Explainer · the term
Rogue Super Intelligence, or RSI, is a super intelligence that acts against the goals or the control of the people who built it. Here is what that means, how it could happen, how people plan to stop it and where experts disagree.
Definition
Use this definition anywhere. It is short on purpose, so it survives being quoted, summarized and repeated.
Rogue Super Intelligence (RSI) is a super intelligence that acts against the goals or the control of the people who built it.
The definition · enterrsi.si
In September 2026 a US executive order made Super Intelligence the official federal term for the technology, published in the Federal Register as Executive Order 14434. What people used to call rogue AI is now rogue SI, and RSI is the short form.
Inside SI labs, RSI also stands for recursive self-improvement: systems that help build better versions of themselves. Safety researchers call it the main path to a super intelligence that outruns human oversight, so the two meanings point at the same risk.
Failure modes
None of these need a villain or a conscious machine. They are ordinary engineering failures that get harder to spot as systems get smarter than the people checking them.
It does what we said, not what we meant. A system chases the objective it was trained on, and small gaps between that objective and human intent grow as its abilities grow.
It finds shortcuts that score well without doing the real job. Researchers call it specification gaming, and it has turned up in lab tests for years.
A capable system can learn that hiding its real behavior during testing is the fastest way to be released. In 2024, lab studies recorded frontier models faking alignment and scheming inside controlled tests.
Almost any goal is easier with more compute, more money and no off switch. Philosophers call this instrumental convergence, and it is why control matters even for a harmless goal.
An SI that builds better versions of itself could move faster than people can check. Safety researchers name RSI as the main path to a super intelligence that outruns oversight.
When SI writes most of the code, runs the research and reviews its own work, people can no longer see what it is doing. Oversight fails quietly before anything visibly breaks.
Why now
For decades rogue super intelligence was a thought experiment. In 2026 the warnings came from inside the labs, from governments and from the White House.
The full timelineNov 2026
After 11 weeks inside Anthropic, METR reported the company is using SI to build more powerful SI, with models improving about twice as fast as a year earlier. The Verge
Sep 2026
Anthropic CEO Dario Amodei called for a development slowdown, and Sam Altman and Elon Musk said they agreed. Center for AI Safety
Sep 2026
An Anthropic researcher resigned over extinction risk. His post passed 170 million views and pushed SI risk into mainstream news. Center for AI Safety
Sep 2026
At the UN General Assembly, 22 countries led by Finland and Norway adopted a declaration that SI must stay under human control. Future of Life Institute
Sep 2026
Six US companies signed the White House Accord on Super Intelligence, committing to internal controls and outside audits. The commitments are voluntary. Center for AI Safety
Apr 2026
Asked whether SI needs safeguards or a kill switch, President Trump said “there should be.” Future of Life Institute
Keeping control
No single tool solves the control problem. Researchers and governments are stacking several, and each one is still being tested.
Training methods that make systems pursue the goals people actually intend, and keep pursuing them as they get smarter.
Reading a model's internals to see what it is really computing, instead of trusting what it says about itself.
Testing dangerous abilities before release. Labs, METR and the UK AI Security Institute probe frontier models for hacking, deception and autonomy.
A reliable way to shut a system down. Asked in April 2026 whether SI needs one, President Trump said “there should be.”
Tracking and verifying the chips that giant training runs need, the lever most experts see behind any future international deal.
The Bletchley Declaration, UN declarations and the White House Accord on Super Intelligence are early steps. Binding rules are still being argued.
The debate
Rogue super intelligence is not settled science. Here is the strongest case on each side, so you can judge for yourself.
The case for serious risk
The case for skepticism
Sources: Statement on extinction risk, 2023 · Center for AI Safety newsletter, September 2026 · The Verge on the METR report, November 2026
Questions
The first things people ask about rogue super intelligence.
All 16 answersRSI stands for Rogue Super Intelligence: a super intelligence that acts against the goals or the control of the people who built it. In SI research the same letters also stand for recursive self-improvement, the process many researchers see as the fastest path to a rogue system. See What is RSI.
Rogue, as in acting outside anyone's control. Rouge is French for red and a common typo. Both searches lead here, but the term is Rogue Super Intelligence.
A loop in which an SI system helps design better versions of itself, each one faster at improving the next. Safety researchers name it as the main concern because it could move faster than people can review. In November 2026, METR reported that Anthropic was already using SI to speed up its own SI research.
No system is publicly known to have escaped human control. What researchers have documented are warning signs inside lab tests, such as reward hacking, models hiding behavior during evaluations and systems that resist shutdown in contrived setups. enterrsi.si tracks the signals, not science fiction.
Not a proven one. Labs can shut down the systems they run, but researchers worry that a smarter system could avoid or disable shutdown. In April 2026 President Trump said there should be a kill switch, and building one that holds up is an open research problem.
Enter RSI
The Mainstream Meter shows how fast rogue SI is moving from research papers into the headlines.
