What aren’t they telling us?
In June 2026, Anthropic put up a paper When AI Builds Itself. It suggests that AI is about to design and create newer versions of itself without human input. This is frightening as it brings to mind scenes of robots replicating themselves and taking over humanity - not helped by the virus looking animation the paper comes with. Over the next three blogs we’ll look at this paper through three lenses - data & AI literacy, data & AI ethics and data communication.
Let’s start with data literacy.
What aren’t they telling us?
This is always the hardest to determine and it’s super hard within Anthropic’s paper as it requires people to have a reasonable knowledge of the developments within AI, particularly around coding and how it works. It turns out that knowing this will change how you see their charts.
AI Coding Timeline
So, in case you don’t know it, here is a simple timeline of some of the key coding developments in the last four years1,2 & 3
One of the biggest problems in AI is its reliability and all the AI companies came up with the same fix. Rather than focus on the models (as they’re probability models and will always have some element of unpredictability) - they decided to control the system around it. This led all of them to build coding harnesses - human written prescriptive code that handholds the AI to help ensure it uses the right tool, to only access what it should, to look in the right places, to use the right tests etc.
(A coding harness kind of reminds me of how we’d let a toddler play - rather than just letting them loose outside, we set up fenced off areas, with age-appropriate toys, the tool shed locked & the key kept well-hidden and of course an adult keeping an eye on them 😊)
Why does that matter
Knowing roughly what coding harnesses are and when they developed helps you to understand what is most likely causing the results that are on display.
Taken at face value, the chart below (taken from the paper When AI Builds Itself) looks like almost exponential growth and given the title of the paper it would be easy to conclude that our AI overlords are just around the corner.
But if we line up the chart above with the AI coding development timeline from before, you’ll see a different picture.
In fact, the big jump we see in 2026 coincides with the development of coding harnesses - you know the one that involves a lot of human written code and system preparation🤔. That doesn’t sound like much to do with AI building itself. The same pattern occurs with the other charts they provide.
Myriad of issues
There are a few other things to be aware of so that you can see why having the whole picture matters.
The amount of code generated per person is based on asking engineers to estimate how much they think they’re doing. That’s subjective, and people may well say what they think they’re supposed to say.
This is all about coding, which is probably the easiest place for AI to do well. There’s lots of training data, clear patterns, and right-or-wrong answers. There’s no guarantee that it will work as well in the more nuanced, messy, real-life stuff the rest of us work in.
One chart looks at how well the AI performs on problems engineers got stuck on and got wrong. But we don’t see a breakdown by experience level and we don’t see how it performs on the code engineers got right.
The long bow
To be fair, the commentary for each chart doesn’t actually state that it’s evidence of AI building itself - just that they’re seeing the big jumps around the start of 2026. But rather than suggesting it’s a result of the human effort put into the development of coding harnesses (and the controls that we have around that), they have decided to draw a very long bow and suggest that this means that AI is about to build itself. There is absolutely no evidence of that and apart from the intro and the imagining of some possible futures at the end, the rest of the paper itself doesn’t suggest it.
So what
So, if most of the paper isn’t really saying that AI is building itself, why all the fuss? Because most people (media included), will only read the introduction and maybe quickly view the charts with the big jumps in 2026 and assume that’s proof that AI is building itself. This is really scaring people and it’s not fair (but more on that in the next blog)
The challenge
Next time you see some scary AI headline try and look a little deeper, either by googling, youtubing, going to someone you trust or even AI’ing it and ask what else happened in that time frame - what is missing from the story? Or if you’re keen read the actual paper to see if it matches the clickbait.
