Who Is Building Superintelligence? How to Read the Claims Carefully

Nobody maintains an official list of who is building superintelligence. The term has no agreed definition, and no institution certifies that anyone has reached it. In practice, people who follow the topic work with three kinds of source: labs and research groups that publish technical work, institutions that assess and report on the field, and public communication from companies and people who sell or promote products. The way to avoid being misled is to weigh each source differently and ask for evidence before turning a headline into a decision.

Key takeaways

• Superintelligence has no operational definition that everyone accepts, so be wary of anyone who says we are "close" without stating the criteria.
• Peer-reviewed papers, technical reports, evaluation documents and interviews carry different weight and should not be read the same way.
• A capability announcement usually signals intent or positioning. It is not proof that a system does what was claimed.
• For your company, the useful question is not who will create superintelligence, but what AI available today can run safely, with logs and human review.
• A short monitoring routine, reviewed on a fixed schedule, prevents impulsive reactions to every news item.

Why "who is building it" is a hard question

Part of the confusion comes from the vocabulary itself. Superintelligence is usually described as a system that would outperform people on most economically relevant cognitive tasks, but different authors use different criteria, and some never specify how it would be measured. The term circulates alongside AGI, agents and language models, so the same news item can be read as "progress toward superintelligence" by one person and as "another model with better benchmark scores" by another. Before reading any map, it helps to understand how these concepts differ. The article on AGI, ASI and AI agents explains the distinctions, and a superintelligence glossary written for executives covers the rest.

The three source types and how much weight each deserves

AI labs and academic groups are the most useful primary sources, provided the material includes method: how the system was evaluated, which limitations are acknowledged, the evaluation data and, where it exists, a safety report. Even then, a technical paper describes what was tested, not what the market will do with it.

Annual reports and indexes produced by universities and international organizations help you see trends and compare approaches. They cite their sources and explain their methodology. They also age quickly and rarely answer a specific question about your operation.

Interviews, posts, talks and press releases are useful for understanding intent and positioning. Anyone selling a product has an interest in making the underlying technology seem imminent. That does not make the information worthless, but it means treating it as a claim until verifiable evidence appears.

How to read a capability claim without being fooled

Before you share or react to a news item, ask these questions:

• Who is speaking, and what do they gain from the statement? Someone selling a product has a direct interest in the outcome.
• Are the criteria written down, or does the sentence rely on vague words like "thinks," "understands" or "surpasses humans"?
• Does the claim point to a published evaluation method, with data, conditions and limits?
• Has someone independent been able to test, reproduce or meaningfully discuss the result?
• Does the statement refer to a system you can buy, an internal prototype, or a long-term goal?
• Are the date and context clear, or is the news recycling older material as if it were new?
• Is the practical change for your process concrete, or does it depend on a scenario that has not been verified?

If you answer "I don't know" to three or more of these, record the item as a hypothesis, not as information for a decision.

A five-step monitoring routine

For people who need to make decisions without following the topic full time, this routine is usually enough:

• Step 1: define the business question. For example: what changes in customer support if a vendor releases a more capable model?
• Step 2: choose three to five primary sources, such as technical publications from the labs whose models you already evaluate, reports from credible institutions, and the official documentation of the providers you use.
• Step 3: for each relevant development, record in a simple document its classification (verified fact, hypothesis or marketing), the original source and the access date.
• Step 4: review the record once a month with the people who make technology and process decisions, without treating each item as urgent.
• Step 5: change a process only when a pilot with an agreed metric and usage logs supports the change. A single news item is not a reason to change anything.

Illustrative scenario: the headline that arrives in the team chat

Illustrative scenario: a coordinator forwards a headline in the work group saying an AI company is "close to creating superintelligence." The operations manager does not reply right away. He finds the original source, which is an executive's remarks at an event, with no published method. In the monitoring record, he classifies the item as a statement of intent, notes the date and asks that next time the link to the primary source come with the headline. No tool in the operation changes that week. The team handles the topic with the same calm it would apply to any vendor's proposal.

What this means for your company today

The useful work for a company is operating the AI available today with proper controls. The NIST AI Risk Management Framework organizes this work into functions for governing, mapping, measuring and managing risk, and the OECD AI Principles highlight transparency, accountability and robustness as general reference points. These frameworks help turn a debate about the future into concrete questions: who owns the process, what data is used, what gets logged, and what the stopping criteria are.

In Brazil, when an agent handles personal data, that processing falls under the LGPD, regardless of who developed the underlying model. International data transfers have their own rules and deserve specific reading, as covered in the article on international data transfer with AI. To follow the national regulatory environment, the article on Brazil's AI legal framework is a good starting point.

In our method, the sequence is: map the current process and its baseline, limit data and actions, pilot a real workflow with logs and an agreed success metric before starting, then hand over and decide whether to expand, adjust or stop. Applied to this topic, the logic is simple: follow the field with a regular review routine, and change operations only when a pilot shows something verifiable.

If you want to separate the background debate from what actually affects your business, the article on superintelligence myths helps dismantle expectations that often get in the way of good decisions.

Frequently asked questions

Who is building superintelligence today? There is no official list. Frontier AI labs, academic groups and technology companies are building increasingly capable systems, but there is no consensus on when a system counts as superintelligence. To decide safely, rely on primary sources and explicit criteria, and treat promises about timing as hypotheses.

How can I tell whether a superintelligence news story is reliable? Find the original source and check whether a method is published, who is speaking, and whether that person or company has a commercial stake in the topic. If the story leads to nothing verifiable, record it as a hypothesis and do not use it as a basis for changing processes.

Do I need to follow this topic if my company only uses common AI tools? Yes, but less often. What affects daily operations is quality, cost, data policy and the provider's terms, and those are best tracked in a monthly review. Future-focused news only becomes important when it turns into a concrete change at a vendor you use.

Is it worth reading technical papers without a computing background? Yes, especially the summaries, the limitations sections and the evaluation documents. You do not need to understand the architecture to see whether a text states what was tested, under which conditions and with which limits. When the conclusion depends on terms nobody explains, take the question to a specialist before deciding.

Sources

• NIST AI Risk Management Framework
• OECD AI Principles
• Stanford HAI: AI Index
• ANPD: Autoridade Nacional de Proteção de Dados