I was working on client team headshots in August 2025. Ran the first photo through my usual background removal tool — the same one I’d used reliably for two years. The edge looked chewed. Like someone attacked it with a chainsaw.
Bad photo, I thought. Tried three more. Same artifacts.
So I did something that felt slightly paranoid at the time: I went back to a project from February 2025, found the original raw photos, and re-uploaded the exact same image to the exact same tool.
Worse results in August than February. Same photo. Same tool. Worse output.
That’s when I understood what was happening.
The Systematic Test
I didn’t want to trust one data point, so I tested properly.
Found 10 photos I’d processed between early 2024 and early 2025. Saved the originals and the AI outputs from the first time I ran them. Re-ran all 10 through the same tool in August 2025 and documented everything side by side.
The pattern was unambiguous. 7 of 10 photos from 2024 gave worse results when reprocessed. 9 of 10 from early 2025 degraded. One photo from March 2025 showed a dramatic edge quality drop by August — five months later, noticeably worse.
Not “I think it’s worse.” Measurably, side-by-side worse.
Same input. Worse output. Over time. That’s not a bug. That’s a system degrading.
What’s Actually Happening
The term is AI model collapse. Here’s what it means without the research paper language:
AI systems train on data from the internet. The internet is now estimated to be 50 to 60% AI-generated content. Which means AI tools are increasingly training on their own output — outputs that were already slightly degraded, slightly wrong at the edges, slightly averaged toward the mean.
Think of it as photocopying a photocopy. Each generation introduces small errors. By the tenth generation the text is barely legible and you’ve introduced artifacts that weren’t in the original. That’s essentially what’s happening, except the photocopier is a machine learning model and the copies are the entire internet.
My background removal tool wasn’t getting worse because someone broke it. It was getting “better” according to its training data — training data that included AI-processed edges that were already slightly wrong. The model learned from its own degraded output and optimised toward it.
This is why AI model collapse is hard to fix: the contamination is already throughout the system. AI-generated content is increasingly indistinguishable from human content, which makes filtering impossible at scale. And users have every incentive to pass AI output as human work — on stock photo sites, design portfolios, inspiration galleries.
Some researchers estimate that by 2026, over 90% of online content will be AI-generated or AI-influenced. The degradation cycle I proved in five months is going to compress further.
Where It Shows Up in Design Work
Background removal was where I noticed it first, but once I understood what I was looking at, I started seeing it elsewhere.
UI generation plugins. The Figma plugins that generate interface components train on design files. Many of those files now contain AI-generated elements. Tested one I’d used six months prior with the same prompt: worse components. The spacing felt arbitrary. The hierarchy was generic. It worked but felt like it came from nowhere in particular, because increasingly it did.
Stock photo and illustration generators. AI image generators trained on stock photos. Then people started uploading AI-generated images to stock sites. Now AI trains on AI images. Look at hands in AI-generated images from 2024 versus 2025. They’re getting worse. More fingers, stranger proportions, subtle wrongness that’s hard to name but easy to feel. I noticed this generating team placeholder images for a SaaS product design project — nothing was shippable.
Inspiration search results. Search “modern dashboard design” or “SaaS landing page” and you’re increasingly looking at AI-generated examples. Which means referencing them for inspiration means being influenced by AI output that was influenced by AI output. Feedback loops compounding.
The specific failure mode across all of these: subtle wrongness that’s hard to articulate. The proportions aren’t quite right. The colour relationships are slightly off. Competent but soulless. You can’t explain what’s wrong, but you know it’s not shippable.
That feeling is AI model collapse creating degradation at the edges of capability — the exact places where good design lives.
What Changed in My Workflow
I still use AI design tools. They’re still faster than doing everything manually. But how I use them changed significantly after August.
AI output is a draft, not a deliverable. My background removal process used to take five minutes. I’d run the tool, accept the output, move on. Now it takes twenty: two minutes for the AI mask, fifteen minutes of manual edge review and cleanup, three minutes of final QA. That’s a 300% time increase. It’s still worth it because AI masking is faster than manual masking from scratch — but I’ve stopped treating any AI output as production-ready without human review.
AI-generated research gets verified. User personas, competitive analysis, market research — if AI generated it, I check it against real sources before it influences a decision. I caught an AI-generated competitive analysis that listed features that didn’t exist. The model had inferred capabilities from product category descriptions rather than actual product behaviour. Research used as a substitute for real understanding is a problem regardless of who generates it. AI-generated research that’s actively wrong is a more expensive version of the same problem.
Reference libraries are human-curated. I stopped pulling inspiration primarily from Pinterest, Dribbble, or Behance — those platforms are increasingly polluted with AI work that looks good in thumbnails and falls apart under examination. I look at actual shipped products, talk to real users, reference physical design. I keep a manual collection of designs I can verify came from human designers working on real products with real constraints.
I save good AI outputs. If AI generated something useful twelve months ago, I keep it. The same prompt today might give worse results. This sounds paranoid. I’ve proved it’s not — same input, worse output, documented across ten photos. The good outputs become reference material.
The Uncomfortable Arithmetic
AI companies aren’t discussing model collapse publicly because it conflicts with the story that AI keeps improving. But the arithmetic is straightforward.
Training requires massive data. The internet was that data source. The internet is now majority AI-generated in many content categories. Every retraining cycle pulls in more synthetic content. Filtering it out is functionally impossible because AI content is indistinguishable from human content at scale, and users are actively motivated to pass AI output as human work.
The result is mathematical, not hypothetical: as synthetic data dominates training sets, models optimise toward their own degraded output. Quality regresses toward the average of increasingly averaged inputs.
My background removal tool didn’t announce this. It just got worse. Five-month degradation cycle for a tool I’d trusted for two years. Discovered by accident, proved systematically.
Why This Makes Human Judgment More Valuable
Here’s the part that’s easy to miss in the anxiety about AI replacing designers: AI model collapse is making human creative judgment more valuable, not less.
The ability to spot when AI output has gone wrong. The ability to articulate why something feels off when the wrongness is subtle. The ability to make the refinements that bring degraded output back to quality. The judgment to know what “good” actually means rather than what matches the training data — these are premium skills precisely because AI can’t perform them.
AI can only recognise patterns in its training data. If the training data is degrading, the pattern recognition degrades with it. The human capacity to compare output against reality — against actual users, actual behaviour, actual design that solved actual problems — doesn’t degrade the same way.
The designers who are well-positioned for what’s coming aren’t the ones who rejected AI or the ones who trusted it completely. They’re the ones who learned to supervise it — to use it for the speed it provides while applying human judgment to everything it produces.
That judgment becomes more valuable as AI tools become less reliable. The photocopy keeps getting worse. Someone needs to be the person who notices.
I noticed it in August. Proved it in an afternoon. Changed my workflow the same week.