all together and nothing scored exceptionally well. And here’s the part most teams wouldn’t have reached: the Van Gogh Museum decided not to move forward.

Sit with that for a second. A novel feature. A big-name sponsor. Hundreds of testers who genuinely enjoyed it. Seven in ten calling it fun. Plenty of organisations would have read those numbers, declared victory, and shipped. Many would argue a 70% approval score is success.

The team saw it differently. They could polish the stories, but better output would never solve the recurring costs or the unresolved privacy questions. And a feature that’s merely fun — mildly connective, expensive to maintain, slightly unclear on data — isn’t good enough to carry the name of one of the most visited art museums in the world. So they killed it.

Van Kregten didn’t hide how that felt. “Heartbroken,” she said. Twice.

That’s the real story here, and it’s rarer than any clever use of AI. The discipline to look at a result most people would call a win and say this is good, and good isn’t good enough — that’s exactly what you’d want from an institution whose whole job is protecting how a complicated, much-misread artist is understood.

Their parting advice was practical: if you’re thinking about building something like this, bring a model like DFV in early, before you fall in love with the idea, so you’re backing things that are sustainable rather than just exciting.

A different visitor need and scale shaped Qatar Museums’ connected AI tour.

The date was lovely. They just weren’t willing to call it love when it wasn’t.

–

This article is based on a presentation by Jolein van Kregten and Fenne de Beer, Van Gogh Museum at the Museum AI Summit on May 27 2026.