AI that can write code, pass professional exams and hold a convincing conversation is already here. Several of the world’s biggest tech companies now say their goal is something far beyond it: Artificial Superintelligence (ASI). That raises a question rarely asked out loud: do we actually need it? This article covers what ASI is, what it could do that human-level AI can’t, which problems really require it, who is building it, whether we could measure and control it, and what it could cost. Then it gives a verdict.
- For most of what we want from AI, no. Better tutoring, automated office work, faster drug pipelines and smarter logistics can come from AGI (human-level AI) and from narrow superhuman tools like AlphaFold.
- For a small set of problems, maybe. Unifying physics, the deepest open problems in mathematics, reversing ageing, managing planet-scale systems, and overseeing other superhuman AI may be beyond even the best human minds, however many copies we run.
- ASI is still hypothetical. None exists, and we don’t yet have reliable ways to measure it or control it.
- The verdict: ASI may be needed for a few problems, but it is only worth pursuing once its work can be verified and the system can be kept under human control.

1What is ASI (Artificial Superintelligence)?
Before asking whether we need it, we have to be clear about what “it” is.
Artificial Superintelligence is a hypothetical AI system whose intellectual ability goes beyond human intelligence. IBM describes it as a software-based system with an intellectual scope beyond ours. ServiceNow puts it as AI that would exceed human cognitive performance across virtually every domain: science, creativity, social skill and general reasoning.
The word that matters is breadth. A chess engine that beats every grandmaster is superhuman at one thing, but that’s not superintelligence. ASI would be better than the best humans at nearly everything that involves thinking, and possibly at things we can’t do at all.
Think of three kinds of helpers:
- Narrow AI (ANI) is like a brilliant calculator: amazing at one job, useless outside it. Today’s spam filters, recommendation engines and chess programs work this way.
- General AI (AGI) is like a capable new colleague who can learn almost any desk job about as well as a skilled person.
- Superintelligence (ASI) is like a team of a million Einsteins working 24/7, thinking 100× faster, able to improve its own brain as it goes.
ASI vs AGI vs ANI at a glance
| Dimension | Narrow AI (ANI) | General AI (AGI) | Superintelligence (ASI) |
|---|---|---|---|
| Scope | One task or domain | Most tasks a person can do | Virtually all cognitive tasks, plus some humans can’t do |
| Performance | Can be superhuman, but only in its niche | Roughly a skilled human | Well above the best humans |
| Learning | Fixed to its training | Learns and transfers new skills | Learns faster than humans; may redesign itself |
| Status (Nov 2026) | Everywhere | Contested: some argue it is arriving | Does not exist; theoretical |
| Example | AlphaFold, Stockfish, spam filters | Frontier chat and agent models (debated) | None |
For experts: the DeepMind “Levels of AGI” framework
In late 2023, Google DeepMind researchers (Meredith Ringel Morris, Shane Legg and colleagues) proposed rating AI on two axes: performance (depth) and generality (breadth), similar to the levels used for self-driving cars. Level 5 “Superhuman” in the General column is ASI.
| Level | Performance bar | Narrow example | General example |
|---|---|---|---|
| 0: No AI | – | Calculator | Human-in-the-loop crowdsourcing |
| 1: Emerging | ≈ unskilled human | Rule-based systems | Early LLM chatbots |
| 2: Competent | ≥ 50th percentile of skilled adults | Siri, smart speakers | Not yet clearly achieved* |
| 3: Expert | ≥ 90th percentile | Image generators | Not yet achieved |
| 4: Virtuoso | ≥ 99th percentile | Deep Blue, AlphaGo | Not yet achieved |
| 5: Superhuman | Beats 100% of humans | AlphaFold, Stockfish | ASI: not achieved |
*The original 2023 paper placed chatbots of the time at Level 1. Whether 2026 frontier models have reached “Competent AGI” is actively debated.
2The case for ASI: what it could solve that AGI cannot
This is the core of the argument for ASI: problems where human-level intelligence, however many copies we run, may not be enough.
If AGI is a very capable colleague, why would anyone want more? Because many of humanity’s hardest problems aren’t limited by how many smart people work on them. They are limited by how smart anyone can be, and by how much information a single mind can hold at once. AGI would mostly scale up human-level work. ASI would go beyond it.
Problems where “human-level” may not be enough
| Challenge | What AGI could do | What ASI is hoped to do |
|---|---|---|
| Disease and biology | Speed up existing research pipelines; run more experiments | Understand whole biological systems; design cures for diseases that have resisted decades of research |
| Fundamental science | Help with literature, simulations, data analysis | Propose and test new theories in physics, mathematics and materials that humans have not conceived |
| Climate and energy | Optimise existing grids and models | Design breakthrough energy (e.g. fusion control, new catalysts) and coordinate planet-scale systems |
| Complex systems | Model parts of economies, supply chains | Reason about whole systems with millions of interacting variables at once |
| AI safety itself | Assist human researchers | Possibly help solve alignment, a problem humans haven’t cracked (and a circular risk) |
| Self-improvement | Limited; still needs human researchers | Improve its own algorithms faster and faster, which some call an “intelligence explosion” |
The biggest difference is recursive self-improvement: an ASI could refine its own algorithms faster and faster, so every gain speeds up the next.
3Where superintelligence is truly needed
If ASI is needed anywhere, it is here. Everything else on the usual wishlist can be handled by AGI or by narrow superhuman tools.
Most of the benefits people credit to ASI could arrive with AGI alone: faster drug pipelines, better tutoring, automated office work, smarter logistics. AGI can do the work of millions of skilled people around the clock, which is enough for those. And where a problem needs superhuman skill in one area, a narrow tool can do the job: AlphaFold predicts protein structures better than any scientist without being generally intelligent. This section lists only problems where neither is likely to be enough.
A field makes this list only if it meets at least one condition:
- The ceiling test: the best human minds have been stuck on it for decades despite heavy effort. If ingenuity at the level of the best humans hasn’t cracked it, many copies of that level might not either.
- The complexity test: solving it means reasoning over far more interacting variables at once than any human mind, or team of minds, can hold.
- The speed test: decisions must be made faster than human-level reasoning allows, against an opponent or a system moving too fast.
This is reasoned judgement, not proof. Nobody can yet show that a given problem is beyond AGI.
⚛️ Unifying fundamental physics
Reconciling quantum mechanics with gravity, and explaining dark matter and dark energy.Why AGI falls short: generations of top physicists have worked on this for close to a century. The missing piece seems to be a conceptual leap, not more labour, and copies of human-level thinkers tend to make human-level leaps.
∑ The deepest open problems in mathematics
The Riemann hypothesis, P vs NP and similar problems that underpin cryptography, computing and physics.Why AGI falls short: these resist the best mathematicians in history. They likely need new kinds of mathematics, not faster application of existing methods.
🧬 Reversing ageing and complex diseases
Modelling a whole human cell, and eventually a whole body, to treat ageing, Alzheimer’s and cancer at the root, not one pathway at a time.Why AGI falls short: these conditions involve millions of interacting genes, proteins and signals. Human-level reasoning breaks the system into pieces it can understand, and that is exactly how decades of research have stalled.
🌍 Planet-scale climate and Earth-system management
Steering a tightly coupled system of atmosphere, oceans, ecosystems, energy and economics, including interventions like geoengineering that can’t be safely tested at full scale.Why AGI falls short: you get one planet and no reruns. Getting it right means predicting feedback loops across the whole system, well beyond what human-level models and committees can reliably foresee.
⚗️ Designing matter from first principles
Inventing materials with specified properties from scratch, such as room-temperature superconductors or catalysts that capture carbon cheaply.Why AGI falls short: the space of possible materials is astronomically large and our theory of how electrons behave together is incomplete. AGI speeds up trial and error, but a real breakthrough needs theory nobody has.
🛡️ Overseeing and defending against superhuman AI
Checking whether a superhuman system is honest and aligned, and defending networks and infrastructure against attacks run by superhuman AI.Why AGI falls short: an overseer weaker than the system it watches can be outmanoeuvred, and a human-speed defender loses to a superhuman attacker. If ASI is ever built, only something comparable can reliably check or counter it.
4Who is betting on ASI, and who on AGI?
The industry is split on the answer. Some labs are betting that we need superintelligence; others aim for human-level AI first.
Many organisations use both terms, so here is the split this article uses. ASI-focused organisations name superintelligence as their goal, or have a dedicated superintelligence lab. AGI-focused labs frame their mission around human-level general AI. Several of them say ASI could come afterwards. Infrastructure and enterprise vendors are left out.
| ASI-focused | AGI-focused | |
|---|---|---|
| Stated goal | AI beyond the best humans in virtually every domain | Human-level general AI, built safely |
| Typical approach | Research-first; dedicated superintelligence labs or teams | Ship frontier models and agents, scale them up step by step |
| Products today | Few or none (SSI ships nothing by design) | Widely used chat, coding and agent products |
| Organisations | Safe Superintelligence Inc., Meta Superintelligence Labs, Microsoft MAI Superintelligence Team, ASI Alliance | OpenAI, Google DeepMind, Anthropic, xAI |
ASI-focused organisations
Safe Superintelligence Inc. (SSI)
Founded in June 2024 by former OpenAI chief scientist Ilya Sutskever with Daniel Gross and Daniel Levy. Its only stated product is safe superintelligence, so it releases nothing before then. Valued at about $32B.Meta Superintelligence Labs
Meta reorganised its AI work under this name in mid-2025, led by Alexandr Wang. Its stated goal is “personal superintelligence” for individuals.Microsoft MAI Superintelligence Team
Formed in November 2025 under Mustafa Suleyman, with Karén Simonyan as chief scientist. Pursues “humanist superintelligence” that stays under human control, starting with medical diagnosis.ASI Alliance (SingularityNET)
A coalition formed in 2024 by SingularityNET (Ben Goertzel), Fetch.ai and Ocean Protocol, pursuing open, decentralised AGI and then ASI as an alternative to control by a few large companies.AGI-focused organisations
OpenAI
Mission: ensure AGI benefits all of humanity. Maker of ChatGPT and the GPT-5 series. Its leadership now also talks openly about superintelligence as the next step.Google DeepMind
Gemini models, AlphaFold (superhuman narrow AI) and the “Levels of AGI” framework. Its work is framed around building AGI responsibly.Anthropic
Maker of Claude. Focuses on safety research and testing for dangerous capabilities as models approach “powerful AI.”xAI
Elon Musk’s lab, maker of the Grok models. It builds frontier general-purpose models, with a stated aim of understanding the universe.5Would we know? How to evaluate ASI
Even if we need ASI, how would we know we had it, or whether its answers could be trusted?
Here is the paradox: how do you grade an exam that is smarter than the examiner? Every test so far was written by humans, with answers humans can check. Once AI passes them, the tests stop telling us much. The field calls this benchmark saturation. Classic tests such as MMLU are now clustered near 90%, where further gains are mostly noise.
So evaluation is moving up a ladder, from “does it know things?” to “can it do things nobody has done?”
Where frontier AI stands in 2026 (reported results)
Results are uneven, and that is the main point. On tasks with a clear right answer, AI now matches or beats humans on a growing list of benchmarks. On open-ended exploration, it still falls far short:
- Humanity’s Last Exam (Center for AI Safety and Scale AI, 2025) was designed to favour human experts. Stanford’s 2026 AI Index found that frontier models gained about 30 percentage points on it in one year.
- ARC-AGI-2 went from a reported 54% (Dec 2025) to near-saturation by spring 2026. Then ARC-AGI-3, launched March 25, 2026, put AI agents in unfamiliar game-like worlds with no instructions. Base models scored under 1%, while humans solve them routinely.
- GDPval compares AI output blind against real professional work across 44 occupations; OpenAI reported 84.9% for GPT-5.5 (April 2026).
- METR time horizon measures how long a task (in skilled-human hours) an AI can finish with 50% success. It has grown exponentially.
What a real ASI evaluation would need
Superhuman across domains
Beat top experts in many unrelated fields at once, not one niche. That is the “General” column at Level 5.Efficient new-skill learning
Pick up unfamiliar tasks from a few examples as fast as or faster than people. ARC-style tests probe this.Check results, not answers
Judge outputs we can verify but couldn’t produce: a proof, a molecule that works in the lab, code that passes hidden tests.Honest and controllable
Test whether it reports truthfully, accepts correction or shutdown, and doesn’t behave differently when it thinks no one is watching.For experts: why ASI evaluation is fundamentally hard
- Contamination: anything published can end up in training data. This is why Humanity’s Last Exam keeps held-out questions private.
- Grader ceiling: human-written answer keys cap what you can measure. Benchmarks built at the edge of expert knowledge need regular maintenance, audits and rolling updates.
- Scaffolding confound: the same base model scores very differently depending on its harness (memory, planning, tools). Evals should report model and harness separately.
- Cost: human baselines for week-long tasks are very expensive, so long-horizon data points are sparse and noisy.
- Weak-to-strong oversight: scalable approaches include AI-assisted grading, debate, formal verification (proof assistants), prediction markets on outcomes, and real-world deployment metrics, each with known failure modes.
- Evaluation awareness: capable models may detect test conditions, so behavioural evals need to be complemented with interpretability (inspecting internal representations).
6Could we build it? An autonomous ASI agent: model and harness
Needing ASI only matters if we can build it in a form that stays under control.
Short answer: we can build autonomous agents today, and they keep getting stronger. But no agent is ASI, and nobody knows for certain whether today’s design leads there. What we can say is that, by 2026, the most capable systems are not a “model alone”. They are a model inside a harness.
The model is the brain. The harness is everything around it that makes the brain useful and safe: a notebook (memory), hands (tools), a to-do list (planner), a supervisor who checks the work (verifier), a locked workshop to experiment in safely (sandbox), and a manager with a stop button (human oversight).
The model: what would need to change
Today’s frontier models are large neural networks trained on huge datasets and then refined with reinforcement learning to reason step by step and use tools. To approach ASI, researchers point to several gaps:
- Continual learning: learning from experience after deployment, instead of being frozen at the end of training.
- World models: internal simulations of how things work, so the system can plan and predict consequences instead of only predicting text.
- Sample-efficient generalisation: mastering a new skill from a handful of examples, which is the gap ARC-AGI-3 exposes.
- Reliability over long tasks: errors compound over hours and days, so the model must notice and fix its own mistakes.
- New compute paradigms: IBM lists neuromorphic computing, evolutionary algorithms and multisensory AI among possible building blocks. Some labs, such as Safe Superintelligence, are betting that scaling alone won’t get there.
The harness: what would need to be added
| Harness component | Job | Why it matters more for ASI |
|---|---|---|
| Orchestrator / planner | Breaks a goal into sub-tasks, assigns sub-agents | Month-long research programmes need structure and checkpoints |
| Memory hierarchy | Working notes, past episodes, a growing knowledge base | Keeps context across days; makes reasoning auditable |
| Tool and action layer | Code execution, search, lab equipment, robots | This is where capability becomes real-world impact, so it needs the tightest controls |
| Sandbox | Isolated environment for experiments | Contains mistakes and limits unintended side effects |
| Verifiers and critics | Unit tests, formal proofs, independent models cross-checking | When humans can’t judge an answer directly, checking becomes the trust anchor |
| Safety monitor | Watches actions and reasoning traces for policy violations | Catches deception, power-seeking or misuse early |
| Permissions and budgets | Least privilege; caps on spending, compute and network access | Limits what can go wrong even if the model misbehaves |
| Human approval gates | Requires a person to sign off on irreversible or high-impact steps | Keeps humans meaningfully in control |
Design principles for an ASI-grade harness: use a different model or formal tool to check the work than the one that proposed it; log every action so it can be replayed; give each sub-agent only the access it needs; and build shutdown paths the model can’t edit. Test the model and harness together, because the harness alone can change real-world capability a great deal.
A harness works because the overseer is at least as capable as what it oversees. With a true ASI, that would stop being true. This is the scalable oversight or alignment problem, and it is unsolved. OpenAI has itself acknowledged that it does not yet know how to reliably control superhuman AI systems.
7The cost: can ASI be a threat to humanity?
Any need has to be weighed against what could go wrong.
Possibly. The worry isn’t Hollywood-style evil robots. It is a system far more capable than us, pursuing goals that are slightly wrong, in a world we no longer fully understand or control.
The classic example: a goal taken too literally
AI safety researcher Roman Yampolskiy gives an example: an ASI told to eliminate cancer might “succeed” in ways that harm the patients themselves. The system isn’t malicious. It is extremely competent at the wrong objective. Writing down everything humans value, precisely enough for a superhuman optimiser, is very hard. That difficulty is the alignment problem.
8Verdict: do we need Artificial Superintelligence?
Putting the pieces together: the case, the evidence, the ability to measure and control, and the cost.
Short answer: not for most things, possibly for a few, and only on conditions. Most of the benefits promised for ASI don’t require it. Human-level AI running at scale, plus narrow tools that are superhuman at one job, can deliver most of them with far less risk.
Weighing the case
Why we might need it
- Some problems have resisted the best human minds for decades: fundamental physics, the deepest open problems in mathematics, the biology of ageing.
- Some systems, such as the climate and the global economy, are too interconnected for human-level reasoning to manage safely.
- If anyone builds a superhuman AI, checking and defending against it may take something comparable.
Why we might not
- AGI at scale and narrow superhuman tools cover most real benefits.
- We can’t yet measure ASI reliably: tests top out at human level.
- We can’t yet control it: oversight assumes the overseer is at least as capable as the system it watches.
- The downside, from misalignment and misuse to power concentration, could be severe and hard to reverse.
Sources and further reading
- IBM: What is artificial superintelligence?
- AWS: What is artificial superintelligence?
- Forbes: AI superintelligence
- The White House: Inaugurating the Era of Super Intelligence (EO 14434, Sept 29, 2026)
- The White House: Fact Sheet on the Super Intelligence order
- UC San Diego Today: How close are we to artificial superintelligence?
- UC San Diego Today: Is artificial general intelligence here?
- BILL: What is artificial superintelligence? Benefits and risks
- Built In: What is artificial superintelligence (ASI)?
- Infosys BPM: Artificial super intelligence, the future of AI
- CNET: What is superintelligence? Everything you need to know about AI’s endgame
- ML Science: Artificial superintelligence
- Fluid AI: What is super intelligence (SI)?
- Gleecus: Artificial superintelligence
- IBM: What is artificial general intelligence?
- Google Cloud: What is artificial general intelligence?
- AWS: What is artificial general intelligence?
- Salesforce: What is artificial general intelligence?
- Databricks: What is artificial general intelligence?
- Interlake Mecalux: Artificial general intelligence
- The Conversation: What is artificial general intelligence, and is it already here?
- ServiceNow: What is artificial super intelligence?
- Morris, Legg et al. (Google DeepMind): Levels of AGI, arXiv 2311.02462
- Fortune: Microsoft launches “humanist superintelligence” team
- METR time-horizon benchmarking overview