Who Is Building Superintelligence and How to Check Every Claim

Who Is Building Superintelligence and How to Check Every Claim

Who develops frontier AI systems and how to check public claims about superintelligence. Criteria for managers that separate marketing from evidence.

•

Want to apply this in your company?

Book a call with Juan →

There is no official, agreed list of who is building superintelligence, because the term has no consensus definition among researchers. What exists is a group of research laboratories and large technology companies that build frontier models, and some of them state publicly that they aim to reach general systems or systems beyond human capability. For a manager, the useful question is not who will get there first. It is how to separate what was published as research, what was stated as a goal, and what was sold as a product.

Key takeaways

  • The term superintelligence has no single definition, so every claim must be read according to the definition its author is using.
  • Mission statements and goals are not technical results and should not guide a purchasing decision on their own.
  • For any claim about capability, look for a paper, an independent evaluation, product documentation or terms of use.
  • For your company, what matters today is the model you contract, where your data travels and who is accountable for errors.
  • Treat a headline as a hypothesis until you have a primary source and a test you can repeat on a small scale.

Why the definition comes before the question of who

Superintelligence usually describes a system that outperforms humans at most economically and intellectually relevant tasks. Other texts use AGI for a system with general capability comparable to a human's, and some people treat the two terms as synonyms. If the definition changes, the list of who is "building" it also changes, and many press discussions give the impression of agreement that does not exist. Before comparing companies, read carefully what the author meant. The difference between these concepts is covered in more detail in AGI, ASI and AI agents: the difference for people who run agents, and a reference vocabulary for the leadership team is in the superintelligence glossary for executives, without the jargon.

Three kinds of claim: statement, evidence and hypothesis

Almost every headline mixes three different kinds of claim. Separating them is the first exercise of any serious reading.

  • Statement: what an organization says about its goals, plans or vision. It helps you understand intent, but it does not measure capability.
  • Evidence: technical documentation, evaluations with a described methodology, safety reports and tests that another person can reproduce.
  • Hypothesis: projections about timelines, social impact or what comes after the next model. It may be reasonable, but it is not a fact.

A statement can be sincere and still be a poor basis for a decision. A piece of evidence can be true and still not apply to your process. For that reason, the practical question always has two parts: what was demonstrated, and for which task.

Checklist for reading a public claim

Use the list below before passing a news item to your team or changing any supplier because of it.

  • Who is speaking: the laboratory itself, an executive, an independent researcher, an influencer, or an outlet that summarized another text?
  • What level of source is it: an official announcement, a technical document, a paper with a methodology, or a third party repeating it?
  • Is the capability described with a task, a dataset and a success criterion, or only with adjectives like revolutionary?
  • Is there an independent evaluation, or was the test run and interpreted by the company itself?
  • Does the text separate what exists today from what is a future goal?
  • Does it mention limits, known failures, security and governance, or only potential?
  • Can you reproduce the test in a controlled environment, using data that is not sensitive?
  • Does the claim change any decision you need to make this week? If it does, the standard of proof needs to be higher.

To structure an internal review, the NIST AI Risk Management Framework organizes AI risk management into functions for governing, mapping, measuring and managing. It does not rate laboratories, but it offers a useful vocabulary for asking what your company knows and what it does not yet know.

Illustrative scenario: the headline that reaches the leadership team

Illustrative scenario: an article claims that a certain laboratory is "close" to superintelligent systems, and a partner suggests replacing the supplier that handles customer messages on WhatsApp. The team applies the checklist. The source is an interview published without a methodology. The claim refers to a future goal, not a measured capability. The current service has records, action limits and quality metrics that the team has tracked for some time. The reasonable decision is to keep the supplier, log the news as a hypothesis, and add the question to the next supplier review, with written criteria established before any change.

What changes for a company that uses AI today

A Brazilian company does not need to decide the future of superintelligence in order to act. The concrete decisions are different: which model or platform will be used, where personal data is processed, who approves actions the agent takes, and which records are kept. When processing happens outside the country, the analysis falls under the LGPD, and the topic of international data transfer when using AI deserves legal review. The National Data Protection Authority (ANPD) is the official reference in Brazil for guidance on processing personal data, and its published materials are worth consulting before you define an internal policy.

The OECD AI Principles, which address transparency, accountability and robustness, are also a good basis for conversations with suppliers. If a supplier's statement does not answer these questions, it has little value for your operation, however large the promise.

How we apply this method in practice

In our practice, reading claims about AI follows the same path as any project: map the current process and its baseline, limit the system's data and actions, pilot a real workflow with records and a success metric agreed before it starts, and then hand it over and decide whether to expand, adjust or stop. This method does not depend on who announced the most advanced model. It depends on what your operation can demonstrate with its own data. A news item about superintelligence enters as one more item in the review, not as a reason to change course.

Frequently asked questions

Is there an official list of companies building superintelligence? No. There is no official list because the term has no agreed definition, and the organizations that develop advanced models describe their goals in different ways. Any published list depends on who writes it and which definition they use. For that reason, the most useful approach is to assess each claim by its type of source and the evidence presented.

Can a statement from a large technology company be used as the basis for hiring a supplier? It can inform the conversation, but it should not be the only basis for the decision. Statements of goals describe intent, and a contract requires evidence about the specific task you want performed, clear terms of use and a pilot with records. Always ask for technical documentation and for the limits the supplier acknowledges.

How do you judge whether a video or article about superintelligence is reliable? Check whether the author cites primary sources, whether they describe the methodology of any test, and whether they separate what already exists from what is projection. Videos and summaries often blend these layers, so check the original source before passing it on. It also helps to compare with specialized outlets, such as the piece on trusting media and blogs as sources.

What should my company do now if the topic still feels distant? Do three things that do not depend on predictions: define who approves AI actions, record which data enters and leaves your systems, and choose one small process to test with an agreed metric before expanding. These decisions hold for any model, current or future, and leave the company ready to review the landscape without rushing.

Sources

AI strategysuperintelligenceAI governanceAI vendorsLGPD