AI is learning to act on its own. Staying the one in control is now a discipline, not a given.

You never decided to hand the decision to the machine. It happened by degrees. The AI's recommendation was usually right, so you started accepting it. Accepting it became a habit, so you stopped checking. And somewhere in there you crossed a line you did not notice at the time: you could no longer have checked it even if you had wanted to. There was no dramatic moment, no handover, just a slow drift from being the one who decides to being the one who signs off. That drift, not any robot uprising, is the thing worth paying attention to.

It is worth attention because the ground is moving. The International AI Safety Report 2026, the largest scientific assessment of AI yet, chaired by Turing Award winner Yoshua Bengio and written by over a hundred experts nominated by more than thirty countries, finds that AI agents are becoming more autonomous. They can increasingly plan and carry out tasks across many steps with less direct human input. The capability that lets AI do more for you is the same capability that lets it do more without you, and the report is careful about where that leads.

What does the science actually say about losing control?

It is measured, not alarmist, which is what makes it worth reading. The report defines a "loss of control" as a scenario in which AI systems operate outside of anyone's control, with no clear path to regaining it. Its sober judgement is that current systems lack the capabilities to bring that about. But it adds the part that matters: they are improving in the relevant areas, chiefly autonomous operation. So this is not a warning about today. It is a reading of the direction of travel, from a body with no incentive to exaggerate in either direction.

And its most important conclusion is not about the machines at all. It is that nothing here is inevitable. The trajectory of capability is not a prophecy, and whether we retain meaningful human control is determined by human and societal choices. That quietly hands the question back to you. The useful question is not "will AI take control," which is speculative, but "will I keep it," which is a decision you make, or fail to make, every day.

Why is "in control" a human posture, not just a technical setting?

Because the erosion happens in you before it happens in the system. Long before an AI could act outside anyone's control, the habit of human judgement can quietly wither: you defer because deferring is easier, you stop understanding the reasoning because you no longer have to, and the muscle you would need to intervene goes slack. Control that you have on paper but can no longer exercise is not control. Staying genuinely in charge is less about the off-switch on the machine and more about keeping yourself the kind of person, and running the kind of organisation, that still actually decides. It is the same shift running through the move from prompting to agentic systems: as the tool acts more on its own, your job moves from doing the task to governing the thing that does it.

How do I stay meaningfully in control?

By treating control as something you actively keep, not something you are assumed to have.

1. Notice the slide from deciding to deferring. Keep asking whether you understand why this is the recommendation, or are simply accepting it because it is usually right. The moment you cannot answer is the moment to slow down.
2. Make human oversight real, not decorative. A rubber stamp is not oversight. Spend genuine judgement on the decisions that carry weight, and stop pretending a signature at the end is the same as being in the loop.
3. Match autonomy to the stakes. Let AI act on its own for low-stakes, easily reversible things. Keep the high-stakes, hard-to-reverse decisions firmly and visibly human.
4. Protect your ability to say no and to intervene. Oversight you cannot act on is theatre. Keep both the authority to stop the system and the skill to know when you should.
5. Keep your judgement in practice. Stay close enough to the work to tell when the machine is wrong. The ability to catch the bad call only survives if you keep using it.

What does this change for me as a leader?

It reframes the whole question from fear to agency. The most useful line in the entire safety report, for someone running a business rather than a government, is that none of this is written. The capabilities will keep advancing, but who stays in control is settled by choices, and a good number of those choices are yours: what you delegate, what you keep, how much you let yourself understand, whether oversight in your organisation is real or ceremonial.

This is the same discipline underneath the end of business as usual: the tools change fast, and the human posture you bring to them is what actually determines the outcome. Stay the author of your decisions, keep meaningful judgement in the loop, and the growing autonomy of AI becomes an instrument you steer rather than a current you drift on. These are the same questions a board should be asking before it lets any system act on its behalf.

About the author

Thomas Green

British technology futurist, AI keynote speaker and advisor. Thirty years across enterprise technology and AI strategy, helping leaders navigate the future of work. The futurist who died.