AGI history: from Turing to today's labs

AGI builds on a longstanding ambition: to create machines that can learn and solve a wide range of problems beyond a predefined specialty. From Turing to today’s labs, its history encompasses competing methods and disagreements over what would count as evidence of general intelligence.

A longstanding ambition, several distinct questions

Artificial general intelligence, or AGI, does not refer to a specific architecture. It is a goal whose meaning depends on the definition being used: performing a broad range of tasks, reaching human-level performance, learning efficiently in unfamiliar situations, or working autonomously.

These criteria are not interchangeable. Software can handle many subjects while still relying on familiar procedures. Another system might quickly learn an unfamiliar game without having the knowledge needed to write a professional report.

Understanding this history means following several threads. The first concerns observable behavior: what the machine can accomplish. The second concerns learning: how it acquires that ability. The third concerns operating conditions: resources, tools, human assistance, and the environment.

This distinction helps avoid a misleading reading of the past. Researchers have not all been working toward the top of the same ladder, with every discovery automatically taking them closer. Sometimes they have pursued different goals using similar language.

Consciousness remains a separate question. An assessment of capabilities cannot, by itself, establish that a system has subjective experience. Conversely, a philosophical discussion of consciousness cannot replace a measurement of reliability.

Our AGI and research section covers these different strands. What is AGI? offers a shorter introduction; this page explains how the problems took shape and why they persist.

Turing: turning an abstract question into a test

In Computing Machinery and Intelligence, published in October 1950, Alan Turing proposes approaching machine intelligence through an imitation game. An interrogator communicates remotely and tries to distinguish between the participants.

The shift is significant: instead of waiting for a universally accepted definition of thinking, Turing examines a setup in which behavior can be compared. His paper also discusses learning machines, so its scope extends beyond conversation alone.

The Turing test is a historical landmark, but it needs careful interpretation. Convincing an interlocutor tells us something about a particular interaction. It does not directly measure every possible ability.

Consider a hypothetical situation: a program gives an excellent explanation of how to repair a mechanism. That answer does not yet tell you whether it could identify the faulty part, investigate an unusual failure, or check that the repair works.

The methodological lesson remains useful: a test must be interpreted within its scope. If you evaluate a conversation, you obtain observations about a conversation. Any broader claim requires a justification for moving from that observation to a general ability.

This tension runs through the history of AGI. An accessible demonstration can make an ability visible to the public, while a scientific account requires explicit conditions and reproducible results. The approaches can complement each other, but they do not answer exactly the same question.

Dartmouth: setting out a broad research agenda

The Dartmouth proposal, dated August 31, 1955, lays the groundwork for a meeting in the summer of 1956. John McCarthy, Marvin Minsky, Nathaniel Rochester, and Claude Shannon describe a research program called artificial intelligence.

The document addresses language, abstractions, problem-solving, and machine improvement, among other topics. It also considers neural networks: the original project was not confined to a single programming method.

The ambition that would become associated with general intelligence is therefore present in the field’s foundational questions, even though the terminology differs from today’s debates. Reading the proposal as a modern specification for AGI would be anachronistic.

The scientific hypothesis must also be distinguished from its realization. Suggesting that an aspect of intelligence could be described precisely enough to simulate establishes a research agenda. It does not supply the algorithm, the necessary resources, or proof that every obstacle can be overcome.

The same distinction helps when reading a contemporary research program. A lab’s mission states what its researchers aim to build. A paper describes a method and its results. An independent evaluation checks particular properties under defined conditions.

Conflating these levels can turn an intention into an apparent accomplishment. Foundational texts are best read as research proposals: ambitious in the questions they ask, but offering no certification of the systems that follow.

From rules to expert systems: the challenge of transfer

One strand of research sought to represent knowledge explicitly and apply rules to solve problems. Expert systems developed this approach within specialized domains. Researcher Nils J. Nilsson traces their development in The Quest for Artificial Intelligence, including work on advice and troubleshooting.

This trajectory illustrates the difference between a general ambition and a bounded solution. Describing useful knowledge in a particular domain does not automatically provide a way to acquire knowledge in every other domain.

A fictional example makes the problem clearer. A diagnostic system knows the relationships between certain symptoms and certain faults. When faced with a differently designed device, its rules may no longer be relevant, even if they remain logically consistent.

You therefore need to distinguish the quality of the rules, the accuracy of the observations, and the situations in which the rules apply. Correct reasoning from an unsuitable assumption can lead to a poor decision. The difficulty is not simply a matter of reasoning power.

This remains a central question for AGI: how can a system reuse knowledge without assuming that a new situation closely matches previous ones? Does it need to recognize a shared structure, build a different representation, or learn a new procedure?

Keep these criteria separate when comparing approaches:

Criterion	Knowledge expressed through rules	Model learned from data
Source of knowledge	Explicit formalization	Adjustment based on examples
Possible correction	Change the relevant rules	Change the data, training, or system
Transfer question	Do the rules apply in the new context?	Are the learned patterns still relevant?
Expected evidence	Results within the intended domain	Results on relevant data and situations

This table provides a framework for comparison, not a ranking: neither column guarantees generality on its own.

Learning from data changes the method, not the underlying question

In a learning-based approach, the designer does not necessarily specify every decision rule. Instead, the designer sets up a process in which the system adjusts its behavior based on examples, objectives, or feedback on its actions.

This shifts the work involved. Data must be selected, success must be defined, and results must be checked beyond the examples used for learning. Replacing written rules with learned parameters does not remove the assumptions introduced by designers.

As an illustration, suppose you train a program to recognize objects photographed against a white background. Testing it on other, very similar photographs evaluates a useful ability. Testing it in a cluttered workshop asks a different question.

The gap between these situations helps explain why generalization is central to debates about AGI. The aim is to determine what remains reliable when appearances, instructions, or the structure of a problem change.

The number of tasks completed successfully is therefore not always enough to judge a result. You also need to examine how diverse the tasks really are, how closely they resemble the training data, and how much assistance the system receives. Exercises that look different can still require the same procedure.

Specialized performance nevertheless retains its practical value. A program does not need to be general to be useful. The history of AGI should help us avoid attributing abilities that evaluations have not established, while recognizing the value of tools with a defined scope.

The Transformer and language models broaden the debate

In June 2017, Ashish Vaswani and his coauthors introduce the Transformer in Attention Is All You Need. The paper proposes an architecture based on attention mechanisms and evaluates it on tasks including machine translation.

That context matters: the published result concerns an architecture and specific experiments. It does not demonstrate general intelligence or prove that every ability can emerge from the same principle.

For context on contemporary tools, see our profiles of ChatGPT, Claude, and Gemini. These names identify products; they are not scientific categories of intelligence.

A shared interface can create an impression of unity: a user requests a translation, an explanation, and then a program within the same conversation. Analyzing that versatility still requires examining the tasks, conditions, and errors involved.

The distinction between a model and a system also becomes necessary. An answer may depend on the model, but also on a search engine, a code interpreter, supplied documents, or a checking procedure. If you want to measure the service delivered, the relevant object of study is the whole system.

Imagine an evaluation in which an assistant can consult documentation. Its success establishes something about that assistant with that access. It does not automatically show that the model has the same ability without documentation.

This reasoning does not diminish the result. It clarifies what has improved: available knowledge, tool use, method selection, or adaptation to the problem. A rigorous history of AGI must preserve these distinctions instead of attributing everything to a single, undifferentiated intelligence.

François Chollet: distinguishing skill from skill acquisition

In On the Measure of Intelligence, published in November 2019, François Chollet proposes characterizing intelligence through the efficiency with which a system acquires skills. He emphasizes the roles of prior knowledge, experience, and generalization difficulty.

The paper also introduces ARC, a collection of tasks designed to examine a form of abstract reasoning. The distinction from simply measuring skill is essential: succeeding at a task does not, by itself, reveal what was needed to learn how to succeed.

This is a position about measurement. It calls for documenting the starting point and the experience a system receives, rather than looking only at its final performance. Our reference page on generalization explores this framework and its implications.

Consider an illustrative example. You give two systems an unfamiliar classification rule. One receives a very large collection of exercises with the correct answers; the other has only a few demonstrations. Identical final results do not tell you the same thing about their learning.

You must still define carefully what counts as prior experience. A problem that is new to the evaluator may resemble exercises the system has already encountered. An appearance of novelty is therefore insufficient.

This framework does not make workplace evaluations useless. If you need to automate a task, the quality of the output matters directly. But if your question concerns general intelligence, you also need to examine how the skill was acquired.

On the Measure of Intelligence, explained presents the argument more directly. It complements this history by examining a measurement proposal in detail, without treating it as a universally accepted definition.

Labs propose different thresholds

The Levels of AGI framework, first published in November 2023, distinguishes dimensions including the depth and breadth of performance. Its authors, including Meredith Ringel Morris and Shane Legg, consider autonomy and deployment separately.

This approach differs from François Chollet’s. It aims to organize observed capabilities along several dimensions, while his approach emphasizes the efficiency of skill acquisition. These questions can complement each other without producing an identical threshold.

In October 2024, Dario Amodei describes a hypothetical form of powerful AI in Machines of Loving Grace: broad intellectual abilities, access to digital interfaces, and the capacity to carry out lengthy tasks. He presents a future scenario, with its uncertainties, rather than a certification of abilities already achieved.

Google DeepMind offers another breakdown in its cognitive framework presented on March 17, 2026. It examines areas including learning, memory, reasoning, and metacognition, with comparisons to human performance.

These publications show why any claim about AGI needs an accompanying definition. A measure of versatility, a measure of learning, and a scenario involving autonomy do not address the same object.

As of this page’s date, October 24, 2026, these texts serve as documentary reference points. Their existence does not establish that a shared threshold has been crossed.

When comparing announcements, ask what is being evaluated, by whom, and under which procedure. An organization can use a coherent definition for its research without that definition being binding on the wider scientific community.

Reading progress without turning history into a timeline of predictions

The history of AGI helps make sense of announcements when you separate three elements: the observed ability, the proposed explanation, and the projection into the future. They do not carry the same evidential weight.

An experiment may show that a system solves a family of problems. The authors may attribute that result to a particular method. Predicting that the method will be sufficient to build general intelligence requires additional assumptions.

The same care applies to failures. A difficulty encountered under a particular test procedure does not automatically prove a permanent impossibility. You need to establish whether it arises from the data, the method, the resources, or the definition of the task itself.

When reading a new paper, you can follow this sequence:

Identify the system being evaluated and the resources available to it.
Examine the tasks, how they were selected, and how closely they resemble the training material.
Check the human comparisons and the conditions under which the tasks were performed.
Look for failures, their frequency, and their severity.
Separate observations from extrapolations.

Our page on AGI benchmarks examines these measurement problems in detail. A score becomes informative when you understand what it summarizes and what it leaves out.

For your own use, this history also suggests a practical rule: evaluate a tool on the work you intend to give it. A label suggesting general ability cannot replace a representative trial, an assessment of errors, or a check on how much supervision is needed.

From Turing to today’s labs, the enduring thread is the effort to design tests that match the claims being made. Methods change; the need to specify what has been demonstrated remains.

Frequently asked questions

Did Alan Turing invent AGI?

Turing raised some of the foundational questions about intelligent machines, but crediting him alone with AGI would overlook distinct research traditions. His paper on the imitation game does not define a single technical path to general intelligence.

Why use AGI instead of simply AI?

AI also includes software designed for a specific function. Adding general signals an ambition to learn and transfer abilities across tasks, although researchers disagree on where to set the threshold.

Does AGI need a body, or does it have to be a robot?

That depends on the definition. A definition focused on intellectual work may allow a software-only system; one that includes physical abilities requires different observations and tests.

Is a system that passes the Turing test an AGI?

A convincing conversation does not establish every ability covered by a definition of AGI. You also need to specify the test procedure, the knowledge required, and the skills that fall outside the conversation.

Why do researchers disagree about when AGI will arrive?

They are not always predicting the same event: success on evaluations, autonomy at work, or a general ability to learn. Their assumptions about data, computing resources, and scientific obstacles also differ.

The history of AGI: from Turing to today's labs