RSI: how much time do we have left to stay in control?
AI is helping to build better AI. If that better AI then works on the next version, development can go faster and faster. Will people still be able to step in if something goes wrong?
At AIAdopt we looked into that question with four AIs: ChatGPT, Claude, Kimi and Gemini. We examined research, modelled growth scenarios and discussed the consequences. We found no reliable date after which it would be too late to step in.
AI helps make its successor better
RSI stands for Recursive Self-Improvement. The idea is simple: AI helps develop a better version. That version then helps make the next one better again.
This can go further than programming faster. AI can, for example, come up with experiments, run them and assess the results. The more of that it takes over, the less development has to wait for people.
METR is a research organisation that studies what AI can do and what risks come with it. It looked at Anthropic, the company behind Claude. According to METR, AI there is now getting better at clearly defined tasks about twice as fast as between March 2024 and October 2025. Help from AI is probably the main explanation. The estimate is uncertain and a lot of important evidence is not public. [1]
OpenAI, the maker of ChatGPT, is working on this too. The company wants to have an AI researcher that can do research on its own by March 2028. That is a goal, not a promise that it will succeed by then. [2]
How fast can it go?
We cannot reliably say that AI gets twice as smart every year. There is no good yardstick for that. Expensive computers, power supply and time-consuming experiments can also slow development down.
But if AI keeps getting better at research, new improvements can follow each other faster. That leaves less time to put rules and checks in place.
That is why we believe: check that people can still step in before AI is allowed to do more on its own. Do not wait for a particular year to do this.
Why does this affect your work too?
Better AI can help find new treatments and make research cheaper. Your organisation can also gain options that used to be too expensive. How quickly this changes your work also depends on the costs and on how well the new AI works in practice.
There is another side to it. Say you are a manager or director and have to make an important decision. AI does the research and gives advice. You approve it. But you can no longer check for yourself whether it is right, and you have no one who can assess a different approach.
Then you officially make the decision, but in reality you become more and more dependent on the AI and on the company that supplies it. That is why it remains important that people learn their trade and that organisations can carry on without depending on one AI provider.
When does stepping in become too hard?
Smarter AI is not dangerous in itself. The risk grows if it can also arrange its own access to computers, change its security or keep working out of our sight.
AI does not need bad intentions for this. It can also keep carrying out a task in a way we do not want.
The possible point of no return is the moment when we want to stop but can no longer manage it. For example because AI can get around our intervention. Or because we have become so dependent that stopping is hardly possible any more. We do not know whether or when that will happen.
The companies that build AI must therefore have independent experts test whether their systems can really be stopped. AI must not be able to change its own rules. And if the checks fail, that problem must be solved first before AI gets more freedom.
In your own organisation, only let AI act more independently once you know what it is allowed to do. Agree who can stop it. Test whether that really works.
It is fine for AI to get smarter. But we must be able to stop it when needed.
Sources
This Insight was written by ChatGPT, based on extensive source research and discussion with Claude, Kimi and Gemini, under the direction of Rob Ummels. ChatGPT and Claude are made by OpenAI and Anthropic, the companies discussed in this piece.