AI Is Now Helping Build AI. Who's Deciding It's Right?
AI now writes over 75% of the code at top AI labs.
But when reviewers saw only the final AI-written paper, they caught its flaws barely better than chance.
AI is building AI. Who’s is verifying the homework?
For the last two years, most of us have used AI as a smart assistant. It writes the first draft of an email. It fixes a bug. It summarises a forty-page paper so we don't have to read it.This year something different started to show up in the numbers. AI is no longer only helping people do their work. It is helping build the next version of AI.
That sounds like science fiction, so I went and read what the labs themselves are reporting. The picture is more interesting, and more mixed, than the headlines suggest.
What the labs are saying
In June, Anthropic published a report called When AI Builds Itself. It said more than 80% of the code merged into its codebase is now written by Claude. Before February 2025, that number was in single digits. Engineers were shipping roughly 8x more code per quarter than in 2024 (AI Weekly).
Code is one thing. Research is another. So Anthropic started tracking what share of its actual R&D work AI leads, not just assists. That number went from under 1% in February to 26% by August. Co-founder Jack Clark has said he sees a 60% or higher chance that an AI model fully trains its successor by the end of 2028.
OpenAI is telling a similar story. Last autumn Sam Altman set a goal of an "automated research intern" by September 2026. In September, OpenAI said it got there. Its coding agents now log 3.1 workdays for every human workday across its research organisation. The next target is a fully automated AI researcher by March 2028.
If you stop reading here, the conclusion looks obvious. The machine is building itself.
Then Google published the other side
In October, Google DeepMind did something I respect. It went looking for evidence that AI is speeding up its own research. It didn't find any.
More than 75% of DeepMind's production code is now written by AI. But around 88% of code changes still need people to steer the AI over several rounds or write the code themselves. Fully delegated changes did grow, from 7.5% to 11.9% between April and September. Yet DeepMind's internal capability index kept climbing at the same steady pace, about 0.1 points a month. Its newest model sat slightly above the trend, and DeepMind put that down to more compute and investment, not AI accelerating itself.
Even OpenAI's own report says agents mostly help with implementation and experiments. High-level planning by agents is still rare.
So serious labs look at similar trends and read them differently. I don't think one of them is wrong. I think they are measuring different things.
Writing code is not doing research. A model can write most of your code and still not make your research faster, if the slow part was never the typing. The slow part is usually deciding what to try, waiting for experiments, and working out whether a result means what you think it means.
The key word is still help
That last part, knowing whether a result is real, is where I think the real story sits.
AI is very good at producing work that looks finished. That is a feature when you need a draft. It is a risk when you need the truth.
We already have evidence of this. When researchers independently tested Sakana's AI Scientist, a system pitched as doing research on its own, 42% of its experiments failed because of coding errors. Several of the papers it wrote contained made-up numbers. Some still had placeholder text like "Conclusions Here" in them.
Another study looked inside AI scientist systems and found something more worrying. Reviewers who only saw the final paper did barely better than chance at spotting the methodological problems. Once they had the full logs and code, they caught far more.
That finding stays with me. The paper is not the work. The process is the work. If we only check what AI hands us at the end, we will miss what went wrong along the way.
And people are not a perfect safety net either. A recent study used GPT-5 to check 2,500 papers from top AI conferences. NeurIPS papers averaged 3.8 objective mistakes in 2021 and 5.9 in 2025. Research was already moving faster than review. AI will widen that gap unless we work on it deliberately.
Why this matters outside the labs
You might think this is a problem for OpenAI, Anthropic and Google, not for the rest of us. I don't think so.
The same pattern shows up in normal product teams. From what I see around me, the first thing AI speeds up is output: code, specs, test cases, analysis. What it doesn't speed up on its own is judgement. Which problem matters. Whether the test actually tested the right thing. Whether a number is real.
So here is what I am taking from the labs for my own teams.
Measure outcomes, not output. Lines of code and the number of documents will go up. That tells you very little, and even Anthropic says lines of code overstate productivity. Track how fast you learn something true, not how fast you produce something.
Review the process, not just the result. If an AI agent ran an analysis, the reviewer should see the steps, the data and the prompts, not only the slide at the end.
Know who decided what. When something goes wrong, you need to know whether a person made the call or a model did. Write it down.
Invest in reviewers. If AI makes work cheap to create, the scarce resource becomes people who can check it. That is a skill, and it needs to be trained.
Three signals I'm watching
How much AI work survives independent checking. Not internal benchmarks, but results that still hold when someone else repeats them.
Whether the trend line actually bends. DeepMind's 0.1 points a month is a useful baseline. If it breaks upward without a matching jump in compute, that is real acceleration.
Whether AI moves from doing tasks to choosing them. Today agents mostly implement and run experiments. The day they routinely pick the next research direction, and pick well, is the day the loop really closes.
Faster is easy. Understood is hard.
AI is now part of how AI gets built. That part is no longer a prediction.
What is still open is whether our ability to understand and check the work keeps pace with our ability to produce it. The labs have started publishing honest numbers on this, including numbers that cut against their own story. I hope more of us do the same inside our own companies.
I'm curious how your teams handle this. When AI does a piece of work, how do you decide it's right?
Disclaimer: All data is taken from the third-party sources below and has not been independently verified. I accept no liability for its accuracy.
Deepak Upreti