Superintelligence Hype Checklist: Claims Leaders Should Question

Superintelligence hype is the habit of presenting a possible future level of AI capability as a present fact, then using it to justify a purchase, a deadline or a fear. The practical defense is to ask what was measured, on which task, by whom, and what happens in your operation if the claim turns out to be wrong. Credible statements name a scope and a test; marketing claims name a destination and a mood.

Key takeaways

• Hype is a claim about the future presented as a fact about the present. Ask what can be shown today, on which task, under which conditions.
• Credible statements name a scope, a method and a limit. Marketing claims name an outcome and a feeling.
• Your decisions should hold up whether or not superintelligence arrives on anyone's schedule. Build for the capability you can observe.
• Test claims on your own work, with your own data and an agreed definition of success, before you change a process or a budget.
• Keep people in charge of identity, permissions and approvals. A system described as superintelligent still needs boundaries.

What counts as superintelligence hype?

Superintelligence usually means a system that would outperform people across nearly all intellectual work. Serious researchers disagree about whether, when and how that might happen, and this article takes no side. The problem for a business leader is narrower. Statements about that possible future leak into sales material about products that exist today, and the leak is where the risk sits. If you want the plain definition first, read what superintelligence means for business leaders.

Hype tends to arrive in three forms. The first is borrowed prestige: a result from a research setting is quoted to suggest that a different product has the same ability. The second is timeline pressure: act now or fall behind, with no description of what falling behind would cost you specifically. The third is category inflation: ordinary automation, a chat window or a scripted workflow is described as general intelligence. All three push you to decide before you can verify.

Not every bold statement is hype. A researcher saying that capability may advance quickly is offering a forecast, and forecasts can be reasonable. It becomes hype when the forecast is used in place of evidence for something you are being asked to sign.

Which claims should leaders question first?

These claims appear most often in pitches and keynotes. Each one has a follow-up question that turns it into something testable.

• 'It can do any job in your company.' Ask: which role, with which tools, under which approvals?
• 'It performs at or beyond human level.' Ask: on which tasks, measured how, and compared with which people?
• 'It improves itself.' Ask: who approves each change, what is logged, and can you roll it back?
• 'No setup required.' Ask: where do its memory, context and permissions come from, and who maintains them?
• 'Competitors will be left behind within months.' Ask: what is the specific cost of waiting for our business?
• 'Fully autonomous.' Ask: which actions can it take without a person, and what stops it when it is wrong?
• 'Safe by design.' Ask: which controls exist, how were they tested, and who reviews incidents?

None of these questions assume bad faith. Most people repeating these claims believe them. The questions simply move the conversation from belief to inspection. For structured language on risk, trustworthiness and accountability, the NIST AI Risk Management Framework is a useful reference when you write your own questions.

A ten-question checklist for any advanced AI claim

Use this list in vendor meetings, board discussions and internal proposals. Write the answers down. An answer given aloud and never repeated in writing is not one you can hold anyone to.

• What exactly is the claim, in one sentence, with the task named?
• Is it an observation (something that happened) or a forecast (something expected)?
• Who measured it, and are they independent of the seller?
• What test was used, and could we run the same test on our own work?
• What data and tools did the system have during the test, and will we have the same?
• How often did it fail, and what did the failures look like?
• What does a person still approve, and what happens when the system is unsure?
• What result would show the claim is wrong?
• What does it cost us if the claim is false, and can we reverse the decision?
• Who in our company is accountable for the outcome?

Count the questions that receive a specific, checkable answer. If most come back as general statements about the future, treat the material as marketing and continue only with a small, reversible step. The article on how to test an AI vendor's claims before you commit shows how to turn the fourth and fifth questions into a short trial.

An illustrative vendor pitch, taken apart

Here is an illustrative scenario, not a real case. An operations director at a mid-sized distributor receives a deck that calls a platform 'a step toward superintelligence' and says it will 'run operations autonomously'. Three slides describe the future. One slide describes the product.

She applies the checklist. First she names a task: answering supplier delay inquiries and updating order status. Then she asks the vendor to run the system on a sample of past inquiries with the real answers hidden, and she compares its output with what her staff actually did.

The system handles routine inquiries well. When asked what it does with an inquiry that involves a contract exception, the vendor answers in general terms about future versions. That gap is the useful finding. She now knows the real boundary: routine status work is a fit, and exceptions need a person.

She proposes a pilot for that one role, with a named approver for exceptions and a success measure agreed before it starts. Writing success criteria for the pilot in advance is what keeps the later conversation honest. The superintelligence slides play no part in the decision. The hype did not need to be refuted. It simply became irrelevant.

What do credible statements look like?

Credible statements about advanced AI share a few habits. You can look for them the same way you would look for good bookkeeping in a supplier.

• They name the task, the conditions and what the system was compared with.
• They separate what was observed from what is expected.
• They state limits and known failure modes without being asked.
• They describe the evaluation well enough that someone else could repeat it.
• They still hold when you shrink them to your own context.
• They allow for being wrong, with a stated way to find out.

A forecast is an opinion with a confidence level; a measurement is a result with a method.

Evaluation practice is the working vocabulary here. Guides such as the OpenAI evals guide and Anthropic's writing on evals for agents explain how teams define test cases, grade outputs and track regressions. You do not need to run these yourself, but reading one shows what a real evaluation contains, which makes thin claims easier to spot. Broad, regularly published reporting such as the Stanford HAI AI Index is good context on how the field is moving, though it describes the field and cannot tell you whether a specific product will work in your operation. Vendor case studies deserve their own scrutiny, and how to evaluate AI case studies without being misled covers that.

How do you plan without betting on a forecast?

The goal is a position that holds up whether capability advances slowly or quickly. A short plan does this better than a prediction.

• Step one: pick one role that is useful today, such as triaging inbound requests or preparing weekly summaries, and write down what doing it well means.
• Step two: list the memory, tools and permissions that role needs, and nothing more.
• Step three: set approval boundaries, meaning which actions need a person every time.
• Step four: choose an observable success measure and a review date, so progress is seen rather than claimed.
• Step five: treat model changes as a controlled event, tested against the same cases before they reach production.
• Step six: revisit the plan on a regular rhythm, and let new capability earn its place with evidence.

Be Human's managed approach follows this shape. A first engagement defines one useful role, the memory and tools it needs, approval boundaries and an observable success criterion, so claims are replaced by something you can watch. Managed operations then include monitoring, recurring evaluations, permission review and controlled improvements. Identity, memory, channels, tools and permissions stay under human control, and a stronger model is a component that can be swapped in through a change process rather than a promise that changes your risk overnight.

Frequently asked questions

Is superintelligence real or just marketing? The term describes a hypothetical level of capability, not a product you can buy. Researchers disagree about whether and when it could arrive, so treat anyone who speaks with certainty as a seller rather than a source. Judge today's products on what they demonstrably do today.

Should we wait for more capable AI before investing? Waiting is also a decision, and it has its own cost. A better approach is to start with one narrow role whose value does not depend on future models, and to design it so a stronger model can be swapped in later. That keeps your options open without staking the company on a forecast.

How can a non-technical leader check a benchmark claim? Ask for the task description, a sample of inputs and outputs, and a way to run the same task on your own material. If the seller cannot share these, the benchmark is a talking point, not evidence. A trusted technical colleague or an outside advisor can review the method in a short session.

Does a vendor mentioning superintelligence mean we should avoid them? No. Many credible teams discuss long-term possibilities. What matters is whether the proposal in front of you stands on its own: a defined role, a test, clear limits and a way to stop. If it only works when you accept the forecast, step back.

Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.