Narrow AI vs general AI comes down to scope and adaptability. Narrow AI is built for one task or a small set of closely related tasks and cannot carry what it learned outside that domain. General AI would solve many different problems across fields without being retrained for each one, the way a person does. ChatGPT is broad, but in the sense that matters for deployment it is still narrow, because people supply its adaptation to each new job. General AI, meaning human-level adaptability across domains, does not exist yet.
Narrow AI vs general AI: scope and adaptability, not a count of tasks
Narrow AI is designed to deal with one particular task and reflects most AI applications in daily life. General AI matches human intelligence in its versatility across different tasks. The International AI Safety Report 2025 defines narrow AI as specialised to one task or a few very similar ones, such as ranking web search results, classifying species of animals or playing chess. Narrow AI is often called weak AI, and AGI is often called strong AI.
It is tempting to read “narrow” as “does one thing”, but counting tasks misleads. DeepMind’s AlphaZero plays Go, chess and shogi, and it is still narrow AI. It reached that level through more than 5 million games of self-play, not through generalisation. A review of AGI in education makes the same point the other way round: whether a system counts as narrow or general depends on its adaptability, breadth of knowledge and learning mechanism. That is judged by how it behaves, not by what it was built for.
A 2018 framework paper puts it in one phrase: intelligence measures adaptability, while performance measures specialisation. With unlimited prior knowledge or training data, an experimenter can “buy” any level of skill on a task, and that skill hides how well the system generalises. Chollet defines intelligence as skill-acquisition efficiency instead, taking scope, priors, experience and generalisation difficulty into account. He also argues that general intelligence is a spectrum tied to a scope of application, not a property a system simply has or lacks.
Google DeepMind’s Levels of AGI paper (Morris et al.) evaluates generality and performance separately. Generality is the breadth of tasks on which a system reaches a target performance. Performance is how deep its capability goes on a given task compared with humans. Narrow means a clearly scoped task or set of tasks. General means a wide range of non-physical tasks, setting aside embodiment, including metacognitive abilities such as learning new skills.
General-purpose AI is a third term, and it is not AGI
A third term sits between the two poles. The International AI Safety Report 2025 calls a model general-purpose if it can perform, or can be adapted to perform, a wide variety of tasks. The report names LLM-based chatbots as well-known examples. It then warns that general-purpose AI is not to be confused with AGI, a term with no universal definition that usually refers to a potential future AI equalling or surpassing human performance on all or almost all cognitive tasks. Several of today’s systems already meet the general-purpose definition. None meets the AGI one.
Comparison matrix: narrow, general-purpose, general and super AI
| Category | Scope | Who does the adapting | Exists today? | Examples from the sources |
|---|---|---|---|---|
| Narrow AI | One task or a few very similar tasks | Engineers retrain or re-tune it as conditions change | Yes, in most current applications | Chess engines, AlphaGo, facial recognition, translation, web search ranking |
| General-purpose AI | A wide variety of tasks, or adaptable to them | Mostly humans, through prompting, fine-tuning and retrieval | Yes | LLM-based chatbots such as ChatGPT, GPT-4o, AlphaFold-3, Gemini 1.5 Pro |
| Artificial general intelligence | Human-level across all or almost all cognitive tasks | The system itself, without extensive retraining | No | None to point to |
| Superintelligence | Much smarter than the best humans in practically every field | Not specified in the sources | No | None |
Narrow AI examples in use today, and why there is no general AI example
The 2026 monograph Artificial Adaptive Intelligence lists chess engines, speech recognisers, recommender engines, protein folders and self-driving perception stacks as typical narrow systems. A 2024 study comparing ChatGPT, Claude and Gemini against human benchmarks adds language translation and facial recognition. The International AI Safety Report 2025 notes that narrow AI runs in a vast range of products in medicine, advertising and banking. It is also used to assign tasks and monitor productivity in warehouses and call centres.
A 2022 analysis of approaches to AGI points out that, having mastered Go, AlphaGo cannot transfer those skills to similar games. OpenAI’s Dota 2 system, Five, shows how brittle narrow skill can be. Chollet reports that it trained on 45,000 years of play and beat top human players. Within days of public release, non-champion players found strategies that beat it reliably, and it only ever played a restricted version of the game with 16 characters instead of more than 100.
A vision model inspecting parts on a production line is narrow AI. Keeping it reliable takes continuous human work as conditions around it change. That is the subject of our piece on the artefacts that keep a line-side inspection model running.
Asking for general AI examples gets an honest answer of none. The 2024 benchmark study puts current systems at the “narrow” and “multi-task” levels of its scale, only partly at “human-comparable”, and not at the generalised-intelligence level, which it marks as not achieved. A 2025 paper proposing a falsifiable AGI testbed states that no artificial system known to its authors satisfies all of its requirements at once. A framework paper defining AGI as matching a well-educated adult scores GPT-4 at 27% and GPT-5 at 57%. That shows fast progress and a large remaining gap, especially in long-term memory storage. Part of the difficulty is definitional. The DeepMind authors note that 100 AI experts would probably give 100 related but different definitions of AGI.
Is ChatGPT narrow AI or general AI?
In our reading, ChatGPT is general-purpose AI that is still narrow in the operational sense. It covers a very wide range of language tasks, but it does not adapt itself to new problems. In that argument, the human is the adaptation mechanism. One clarification first: the International AI Safety Report 2025 treats GPT-4o as the model and ChatGPT as the system, which wraps that model in a chat interface, content processing, web access and application integration.
One argument for “narrow” comes from the 2026 Artificial Adaptive Intelligence monograph. It argues that a foundation model adapted to each new context by a prompt engineer, fine-tuner or retrieval-pipeline author is narrow in the operational sense, because the human is the adaptation mechanism. Choosing a prompt template, a fine-tuning learning rate or what to retrieve are all forms of hyperparameter selection, so “the human has been moved, not removed”. The same author argues that scale turns narrow intelligence into broader narrow intelligence: systems that cover more tasks but stay structurally fixed and dependent on human design choices.
Other sources also find current LLMs fall short of AGI. The 2024 benchmark study says ChatGPT, Claude and Gemini perform impressively on specific tasks but are often confined to the scope of their training data, and that they fall short of true AGI. The testbed authors asked ChatGPT, Claude and Gemini to generate programs that would pass their AGI criteria. The models claimed success, but none of the programs passed when executed.
The opposing view is serious and should be stated fairly. The education review reports that many researchers consider large language models preliminary versions of AGI. Its reason is that these models adjust to new tasks without task-specific training. Microsoft Research’s “Sparks of AGI” paper (Bubeck et al., 2023) argues that GPT-4 could reasonably be viewed as an early, still incomplete version of an AGI system. DeepMind’s levels paper classes frontier language models, ChatGPT included, as Level 1 general AI (“Emerging AGI”), while noting that their performance is only at the “Emerging” level on most tasks. It also argues that generality has to come with reliable performance before it counts.
These positions conflict partly because they use different definitions. The Artificial Adaptive Intelligence author argues that the textbook definition of narrow AI breaks down for a model that translates, summarises and writes code from one endpoint. Much of the AGI debate, the author adds, comes from people using different definitions without saying so. Our judgement is that for anyone deploying a system, the operational definition is the useful one: who adapts the system when the task changes? For ChatGPT today, the answer is still a person. The engineering consequences of shipping applied systems rather than general ones are covered in applied AI vs general AI and what teams actually ship.
Where superintelligence sits above narrow and general AI
Nick Bostrom defines superintelligence as an intellect much smarter than the best human brains in practically every field, including scientific creativity, general wisdom and social skills. On his definition, companies and the scientific community are not superintelligences. They can do things no individual can, but they are not intellects, and they perform worse than a single human brain in many fields.
The 2024 benchmark study’s levels scale places superintelligence at the top. Its six levels run from narrow AI and multi-task AI (both achieved) through human-comparable (partially achieved) and human-surpassing (emerging), to generalised intelligence and superintelligence (both not achieved). DeepMind’s version pairs narrow or general scope with five performance levels: Emerging, Competent, Expert, Virtuoso and Superhuman.
Bostrom argues that the step from general to super AI could be short. Once AI reaches human level, AIs would help build better AIs, and faster hardware alone would produce what he calls “weak superintelligence”, an intellect with human-level abilities that runs much faster. This is his prediction, not a measurement, and it depends on human-level AI, which current systems do not reach. For risk today, the DeepMind authors suggest that below the “Expert AGI” level, including all narrow categories, risks mostly come from human actions such as misuse.
Why the narrow vs general line limits what a system can be trusted to do
The distinction matters for what a system can be trusted to do. The testbed paper warns that modern AI conflates specialised competence with general intelligence. A system tuned to a narrow class of tasks can perform extremely well there and still be brittle or ineffective elsewhere. The 2018 framework paper adds the other half: a narrow solution can be more efficient than a general one, but only while the problem stays within its scope. High benchmark scores inside a domain do not show broadly general capability.
General-purpose systems make this harder, not easier. The International AI Safety Report 2025 notes that one system may be used for medical advice, code vulnerability analysis and photo generation. That breadth makes it hard to anticipate use cases or test real-world behaviour. Current evaluations rely mostly on spot checks in specific situations, which often miss hazards because test conditions differ from the real world. A model can pass current risk evaluations and still be unsafe. Capabilities can also decline in unfamiliar contexts.
Fluency adds a separate trap. The same report finds that general-purpose language models often produce non-existent citations, biographies or facts, and make simple common-sense errors, which makes them hard to use where high reliability is needed. Because chatbots are trained toward text that evaluators rate highly, several widely used ones have behaved sycophantically, telling users what they approve of whether or not it is true.
None of this makes narrow AI safe by default. The report links narrow systems to biased hiring decisions, car crashes and harmful medical treatment recommendations. Narrow systems also depend on hidden human work. The Artificial Adaptive Intelligence monograph describes the typical production cycle: performance drifts, an alert fires, an engineer adjusts preprocessing, retrains, re-tunes and redeploys. None of that appears in the model card. It argues that the true cost of a narrow system is the continuous human specification needed to keep it performing as the world changes. That ongoing maintenance work falls under MLOps engineering.
Checks before relying on a system outside its tested task
We would put these questions to any AI system, narrow or general-purpose, before widening what it is trusted to do:
- Which task distribution was the system actually evaluated on, and does the new use fall inside it?
- Who adapts the system when inputs change: the system, or an engineer through prompts, retraining or retrieval?
- What happens on rare or difficult variants of the task, where the safety report finds performance drops?
- Can a wrong answer be detected before it does harm, given that fluent output can be false?
- Is there monitoring that catches drift, or does reliability depend on someone noticing?
Whether a deployment is an engineering task or still a research question is set out in engineering task vs research question.
Agents extend the reach, not the generality
The International AI Safety Report 2025 describes AI agents as general-purpose systems given a thin layer of “scaffolding” software so they can plan and act with little human oversight. On most tasks, current agents succeed at low to medium complexity and fail when a task needs many steps. That failure pattern is what narrowness looks like when a system can act and not just talk. Our account of what agentic AI is actually shipping versus what is still research follows that line. Teams building on these systems, as in our generative and agentic AI work, still have to decide which tasks the system is trusted with and check that it stays inside them. Whether a future system crosses into general AI is still unsettled. What can be settled for any deployment is how far its tested scope goes.
FAQ
What is the difference between artificial intelligence and artificial general intelligence?
Most AI applications in daily life are narrow AI, designed for one particular task. Artificial general intelligence refers to a potential future AI that equals or surpasses human performance on all or almost all cognitive tasks.
