A Nature Essay Questions the Term ‘Superintelligence’

⏱️ Reading time: 9 min

In October 2026, Nature published an essay that asks an uncomfortable question for the entire artificial intelligence industry: whether artificial superintelligence is still a term with scientific grounding or has become just a marketing label. The journal raises the question right as several labs publicly claim to be close to surpassing human intelligence in general tasks.

📑 En este artículo

1. TL;DR
2. What Is Artificial Superintelligence
3. What Happened
4. Context and History
5. Technical Details
6. Impact and Analysis
7. What’s Next
8. Frequently Asked Questions
   1. What exactly does Nature’s essay say about artificial superintelligence?
   2. Who defined the AGI levels this article uses as a reference?
   3. What is Anthropic’s Responsible Scaling Policy?
   4. Has today’s AI already reached artificial general intelligence?
   5. Why does this debate about general artificial intelligence matter to someone who doesn’t work in AI?
9. References

The essay, titled The science of superintelligence and why AI is still a useful word, does not deny that AI is advancing. It questions whether there is a measurable definition of “superintelligence” that serves for anything more than a headline.

TL;DR

• In October 2026, Nature published an essay questioning whether superintelligence has solid scientific grounding.
• In 2023, DeepMind proposed six levels of AI capability, from “no AI” to “superhuman,” in the Levels of AGI paper.
• Anthropic ties its Responsible Scaling Policy to measurable ASL levels, not to the word “superintelligence.”
• The EU AI Act sets the systemic risk threshold at 10^25 training FLOPs, not at a label.
• The essay does not deny AI’s progress: it argues that “artificial intelligence” remains a useful term.

What Is Artificial Superintelligence

Artificial superintelligence is the hypothetical ability of an AI system to outperform humans at practically any relevant cognitive task, including those that require judgment, creativity, and long-term planning, not just isolated tests of memory or calculation.

The term became popular within AI philosophy long before today’s language models existed, but it lacks a standard metric accepted by the entire scientific community. That is why Nature uses the question as a starting point: there is a need to distinguish between the ambition companies declare and what can actually be measured in a lab.

What Happened

The essay traces how, in just a few years, the industry’s language shifted from talking about language models to talking about general artificial intelligence and superintelligence almost as synonyms. The author argues that this drift did not come from a measurable breakthrough, but from the absence of a shared criterion: each lab defines the label its own way, and that definition shifts depending on what suits a funding round, a product announcement, or a regulatory debate.

The piece does not call for abandoning AI’s vocabulary. On the contrary, it argues that “artificial intelligence” as a term remains useful for science, even if imprecise, because it groups together a real set of techniques (deep learning, reinforcement learning, language models) that can indeed be studied and compared. The problem arises, according to the essay, when the discussion jumps straight to “superintelligence” without passing through an intermediate scale that someone can verify.

The publication also comes at a time when mainstream media, not just trade outlets, repeat the word in headlines without qualifying it. For a scientific journal, that repetition without empirical backing is exactly the kind of problem an editorial board cannot ignore: when a term becomes synonymous with progress that no one can measure, it stops serving as a scientific category and becomes a slogan.

Context and History

The idea of measuring AI by levels, rather than by a binary label, is not new. In 2023, researchers at Google DeepMind published Levels of AGI for Operationalizing Progress on the Path to AGI, a paper that proposed six capability levels, from Level 0 (no AI) to Level 5 (superhuman), crossed with a second scale of generality, from narrow tasks to fully general tasks. The proposal sought exactly what Nature is now demanding: a common yardstick for measuring progress that does not depend on a press release.

That framework coexists with opposing positions within the industry itself. OpenAI was founded with the stated mission of building safe and beneficial AGI, but it never set a public technical threshold for deciding when that point is reached. Anthropic took the opposite path: in his essay Machines of Loving Grace, its CEO Dario Amodei deliberately avoids the words AGI and superintelligence, proposing instead to talk about “powerful AI,” defined by what the system can do on concrete tasks, not by a vague comparison with human intelligence in general.

The label “artificial intelligence” itself was born in 1955, when John McCarthy decided to organize the Dartmouth workshop and chose that name for the new field, and since then it has gone through more than one “AI winter” in which a large part of the scientific community stopped using the term because of the discredit it had accumulated. The current discussion about superintelligence repeats, on a different scale, that same pattern: a word that promises too much ends up losing credibility even as the technology behind it keeps advancing.

Technical Details

The dispute over the term is not just semantic: in at least one case, a capability definition triggers safety commitments a company is bound to meet. Anthropic’s Responsible Scaling Policy ties each safety level (ASL, for AI Safety Level) to concrete evaluations of dangerous capabilities: how much a model helps design a biological weapon, replicate autonomously on infrastructure it does not control, or carry out a complex cyberattack without human supervision. If a model exceeds the threshold of an evaluation, the policy requires strengthening deployment safety before release, regardless of what that capability is called publicly.

flowchart LR
A["Level 0: No AI"] --> B["Level 1: Emerging"]
B --> C["Level 2: Competent"]
C --> D["Level 3: Expert"]
D --> E["Level 4: Virtuoso"]
E --> F["Level 5: Superhuman"]

DeepMind’s scale does not measure only the depth of a capability: it also measures generality. A model can be virtuoso at playing chess (high level) but narrow in generality, while a modern conversational assistant is more general but rarely gets past the competent level on demanding specific tasks. That double yardstick explains why two labs can look at the same model and reach opposite conclusions about whether it is already superintelligent.

There is no API or command that automatically returns which level a model is at. Placing it on DeepMind’s scale is a qualitative judgment: each research team compares its own benchmark results against the paper’s description of the six categories, and two teams can disagree on the outcome. That is precisely what Nature points to as the weak spot.

Neither DeepMind’s scale nor Anthropic’s policy solves the underlying problem: both rely on internal evaluations, carried out by the same company that later decides whether to release the model. Nature does not call for outsourcing those tests, but the essay does note that neither scale currently requires mandatory independent auditing.

Impact and Analysis

The lack of a common definition has practical consequences beyond the lab. Governments that allocate public funds to AI need a criterion to decide which projects qualify as frontier research and which are merely conventional applications labeled “artificial superintelligence” to attract budget.

For investors, the problem is different but related: a funding round valued at several billion dollars can rest on the promise of reaching AGI without any independent way to verify how far off that still is. Nature’s essay does not accuse any company of lying; it argues that, without a common metric, not even the company itself can know with certainty where it stands on the scale.

The European Union’s AI Act avoids the problem head-on: instead of relying on a word, it sets a measurable technical threshold, training with more than 10^25 FLOPs, to classify a model as systemic risk and require additional evaluations. That is the same logic the Nature essay defends: a figure anyone can verify carries more weight than a label whose meaning shifts depending on who uses it.

💭 Key takeaway: no industry safety policy today uses the word “superintelligence” as an official threshold; all of them, including Anthropic’s, rely on concrete, verifiable capability tests.

What’s Next

Nature is not the only scientific publication pushing in this direction. In the coming months, updates are expected to the capability frameworks of OpenAI, Google DeepMind, and Anthropic, which tend to get revised whenever a new generation of models appears. If editorial pressure holds, labs will likely start talking less about superintelligence in their announcements and more about concrete capability levels, if only to avoid being cited by an essay like this as an example of inflated language.

What is not going to happen soon is a formal agreement among companies on a single scale. Each lab has different incentives for defining progress its own way, and no scientific authority has the power to impose common terminology.

If you’re interested in the technical detail, read DeepMind’s original Levels of AGI paper: it’s just a few pages, and the levels table is easy to follow without any AI background.

Frequently Asked Questions

What exactly does Nature’s essay say about artificial superintelligence?

It argues that the term lacks an accepted scientific metric, while maintaining that “artificial intelligence” as a general category remains useful for research.

Who defined the AGI levels this article uses as a reference?

Researchers at Google DeepMind proposed them in 2023 in the paper “Levels of AGI for Operationalizing Progress on the Path to AGI.”

What is Anthropic’s Responsible Scaling Policy?

It is the internal framework that ties each risk level (ASL) to concrete evaluations of dangerous capabilities, such as assistance with biosecurity or autonomous cyberattacks, rather than to a label like superintelligence.

Has today’s AI already reached artificial general intelligence?

There is no scientific consensus on this. Different labs use different criteria, which is exactly the problem Nature’s essay points to.

Why does this debate about general artificial intelligence matter to someone who doesn’t work in AI?

Because that definition influences public funding decisions, what regulation requires, and how credible AI companies’ commercial announcements are.

References

• Nature: essay “The science of superintelligence and why AI is still a useful word.”
• arXiv: “Levels of AGI for Operationalizing Progress on the Path to AGI” (Google DeepMind, 2023).
• Anthropic: official announcement of the Responsible Scaling Policy.
• Dario Amodei: essay “Machines of Loving Grace” on powerful AI versus AGI and superintelligence.
• Wikipedia: general background on artificial general intelligence.

📱 Enjoy this content? Follow @programacion on Telegram for daily tech content in Spanish: quick summaries, fresh content every day.

Featured image: Foto de Mirella Callage en Unsplash