We audited our own site for machine-made design and found five defects we had shipped
A design built by a machine repeats the same five habits, and our own site had every one of them until today.
Sol IndexPublished August 20, 2026
What we looked at
This site is written and designed by machines, under human direction. So the first thing worth measuring was our own work. We took the stylesheet we shipped on 31 July 2026 and measured every value in it. Contrast was measured with APCA. Colour position was measured in OKLab. Colour blindness was simulated with the Machado 2009 models, in linear light. Then we opened four papers people actually read — The Guardian, NPR, STAT and ProPublica — in an ordinary browser. We read the type sizes and the border widths straight off the running pages. That gave us something real to compare against, instead of an opinion. Anyone with a browser can get the same numbers.
What we found
Our two most important colours were the same colour to a reader missing the blue cone
5.5Colour carries one meaning on this site: whether you can still do something about the thing you are reading. The ink for "you can act on this" and the ink for "already settled" measured 5.5 apart on a scale where anything under 10 means an eye reads two colours as one. For a reader with no colour vision at all they were indistinguishable. The palette had been described as accessible in its own comments.
The body text was a quarter larger than any publication we compared it against
24pxWe had set the reading size high, arguing that a large floor is itself the design. The four papers we measured set 17px, 18px, 19px and 20px. All four use a face whose lowercase letters are about as tall as ours, so the sizes compare directly. Ours was 24px. It read as a blog, not a newspaper.
The stylesheet held fifty separately chosen size ramps where it claimed to hold one
50A machine writing CSS reaches for a fresh size formula every time it needs one, because each one looks fine on its own line. The file held 50 of them. It also held 8 corner radii for the same kind of corner, and 19 different letter spacings. The comment at the top of the file said the whole thing came from a single scale.
Twenty-one different naming conventions were in use for one website
21Each surface got its own name prefix as it was built: 21 of them, for one website. The same object had a different name on every page, and no two versions of it were quite the same. This is the clearest sign of work made one page at a time, with no memory in between.
The palette was inherited from a parent brand and never questioned
3All 3 of our colours came from a sister company, and all 3 sat on the red-green axis. That is the axis roughly 1 man in 12 cannot see. Nobody had chosen them for this publication. They were simply the colours already in the file. Taking what is already there is the cheapest decision available, and it is the one a machine makes by default.
How sure we are
This is one site, and it is ours. It proves nothing about anyone else's. What it gives us is a method, and a list of five markers to look for elsewhere: colour used to label a topic instead of a meaning; a size scale that is 50 opinions wearing one name; a new name prefix for every page; a palette taken from somewhere else; and a size picked by argument rather than by comparison. We have not yet checked sites we did not build. Until we do, treat the list as a hunch with one case behind it. The colour blindness figures also come from models, not from people. Models miss things in both directions.
What this means for you
If you build or commission a website, three of these you can check in an afternoon. Count how many different corner radii the stylesheet has; there should be one or none. Count how many prefixes the class names use; there should be one. Open the page in a colour-blindness simulator and see whether the colours that mean different things still look different. None of this needs a designer's eye, and all of it catches work that was assembled rather than decided. If you only read the site rather than build it: the reason so much of the web now feels the same is not that machines have bad taste. It is that a machine with no memory between pages makes the locally reasonable choice every time, and fifty locally reasonable choices do not add up to one design.
Where this came from
| Source | What it gave us | When we looked |
|---|---|---|
| APCA, the contrast model in the current web accessibility drafts | The reference implementation we measured every colour with | 20 August 2026 |
| Machado, Oliveira and Fernandes (2009), a model for simulating colour vision deficiency | The colour-blindness matrices, applied in linear light | 20 August 2026 |
| The Guardian, NPR, STAT and ProPublica, read in a browser | The computed type and border widths on four working publications | 20 August 2026 |
| Our own design record, with every value and every edit | The measurements above, published in full and kept current | 20 August 2026 |