Superhuman AI: Every Time AI Beat the Best Humans (1997–2026)

Superhuman AI beats the best humans at a specific task. See every milestone from Deep Blue to AlphaGo, AlphaFold and IMO gold, plus where it still fails.

By Izhar Ahmad Danish · · 5 min read

Superhuman AI: Every Time AI Beat the Best Humans (1997–2026): Time of Superintelligence cover image
On this page
  1. What is superhuman AI?
  2. Timeline: when AI became superhuman
  3. What are the best-known examples of superhuman AI?
  4. Is superhuman AI the same as superintelligence?
  5. Can humans still beat superhuman AI?
  6. Why superhuman AI matters for the road to superintelligence
  7. The bottom line
  8. Frequently asked questions
  9. Sources

Superhuman AI is not a future idea. It arrived in 1997, when IBM’s Deep Blue beat the world chess champion. Since then AI has overtaken the best humans at Go, poker, protein structure prediction and, by 2025, olympiad mathematics. Each of these systems is superhuman at one thing, which is what separates them from the hypothetical super intelligence that beats us at everything.

Short answer: Superhuman AI is AI that outperforms the best humans at a specific task. Real examples include Deep Blue (chess, 1997), AlphaGo (Go, 2016), Pluribus (six-player poker, 2019) and AlphaFold (protein structures, 2020). These systems are narrow: superhuman in their domain and useless outside it. No AI is yet superhuman at everything.

What is superhuman AI?

Short answer: AI that beats every human, including the world’s top experts, at a defined task.

Google DeepMind’s 2023 paper Levels of AGI gives the clearest working definition. It rates AI on two axes: how good it is (performance) and how many tasks it can do (generality). The top performance level, “Level 5: Superhuman,” means outperforming 100% of humans. On the narrow side of the table, the paper places AlphaFold and the chess engine Stockfish at that level. On the general side, the superhuman cell, which the paper calls artificial superintelligence, is empty.

That table explains most of the confusion about the term. “Superhuman” describes a level of skill; “superintelligence” describes that skill across virtually every domain.

Timeline: when AI became superhuman

Short answer: chess in 1997, Go in 2016, multiplayer poker in 2019, protein structures in 2020, olympiad math in 2025.

Year System (developer) Domain What happened
1997 Deep Blue (IBM) Chess Beat world champion Garry Kasparov 3.5–2.5 in a six-game match ending 11 May 1997
2016 AlphaGo (DeepMind) Go Beat Lee Sedol 4–1 in Seoul, 9–15 March 2016
2017–18 AlphaZero (DeepMind) Chess, shogi, Go Learned all three from the rules alone by self-play and beat the strongest programs, including Stockfish
2019 Pluribus (Carnegie Mellon / Facebook AI) Six-player no-limit Texas hold’em Beat elite professionals; the first superhuman result in multiplayer poker
2020 AlphaFold 2 (DeepMind) Protein structure prediction Predicted structures accurately enough to transform biology; Nobel Prize in Chemistry 2024
2025 Gemini Deep Think (Google DeepMind) and an OpenAI model Olympiad mathematics Gold-medal level at the IMO; Gemini solved 5 of 6 problems for 35 points

Sources: HISTORY, Wikipedia, DeepMind, Science, Google. Full list below.

Two things stand out. The gaps are shrinking, and the domains are moving from games with fixed rules toward open-ended science and mathematics.

What are the best-known examples of superhuman AI?

Short answer: AlphaGo for games, AlphaFold for science, and Pluribus for hidden-information strategy.

AlphaGo and AlphaZero

Go was long considered too complex for computers: there are too many possible moves to search them all. AlphaGo combined deep neural networks with tree search and self-play reinforcement learning, and won four of five games against Lee Sedol in 2016. Its successor AlphaZero started with nothing but the rules. DeepMind reports it first outperformed Stockfish after about 4 hours of chess training, and beat the Lee Sedol version of AlphaGo after 30 hours of Go training.

AlphaFold

Proteins fold into 3D shapes that determine what they do, and working out a shape used to take years of lab work. AlphaFold 2 predicts these structures from amino acid sequences. The Nobel committee said it has made it possible to predict the structure of virtually all of the roughly 200 million known proteins. Demis Hassabis and John Jumper shared half of the 2024 Nobel Prize in Chemistry for it.

Pluribus

Poker hides information and involves bluffing, which makes it harder than chess in a different way. Earlier poker AIs were superhuman only head-to-head. Pluribus, by Noam Brown and Tuomas Sandholm, beat top professionals in six-player games in 2019, mostly by learning from games against copies of itself.

IMO gold (2025)

In July 2025, an advanced version of Gemini Deep Think solved five of six International Mathematical Olympiad problems in natural language, scoring at gold-medal level. OpenAI announced a similar result days earlier. Strictly speaking this is “top human level” rather than superhuman, since some human contestants scored higher, but it is the first time a general-purpose model reached that level.

Is superhuman AI the same as superintelligence?

Short answer: no. Superhuman AI is narrow; superintelligence would be general.

Superhuman AI Superintelligence
Scope One task or a small set Virtually all intellectual tasks
Exists today? Yes No
Examples Deep Blue, AlphaGo, AlphaFold, Pluribus Hypothetical; see our super intelligence examples
Can it transfer skills? Rarely: AlphaGo cannot play poker Yes, by definition
Main risk Misuse, over-reliance Loss of human control

For a broader overview of what “super AI” means in everyday speech, see super AI. For the step in between, read what is AGI and AGI vs ASI.

Can humans still beat superhuman AI?

Short answer: sometimes, by exploiting blind spots rather than by playing better.

In 2023, researchers led by Tony Wang showed that an adversarial program could beat KataGo, an open-source superhuman Go AI, by luring it into positions it misjudged. Human amateurs could then learn the trick and beat KataGo themselves. The lesson is that “superhuman” on average does not mean robust everywhere, a point that matters a lot more for AI systems deployed in the real world than for board games.

Why superhuman AI matters for the road to superintelligence

Short answer: it shows how quickly AI can go from weaker than humans to far stronger in a domain.

AlphaGo went from beating the European champion in late 2015 to beating Lee Sedol a few months later, and AlphaZero reached superhuman chess in hours. If the same jump happened in general skills such as research or engineering, the gap between AGI and superintelligence could be short. That is why pioneers such as Geoffrey Hinton treat narrow superhuman results as early warnings rather than curiosities.

The bottom line

Superhuman AI already exists and has for almost 30 years, but only in narrow slices: games, protein structures, competition math. The open question is not whether AI can beat the best humans; it is whether one system can do it everywhere at once.

Enjoyed this? Get the SI Briefing

One short email a week: what changed on the road to super intelligence, and why it matters. No spam, unsubscribe anytime.

Frequently asked questions

What is superhuman AI?

Superhuman AI is an AI system that performs a specific task better than any human, including the world’s best experts. Chess engines, AlphaGo and AlphaFold are examples. It is narrow: being superhuman at Go says nothing about other skills.

Does superhuman AI exist?

Yes, in narrow areas. AI has been superhuman at chess since Deep Blue beat Garry Kasparov in 1997, at Go since AlphaGo in 2016, and at predicting protein structures since AlphaFold in 2020. No AI is superhuman across all tasks.

What is the difference between superhuman AI and superintelligence?

Superhuman AI beats humans at one task or a few. Superintelligence would beat the best humans at virtually every intellectual task. All superintelligence is superhuman, but almost all superhuman AI today is narrow.

Is ChatGPT superhuman?

In some respects. It writes faster and knows more facts than any person, but it still makes errors that experts would not. Google DeepMind’s Levels of AGI paper rates chatbots of its time as only “Emerging AGI”.

Can humans still beat superhuman AI?

Sometimes. In 2023 researchers found adversarial strategies that let amateur players beat KataGo, a superhuman Go program, showing that superhuman systems can have blind spots no human expert would have.

8 Sources

  1. Morris et al., Levels of AGI for Operationalizing Progress on the Path to AGI (Google DeepMind)
  2. HISTORY: Deep Blue defeats Garry Kasparov in chess match (11 May 1997)
  3. AlphaGo versus Lee Sedol (Wikipedia)
  4. Google DeepMind: AlphaZero, shedding new light on chess, shogi and Go (Dec 2018)
  5. Brown & Sandholm, Superhuman AI for multiplayer poker (Science, 2019)
  6. Google: Demis Hassabis and John Jumper awarded Nobel Prize in Chemistry (Oct 2024)
  7. Google DeepMind: Gemini Deep Think achieves gold-medal standard at the IMO (21 Jul 2025)
  8. Wang et al., Adversarial Policies Beat Superhuman Go AIs (arXiv 2211.00241)
Before you go

Super intelligence is moving fast.

Don’t miss what happens next

Get one short email a week with what changed and why it matters. No spam. Unsubscribe anytime.