Think-Reason-Learn spots AI slop in web design, and says why

AI coding tools build web pages that look alike. Designers call them AI slop. Think-Reason-Learn can tell an AI page from a page a person made. It gets 99.6 of 100 pairs right, where a coin flip gets 50. It also shows which yes-or-no checks gave the AI page away. A detector that only looks at screenshots is as accurate, but it cannot explain itself. Top AI models asked to judge the pages get about 75 of 100. These are early results from experiments at Vela Partners.

AI coding tools now build a full website from one line of instructions. With no design direction, their pages converge on one look. A centred headline, three feature cards, a gradient button and emoji icons. Platforms want to filter these pages. Design teams want to grade them. The people who build the tools want a signal to improve on.

Think-Reason-Learn is an open-source library that predicts outcomes from unstructured and multimodal data, such as text, images and video.¹ It works with questions a person can read. We asked it two things. Can it tell an AI page from a human-made site? And can it say why?

We collected 189 pages that four AI coding tools built from 50 one-line briefs, such as "the website of a neighbourhood sourdough bakery". We also collected 982 sites made by people, from Awwwards, a web design award site.² We kept a quarter of each set hidden for the final test. Then we compared three ways to tell the two groups apart.

One caveat up front. The AI pages were built with no design guidance and no images or video to use. That makes this the easy version of the task. The experiments are also small. Image and video models are costly to run, so we capped the number of pages. Treat the results as directional guidance on what Think-Reason-Learn can do. The section on better-designed pages says what the test does not show.

What gives an AI page away?

AI pages are static. Everything is in place on the first frame. Human-made award sites open with a short scene. They animate many elements. They use motion and video to carry an idea. Figure 1 shows the six checks where the two groups differ most.

Part of this gap is a limit of the test, not a sign of who made the page. The AI tools got only a one-line brief. They had no video, photos or brand assets to use. Loading screens usually exist to load heavy media. So some of the strongest checks may measure missing assets as much as AI authorship.

Why does a readable answer matter?

A readable answer tells a designer what to fix. On this test, it costs nothing in accuracy. One check alone does most of the job. It asks whether the page loads with no loading screen, pop-up or gate. A detector built on that one check gets 93 of 100 held-out pairs right, where a coin flip gets 50. The check for a video on the page gets 90 on its own. The five best checks from the training pages, counted equally, get 95. All 386 checks together get 99.6, with each check weighted by how much it helped in training.

A detector that gives no reasons reaches the same score. So readable checks give up nothing here. What they add is a diagnosis. A flagged page comes with the checks that flagged it. For example, nothing moves on load and there is no video. One warning. An AI tool that copies these traits could slip past the detector without designing any better. The checks are a diagnosis, not a recipe.

How accurately does Think-Reason-Learn spot AI pages?

The learned checks get 99.6 of 100 held-out pairs right. An image classifier gets 99.5. Both keep that score on 24 pages from three AI tools that were not in training (Figure 2). Top AI models asked to judge do much worse. OpenAI's GPT-6.1 Sol gets 75 of 100 and Moonshot AI's Kimi K3.1 Pro gets 73. An open-source design rulebook, applied by Google's Gemini models, scores 44 and 33. That is no better than a coin flip, and those samples are small.^{3, 4}

The image classifier is a strong comparison. It turns three screenshots of each page into a list of numbers with OpenAI's CLIP, an image model.⁵ Earlier work found that CLIP spots AI-made images well, even from new tools.⁶ The classifier then learns which numbers point to AI, in the same way the checks are weighted. It is free to run once the screenshots exist. The checks cost about one cent a page.⁷ But the classifier's numbers mean nothing to a person. Both detectors are near the ceiling of this test, so the test cannot say which one is stronger.

Pages from AI tools the detectors never saw. Scores are human-made sites ranked above AI pages, out of 100 pairs. A coin flip scores 50. "Caught" uses an alarm level that wrongly flags 5 in 100 of the held-out human-made sites. MiniMax's M3 stopped partway through six of its ten pages, so four could be tested. With so few pages, a perfect score is encouraging, not conclusive.

How does Think-Reason-Learn judge a web page?

In three steps. First, a vision model describes the page. Google's Gemini 3.8 Flash watches a recording of the page being scrolled.⁸ It writes a structured description of the layout, type, colour, imagery and motion. Second, a decision model answers 386 yes-or-no checks about that description. The model is TypeSafe AI's Jev.⁹ It answers a fixed question with a probability, not with text. Of the checks, 307 come from two open-source design rulebooks.¹⁰ Think-Reason-Learn wrote the other 79 from training pages only. Third, a standard scoring formula, a logistic regression, learns from the labelled pages how much each answer matters. No step asks a model for an overall verdict. Every answer does depend on one model's description, though.

Splitting the judgement into checks makes the result readable. Earlier work shows that checklists make model judgements agree better with people.¹¹ It also shows that predicting through concepts a person can read keeps a model honest.¹² The difference here is who decides which checks matter. The labelled pages decide, rather than a judge.

Would the detectors catch AI pages with better design?

Nobody has tested this yet. The pages here came from one-line briefs with no design guidance and no assets. Any method that learns from examples finds them easy. Pages built with design direction, or with a design skill, will be harder to catch. On those pages the two detectors may start to differ. Three more limits apply. Both detectors trained on the same award-gallery sites, so either may flag a plain human-made page with no motion. All human-made sites come from one gallery. And the checks are answered from descriptions that one model wrote. Larger experiments are capped by the cost of image and video models, so these results are directional.

What can you do with this today?

Check your own landing page against the six questions in Figure 1. That is what a reviewer would do with the flagged checks. Build the same kind of detector for your own pages with Think-Reason-Learn. It turns labelled examples into readable checks and weights. A ready web design decision model, next to the decision models already in the Think-Reason-Learn app, could be worth testing. Releasing the checks, the generated pages and the evaluation code with a paper could also be worth doing. Send results and questions to benchmark@vela.partners.