Default drift

When interfaces, copy and flows are generated, every detail nobody specified is filled with what the model has seen most often. That is what generation is for, and much of the time it is fine. The pattern starts when defaults are accepted without being examined: the layout every generated product shares, the onboarding that asks for the same permissions, the copy that sounds like everyone else’s.

The convergence comes from the default, not the person. When writers worked with an instruction-tuned model, their essays grew more alike, and the effect was ‘mainly attributable’ to the text the model contributed; what the writers wrote themselves ‘remains unaffected’. People working with ChatGPT also ‘felt less responsible for the ideas they generated’, which is where this pattern hands over to Unowned decision.

Some defaults carry more than taste. In a test of 80 coding tasks across more than 100 models, generated code was insecure in 45% of cases, and left cross-site scripting open 86% of the time. Three commercial models produced 300 interfaces with 541 semantic accessibility faults between them, and a separate benchmark found generated HTML ‘frequently fail to meet full accessibility compliance without further prompting’: the safe version was available, and nothing asked for it. A default for consent, payment or safety ships as readily as a default for a button colour, and it looks finished, so nobody remembers it was a choice.

The pull towards sameness is measurable in creative work. In a study of nearly 300 writers, those given story ideas from generative AI wrote stories that judges rated more creative, but the stories became more similar to one another: better individually, narrower collectively.

The first direct evidence from generated products arrived in 2026. Researchers embedded 73 promotional websites built by graduate students with Lovable and found that 88% fell into six visual clusters. The makers could not see it: their sense of control over the result, their satisfaction and their judgement of its quality bore no relation to how typical the site was. The authors call this ‘authored ignorance’. It is one cohort using one tool, and a preprint; no study yet shows the same convergence across shipped commercial products.

Researchers studying how people who do not code build websites with language models map the points where homogenisation enters, find that ‘the push for frictionless generation can exacerbate’ it, and propose ‘productive friction’ as the counter. NIST’s profile for generative AI lists ‘harmful bias and homogenization’ among the risks a deployer should manage. The friction they describe is the third design move above: generate the alternative, and compare, before accepting the first plausible answer.