Type “artificial general intelligence examples” into a search box and you get two contradictory answers. One list points at ChatGPT, AlphaFold and Gemini and implies AGI has quietly arrived. Another insists no such example exists, full stop. Both are half-right, and the gap between them is where the real story sits.
The useful split is between systems that show flashes of generality and scenarios that describe what genuine AGI would actually do. Look at the first group to see how far the technology has come. Look at the second to understand why researchers still hesitate.
Why you won’t find a confirmed AGI example
AGI, as most labs define it, means a system that can learn and perform at roughly human level across essentially any cognitive task, including tasks nobody trained it on. No company claims to have built one. OpenAI’s charter, Google DeepMind’s published research and every serious academic definition treat it as a target rather than a product.
What exists is narrow AI, and plenty of it. Spam filters, recommendation engines, protein folders and chatbots all do particular things extremely well. The spectrum from spam filters to superintelligence is worth keeping in mind, because most systems mistaken for AGI sit far closer to the narrow end than the headlines suggest. If the terminology itself is fuzzy for you, this breakdown of what artificial general intelligence actually means clears up most of the confusion.
Real systems that show glimmers of generality
Large language models juggling dozens of domains
In 2023, GPT-4 passed the Uniform Bar Exam around the 90th percentile, scored near the top of the SAT verbal section, and wrote code that shipped to production at startups. Same model, same weights, no retraining between tasks. That range is genuinely new.
The cracks show up under pressure. Ask a language model to reason about a physical scene it has never seen and it can still stumble in ways a twelve-year-old wouldn’t. Breadth without reliability is the pattern: impressive across many domains, brittle inside each one.
AlphaFold: superhuman, but narrow
AlphaFold predicted structures for roughly 200 million proteins, a body of work that shared the 2024 Nobel Prize in Chemistry. No human can match it at protein folding. It also cannot plan a clinical trial, write a grant application or notice that a colleague’s hypothesis is flawed. This is the clearest illustration of why raw capability and generality are different things.
Gato: one network, 604 tasks
DeepMind’s Gato, unveiled in 2022, was trained on 604 separate tasks with a single transformer. It captioned images, played Atari, stacked blocks with a robot arm and controlled a simulated robot. It was rarely best-in-class at any of them. What made it notable was the breadth held inside one set of weights.
Agentic models that chain steps together
By 2025, models were routinely running coding sessions lasting hours, browsing the web, filling in forms and correcting their own errors mid-task. METR, a research nonprofit that measures this, found the length of task a model could complete with 50% reliability had been doubling roughly every seven months. That curve is one of the more honest measures of progress we have.
Robots that turn instructions into movement
Google’s RT-2 and follow-on systems map plain-language instructions directly to motor actions, and humanoid projects at Figure and elsewhere have pushed the same idea into warehouses. They remain fragile. Ask one to unpack a dishwasher in an unfamiliar kitchen and it will likely fail on something a five-year-old manages without thinking.
What a genuine AGI example would look like
Since nothing on the shelf qualifies, the more useful exercise is describing what would. These scenarios are hypothetical, but each one tests something current systems cannot do:
- The clinic test. A doctor in a rural practice hands over an unfamiliar scan format, a drug interaction database she has never used and a patient who speaks a language the system was not trained on. It reads all three, asks sensible clarifying questions, and drafts a referral by the end of the afternoon.
- The small business test. A two-person bakery turns over payroll, supplier negotiation, a hiring decision and a tax filing. The system handles each one, flags the legal risk in the contract it was asked to sign, and remembers the owner’s preferences three months later.
- The novel problem test. A research group asks for a proof nobody has published, or an experimental design that would distinguish between two competing theories. The answer needs to be right, not merely plausible.
- The job-change test. Point the system at a profession it has never encountered and give it a week. A general intelligence would be productive in days. A narrow one needs a new training run.
Behaviours to watch instead of product names
Tracking which chatbot ranks highest tells you less than tracking a handful of capabilities. These are the signals that separate real generality from clever marketing, and this piece on the traits that make an AI system genuinely general expands on each one:
- Transfer to a domain it was never trained on, without fine-tuning.
- Sample efficiency: learning a new skill from one or two examples rather than millions.
- Self-correction, spotting its own errors without a human pointing them out.
- Acquiring new tools on its own initiative when the current ones fall short.
- Holding performance steady when the input format, language or context shifts.
The benchmarks quietly tracking the gap
ARC-AGI, designed to resist memorisation, became the standard test of fluid reasoning. In late 2024, OpenAI’s o3 hit 87.5% on the original set with heavy compute, a jump from single digits a year earlier. On the harder ARC-AGI-2 set released in 2025, frontier models fell back to double digits against a human baseline near 85%. That contrast tells you more than any single score.
Other measures matter too. FrontierMath, built from unpublished research problems, had top models scoring under 10% when it launched. OpenAI’s GDPval compares model output against human experts on real occupational tasks, and the results land in the range of “competitive on some, clearly behind on others.” If you want to understand why these benchmarks measure something different from ordinary machine learning metrics, AGI versus machine learning is a good primer on the distinction.
The line keeps moving, which is the whole problem
Chess was the benchmark once. Then Go. Then translation, then driving, then passing exams. Every time a machine clears one, the goalposts shift and the achievement gets reclassified as narrow AI. That pattern is not cynicism; it reflects a real intuition that generality means handling the next thing, not the last one.
The reasons nobody can agree on a definition run deeper than semantics, and this piece on why the field can’t settle on what we’re building gets into the disagreements. What counts as the boundary between AGI and narrow AI depends partly on which capabilities you consider essential and partly on how generous you feel that week.
How to read the next AGI claim you see
When a company or a commentator announces that AGI has arrived, run three checks. First, ask what the system did that it was not trained for, and whether anyone independent verified it. Second, ask whether the demonstration involved a task with a known answer that the model might have seen, because benchmark contamination is rampant. Third, ask what happened on the tenth attempt, not the first.
Those questions cut through most of the noise. The systems on the market today are extraordinary tools with unusually wide reach. They read, write, code, plan and argue across domains their designers never anticipated. None of them will adapt to an entirely new profession by Friday, and until one does, the honest answer to “give me an example of AGI” is that we are still collecting candidates.