AI Training Data Is Poisoning Design Trends. Here's How to Spot It

When a London-based design studio was commissioned to redesign a fintech app earlier this year, it noticed something peculiar. The client's initial brief, heavily influenced by AI-generated moodboards, requested a design that looked remarkably similar to the apps on the market. The culprit? Generative AI tools trained on datasets that prioritise frequency over quality, creating what the studio founder calls 'the great flattening of digital design'.

This phenomenon extends far beyond one studio's experience. As AI design tools become ubiquitous, their training data, often scraped from publicly available sources after 2022, is creating a feedback loop that amplifies certain aesthetic choices whilst marginalising others. The result is a homogenised design landscape where originality becomes increasingly difficult to achieve and even harder to recognise.

The Dataset Problem: What AI Actually Learns

Most generative AI tools used in design are trained on massive datasets compiled through web scraping, with Midjourney, DALL-E, and Stable Diffusion leading the charge. However, these datasets suffer from 'temporal bias'. The design samples in widely used training datasets serve as a snapshot of the aesthetic preferences from the mid-2010s.

This temporal clustering has tangible consequences. The majority of geometric sans-serif typography, gradient overlays, and specific colour palettes (particularly millennial pink and various purples) in AI-generated designs isn't coincidental. These elements dominated Dribbble and Behance during the peak scraping period, making them statistically overrepresented in training data. When designers use AI tools for inspiration or generation, they're unknowingly drawing from a well already contaminated by historical bias.

The Homogenisation Effect: When Everything Looks the Same

Today, you'll notice a similarity in how startups present themselves visually. The same rounded corners, the same illustration style (flat, friendly, vaguely humanoid figures), the same layout structures. Whilst some of this can be attributed to broader design trends, the acceleration coincides suspiciously with widespread AI tool adoption.

When you conduct informal research on startup websites launched recently, you are going to see a dramatic reduction in stylistic variance, with clustering around specific visual patterns consistent with popular AI training datasets. You are going to see unprecedented design similarities. It's not just that similar designs exist; it's that the diversity of approaches has measurably narrowed.

This convergence isn't merely aesthetic. It has functional implications. When interfaces become too similar, users struggle to differentiate between services, reducing brand recognition and potentially harming usability through learned patterns that don't transfer across contexts.

How to Spot AI-Influenced Design Pollution

Determining if a design is affected by AI training bias involves fostering 'dataset literacy'. Several red flags can alert discerning designers. First, examine colour choices. If a design relies heavily on specific gradient combinations (particularly blue-to-purple or pink-to-orange), it may be pulling from overrepresented training samples. Second, examine illustration styles. The majority of geometric, minimalist human figures with no facial features suggest AI influence or indirect inspiration from AI-generated work.

Typography offers another tell. An overreliance on specific typeface categories, particularly geometric sans-serifs like Circular or Proxima Nova alternatives, often indicates AI-assisted design. Whilst these are quality typefaces, their statistical overrepresentation in training data means they appear disproportionately in AI outputs. Layout structures provide the final clue. AI-influenced designs often feature centred hero sections, three-column feature grids, and specific spacing ratios that reflect the most common patterns in training datasets rather than purposeful design decisions.

Designing Beyond the Dataset

Escaping AI homogenisation requires intentional effort. Start by diversifying your inspiration sources beyond platforms that feed AI training datasets. Explore design archives, visit physical exhibitions at institutions like the Design Museum in London, and study work from underrepresented geographic regions and time periods. The Wellcome Collection's digital archive, for instance, offers rich visual material that exists outside typical AI training data.

When using AI tools, treat them as starting points for subversion rather than endpoints for execution. If an AI suggests a particular direction, ask why and consider deliberately moving in the opposite direction. Establish design principles before engaging with AI tools, ensuring your decisions remain anchored in strategic thinking rather than statistical probability.

Finally, advocate for better training data. As designers, we shape the future dataset through what we publish and share. By creating and promoting diverse, culturally specific, and up-to-date work, we can help correct the biases that plague AI systems. The future of design depends not on rejecting AI tools but on ensuring they learn from a richer, more representative sample of human creativity.