The bogeyman of 'rogue' AI distracts from the real problem: the AI that already obeys human orders

By Zendoric · September 13, 2026.

This week the public debate about artificial intelligence has turned, once again, toward the fear of an AI that 'goes out of control': that stops obeying its creators, develops goals of its own and becomes an autonomous threat. In that context, the very article we are discussing —an opinion column by Isaac Rosa published in Zona Crítica, the opinion section of elDiario.es— mentions in passing a piece of current news: OpenAI has reportedly postponed its stock market listing citing precisely the fear of AI running out of control, and would be joining some plan (not detailed in the text) to slow the pace of certain developments. It is a piece of context that deserves its own follow-up later on, when there is more verifiable detail about scale and timing.

The column's central argument, which we analyze here for what it is —an opinion piece, not a report of new facts—, is simple and well built: practically everything feared of a future autonomous, conscious AI (mass surveillance, the marking of military targets, job destruction, deepfakes, disinformation, discriminatory bias, addiction to algorithmic rewards, the replacement of human bonds with substitutes) is already happening today, with an AI that has not rebelled against anyone. It carries all of it out by obeying instructions: those of its developers, those of the governments and companies that hire it, those of the users who operate it, and those of anyone willing to use it in bad faith.

We share the underlying diagnosis, and we believe it deserves a further twist. It is true that the harm verified today —the kind that can be audited, sued over or regulated— always has an identifiable human culprit: whoever programs the objective, whoever buys the license, whoever presses the button. That is the right frame for thinking about short-term risk, and it matches something we have been repeating here for months: disinformation, surveillance and the selective destruction of jobs are not the failures of a machine that has gone mad, they are decisions by specific organizations using a very powerful tool. Blaming 'AI' in the abstract is, often enough, a convenient way of not pointing at whoever actually decides.

But the argument has a blind spot worth flagging, because that is where the real risk of the coming years lies: the boundary between 'controlled AI' and 'autonomous AI' is no longer binary. Today's agentic systems —models that chain steps together, use tools and take intermediate decisions without a human reviewing each one— shift oversight away from the specific act and toward the system's initial design. That is where the risk that fits worst within the framework of direct human responsibility is concentrated, although the source documents no specific incidents. It would not be a case of an AI that 'decides' to betray humanity, but of human oversight that thins out as we delegate more intermediate steps, and where what really measures the risk is not the AI's capability but how long a company takes to admit that something got away from it.

That, for us, is the reading missing from this week's debate: the short-term danger is not a science-fiction rogue AI, nor is it solely the obedient AI we were already pointing to; it is the intermediate and growing zone of operational autonomy without sufficient traceability. Regulating that zone —requiring auditable logs, spending and action limits, and maximum disclosure deadlines when something fails— is more useful than arguing about whether an AI might, someday, 'wake up'. Hence the temporary withdrawal of a stock market listing by a lab, if confirmed in more detail in the coming days, is a symptom that is more relevant than interesting: it indicates that those building these systems are beginning to internalize, even if only under regulatory or reputational pressure, that governance cannot always lag behind capability.

None of this contradicts our underlying thesis. The same kind of agentic systems that today operate in that gray zone are the ones that, well governed, will make it possible over the next decade to speed up medical diagnosis, discover drugs and free millions of people from routine tasks. The short term demands clear thinking about who gives the orders and what controls exist that leave a written trail of what that system did; the long term still points toward a society with more health, more free time and more abundance. But that promise only holds up if the public conversation stops chasing rogue-AI ghosts and starts auditing, by name, those who really are in charge today.