What Is Superintelligence?: The Research Idea, and Why This Week's Headlines Use the Word for Everything

What researchers mean by the word

The standard picture comes from Nick Bostrom, who described superintelligence as a system that would surpass human intelligence, not as a label for the chatbots and filters that already exist. Mashable's account of this week's accord uses that line to separate the philosophy from the press conference, and it is the right split: a superintelligent system would be better than expert humans at science, strategy, and the design of the next system.

People often collapse three rungs into one word. A language model predicts the next piece of text and can be very good at that. Artificial general intelligence, in the way labs use it, would handle unfamiliar tasks about as well as a person. Superintelligence sits above that, a system that would beat the best human groups at most of those tasks and could keep widening the gap by improving itself. Crossing the last rung is the claim. Shipping a chatbot is not the crossing.

That is also why the timeline arguments sound so different from each other. Some researchers told the New York Post this week that the pace now looks faster than they expected, and that the turning point is when a system gets better than humans at AI research. Others still treat that point as a hypothesis. Both can be said out loud without pretending the tool on a student's laptop has already arrived there.

What happened on September 29

On September 29, 2026, President Trump and the leaders of Google, Anthropic, Meta, OpenAI, xAI, and Nvidia signed a one-page voluntary pledge, the White House Accord on Super Intelligence, also described as a joint commitment on frontier responsibilities. Mashable's explainer is the clearest public walkthrough: the document asks each signer for internal monitoring, an internal team that maintains those controls, independent auditors, and an independent committee at the board. It does not create a new regulator, and it does not bind a company that prefers not to follow through.

The same day, an executive order told federal agencies to say Super Intelligence, or SI, instead of artificial intelligence in documents that are not already fixed by statute. That is a vocabulary instruction. It calls on Congress to consider writing the new term into law. It does not certify that a lab has built a system smarter than every researcher in the building.

Other executives were in the room, including leaders from Microsoft and Amazon, and reporting so far says those two companies did not sign. OpenAI was represented by Greg Brockman, not by Sam Altman. The useful fact is the short list of signatures and the voluntary wording, not the guest list.

What the word does not mean yet

ChatGPT, Claude, Gemini, and Grok are not superintelligence under the research definition. They draft, summarize, and miss. They also get used, badly, as ghostwriters. A course that treats a fluent paragraph as the arrival of a new species is making the same mistake as a headline that uses the White House's new label for a spam filter.

Labs have been aiming at the stronger claim for a while. Ilya Sutskever left OpenAI in 2024 to start Safe Superintelligence, a company whose only stated product is a safe version of that system, and which still had no public model to try as of this fall. Meta has used the same word for its own lab. Ambition and a product you can paste into a box are different objects.

Safety talk has not waited for the product. The accord's four controls are a promise to watch the work, audit it, and put a committee near the board. Researchers who worry about extinction risk want something harder than a voluntary page. Neither side has demonstrated a machine that can replace a research group and then redesign itself without help.

Why it still matters for a writing score

The practical problem on a campus is older than this week's rename. A detector does not ask whether the lab that trained the model is chasing superintelligence. It asks whether the finished sentences are even, safe, and flat. How to read an AI writing score is that narrower job: a percentage is a share of the prose, not a verdict about the future of the species.

A high score can still land on writing a person typed. Template essays and very regular English look like model prose for statistical reasons, which is why a high AI score does not mean the student did not write it. Superintelligence, if it arrives, would make generated text harder to spot by eye. It would not turn today's percentage into a recording of who sat at the keyboard.

If the worry is a course report rather than a philosophy argument, the campus tool and a public checker remain different classifiers. Can Turnitin be wrong is the page for a surprising badge. DetectGPT will not reprint that badge, and a kinder number here is not proof that a future system is harmless or that a current draft is clean.

FAQs

What is superintelligence?

In AI research it means a system that would outperform the best humans across almost every important cognitive task, including scientific work and improving AI. It is a threshold, not a brand name for today's chatbots.

Do we have superintelligence now?

No. Current models are powerful text and image systems. Labs talk about building something far beyond them. Talking about it is not the same as shipping it.

Why is the White House using the phrase for all AI?

An executive order on September 29, 2026 told agencies to say Super Intelligence instead of AI in ordinary documents. That is a naming change. It does not match the research definition.

What did the tech companies actually sign?

A voluntary one-page accord: internal monitoring, an internal controls team, independent auditors, and a board-level committee. Google, Anthropic, Meta, OpenAI, xAI, and Nvidia are the reported signers. It is not a law.

Does a writing detector measure superintelligence?

No. DetectGPT scores how machine a draft sounds. It does not rank a lab's progress toward a system smarter than its own researchers.

Will a smarter model make AI writing impossible to notice?

Harder to notice by vibe, maybe. A score is still a style estimate, and careful human writing can look even. Read the lines, and do not treat the percentage as a confession.