Since the invention of the modern computer, there has been debate over how to define artificial general intelligence (AGI), how to test a machine to see whether it meets that definition, and what the benefits and drawbacks of AGI will be for human work, creativity, and scientific discovery.
This article explains what AGI is, explores its history, key challenges, and whether it already exists or remains a distant goal.
Table of contents
Understanding artificial intelligence (AI)
What is artificial general intelligence (AGI)?
Potential applications of general AI
Ethical considerations and challenges
Understanding artificial intelligence (AI)
To understand AGI, it’s important to distinguish it from other forms of artificial intelligence (AI). AI is generally categorized by how broadly it can apply its intelligence and how well it performs compared to humans.
What is artificial intelligence?
AI refers to technology that enables machines to solve complex problems, often mimicking or surpassing human abilities. It powers tasks like language processing, speech and image recognition, data analysis, and code generation. However, AI varies in capability and can be classified into three main types:
- Narrow AI (weak AI): Specialized systems designed for specific tasks, such as spam filtering, recommendation algorithms, and chess-playing programs. These systems excel in their designated functions but cannot adapt beyond them. All current AI falls under this category.
- Artificial general intelligence (AGI): A theoretical AI that can learn, reason, and solve problems across a wide range of domains, similar to human intelligence. Unlike narrow AI, AGI would not require retraining for new challenges.
- Artificial superintelligence (ASI): A hypothetical AI that surpasses human intelligence across all disciplines, including creative problem-solving and strategic thinking. ASI remains speculative but is often discussed in relation to AGI’s long-term evolution.
While today’s AI is impressive, it remains narrow, excelling only within predefined boundaries. The pursuit of AGI is the quest for a true machine intelligence—one that can think, learn, and adapt like a human.
What is artificial general intelligence (AGI)?
There is no universally accepted definition of AGI, also known as general AI. However, many definitions suggest that a system qualifies as AGI if it can do the following:
- Learn adaptively without requiring human intervention
- Generalize knowledge to solve unfamiliar problems
- Perform comparably to humans across a broad range of tasks
Beyond these broad attributes, definitions of AGI vary, often reflecting the goals of those attempting to develop it:
- In their 2007 book, Artificial General Intelligence, Ben Goertzel and Cassio Pennachin define AGI as AI systems that possess “a reasonable degree of self-understanding and autonomous self-control” and can solve a variety of complex problems in various contexts.
- OpenAI defines AGI as “a highly autonomous system that outperforms humans at most economically valuable work.”
- François Chollet, former Google AI researcher and the creator of the ARC-AGI benchmark, defines AGI as a system capable of efficiently acquiring new skills outside of its training data. He emphasizes that intelligence is marked by skill acquisition and generalization, rather than skill itself.
Key traits of AGI
While definitions of AGI vary, they generally distinguish it from narrow AI by emphasizing its ability to function across diverse domains. Regardless of the specific definition, an AGI would need to possess several core traits to achieve these capabilities:
Autonomous decision-making
An AGI must be able to determine when to seek new information, request assistance, or take independent actions to solve problems. For example, if tasked with modeling a complex financial market, an AGI would need to identify relevant data sources, analyze historical trends, and determine how to acquire the necessary information—all without human guidance.
Problem-solving in unfamiliar domains
AGI must be able to generalize knowledge from one domain and apply it to new, unfamiliar tasks. This ability to transfer learning through analogy is similar to how a musician trained on one or two instruments can quickly learn a third. In the same way, an AGI must leverage prior knowledge to solve problems it was not explicitly trained for.
Continuous self-improvement
An AGI must be capable of evaluating its own performance and adapting to new situations. One approach to recursive self-improvement is self-generated training data, as seen in DeepMind’s RoboCat. Another potential capability is modifying its own code and architecture. However, such self-modification could introduce safety risks if AGI makes changes that humans cannot fully understand or control.
History of general AI
The history of AGI is best understood within the broader history of AI. Research has evolved through several distinct eras, each shaping the path toward more capable and general AI systems.
Early AI: Symbolic AI (1950s–1980s)
The first attempt to build AI in the 1950s and 1960s was based on the idea that you could teach a machine to think by programming rules and logic (represented as symbols) into the computer and asking it to solve problems using these rules. This produced expert systems that could beat humans at board games and do specialized tasks (IBM’s chess champion computer Deep Blue is one example), but they were unable to learn anything outside of their programmed knowledge.
The shift to machine learning (1990s–2010s)
A major shift occurred in the 1990s with the rise of machine learning (ML), which took inspiration from how biological neurons function in the brain. Rather than using hard-coded rules, these connectionist systems use neural networks that use many layers of artificial neurons that learn by training on large datasets and improving their outputs incrementally over many training runs.
The deep learning revolution (2010s–present)
The modern deep learning revolution began in 2012 when researchers began using graphics processing units (GPUs) to create neural networks with trillions of parameters. This provided a huge boost in computational power that gave these machine learning models—including contemporary large language models (LLMs) like ChatGPT—the capacity to learn more and generalize some knowledge to similar tasks.
Defining AGI: Beyond traditional AI benchmarks
As AI systems became more sophisticated, researchers proposed new benchmarks to assess whether an AI system had reached human-level intelligence. The most famous early benchmark, the Turing test, was designed to determine if a machine could mimic human conversation convincingly. However, as LLMs like ChatGPT and Claude can now pass this test, many researchers consider it outdated.
More recent benchmarks, such as the ARC-AGI test, focus on an AI system’s ability to generalize beyond its training data. While current AI models still fall short of humanlike reasoning, some, like OpenAI’s o3 model, have achieved breakthrough results, reigniting debates on AGI’s feasibility.
How might AGI work?
There is no consensus among AI researchers on which approach will ultimately lead to AGI. Both symbolic AI and deep learning have limitations when it comes to building systems that can generalize knowledge across different domains. Current research focuses on developing models with metacognitive abilities—the capacity to evaluate and improve their own reasoning processes.
Limitations on current approaches
Symbolic AI systems rely on human programmers for knowledge and can’t obtain new information on their own, while deep learning systems, including generative AI, require vast datasets and long training periods to learn new tasks. Humans, on the other hand, readily absorb new information and can learn to do new things quickly with very few examples.
Even with these challenges, however, researchers are exploring many avenues to create machines capable of learning, generalizing, and making decisions at a human (or better) level. Some recent approaches that have elements of AGI include neuro-symbolic AI, agentic AI, and embodied AI.
Neuro-symbolic AI
Some AI researchers, including Gary Marcus and Ben Goertzel, argue that neuro-symbolic systems are the way to AGI. These systems combine different types of AI systems to compensate for the shortcomings of any one approach.
For example, in 2023, Goertzel and his collaborators released OpenCog Hyperon, an open-source AGI effort that provides a software framework for combining AI systems from various disciplines, including natural language processing (NLP), formal logic, and probabilistic reasoning. Google DeepMind recently achieved silver medal–level performance at the International Mathematical Olympiad with two neuro-symbolic systems, AlphaProof and AlphaGeometry 2.
Agentic AI
AI agents are considered a possible step on the road to AGI because they can evaluate and respond to their environments, understand context, and make decisions independent of humans to accomplish objectives. Like the neuro-symbolic approach, agentic AI systems work by combining multiple kinds of AI to accomplish different tasks. However, research into agentic AI is still in its early stages, and many of the more advanced capabilities attributed to agentic AI are still theoretical.
Embodied AI
Leading AI thinkers, including OpenAI co-founder Andrej Karpathy and scientist Melanie Mitchell, have said some form of embodiment may be necessary to reach AGI. This is rooted in the idea that it would be difficult for an AI to learn basic cognitive skills like understanding causality or object permanence without the ability to receive sensory inputs.
Embodied AI is implicitly required to meet some popular definitions of AGI. For example, Apple co-founder Steve Wozniak has proposed a benchmark called the Coffee Test, in which a machine could be considered to possess AGI if it was able to enter the home of an arbitrary person and figure out how to brew a cup of coffee.
Potential applications of general AI
Because of the nature of generalized intelligence, the potential applications for AGI are virtually limitless. Some industries that may particularly benefit from the adaptiveness and autonomy that AGI will offer include healthcare, education, manufacturing, and finance.
Healthcare
AGI has the potential to affect many areas of healthcare where it would be advantageous to have an intelligent system with access to vast amounts of data, including diagnostics and drug discovery, and the ability to create individualized treatment plans that reflect the full picture of a patient’s health history.
Education
AGI systems in education can be used to help personalize learning pathways for students to meet their specific needs, assist teachers with administrative tasks and lesson planning so they can spend more time on teaching, and help teachers analyze student performance to identify gaps where students may be falling behind.
Manufacturing
Manufacturers have a constant need to optimize the processes that underlie complex supply chain logistics, production schedules, and quality control. AGI has the potential to aid in making decisions about how to improve processes and automate them.
Finance
Because financial sector companies deal with a vast amount of data, AGI will be able to analyze and make decisions about that scale of information much faster than humans can. This has the potential to speed up data-heavy tasks like risk assessment, compliance, and market analysis.
Ethical considerations and challenges
As progress toward AGI continues to advance, there are legal issues and ethical concerns that will have to be considered by both those building and those using AGI systems.
Bias
In the same way that narrow AI systems can suffer from a lack of diversity in training samples, AGI systems have the potential to exhibit racial, gender, or other types of bias based on skewed or incomplete training data. Algorithms can also introduce bias by weighting certain variables to privilege one group over another.
Legal responsibility for AGI actions
AI systems have already been the subject of legal disputes over violations of privacy and fair housing laws. However, existing legal frameworks do not always clearly define who is liable for harm caused by AI. The emergence of advanced intelligent agents will further complicate questions of accountability when machines act in ways that break the law.
Alignment challenges
AGI systems could have access to vast amounts of data and the autonomy to make impactful decisions. Ensuring that these systems align with human values and ethical principles is a key focus of AI alignment research. Experts are working to develop methods that enable AGI to interpret and adhere to desired goals and constraints, minimizing unintended or undesirable outcomes.
Future of general AI
As AI advances, it presents both challenges and opportunities. While concerns around employment and safety must be addressed, AGI has the potential to bring significant benefits in fields such as data analysis, automation, optimization, healthcare, and security.
AGI could accelerate progress on complex scientific and social issues by solving problems at a scale beyond human capability. By handling repetitive tasks, AGI may also free people to focus more on meaningful work and personal interests. Ultimately, its development will reshape not only industries but also how humans perceive intelligence and their role in the world.
AGI FAQs
What is the difference between AI and AGI?
AGI is a subtype of AI that differs from narrow or weak AI, which is designed to perform specific tasks within a limited domain. In contrast, AGI refers to a hypothetical stage of AI development in which systems possess humanlike flexibility, adaptability, and reasoning, allowing them to learn and perform a wide range of tasks across different domains.
What’s the difference between generative AI and general AI?
Generative AI is a type of AI that analyzes large datasets to generate predictions, content, or responses based on learned patterns. General AI, or AGI, refers to AI capable of human-level intelligence and reasoning across multiple domains, allowing it to learn and perform a wide variety of tasks without being limited to a specific function.
Is ChatGPT considered AGI?
Some experts suggest that LLMs like ChatGPT and Claude could already be considered AGI. However, this view is not widely accepted among AI researchers. ChatGPT lacks a true understanding of the text it generates, struggles with reasoning, and cannot generalize its knowledge across different domains, such as controlling a physical system like a self-driving car. These limitations mean it does not meet the criteria for AGI.
Is o3 considered AGI?
While OpenAI’s o3 reasoning model achieved an impressive 87.5% score on the ARC-AGI benchmark on December 20, 2024, the benchmark’s creator, François Chollet, does not consider it to have reached AGI.
Observers point out that o3 relied on extensive pre-training with public test samples and required massive computational resources to achieve its score. Chollet also noted that some lower-compute models scored as high as 81%, suggesting that o3’s success was driven more by brute-force computation than by true general intelligence.
What are the main challenges in building general AI?
- Trustworthiness: AGI systems must be consistently accurate and reliable for users to depend on their outputs in critical applications.
- The long-tail problem: No matter how much training data an AI system has, it will inevitably encounter rare or unforeseen scenarios. For example, self-driving cars will face situations not covered in their training, requiring them to generalize effectively.
- Energy consumption: Advanced AI models already require vast amounts of energy and water for computation. AGI could demand even greater resources unless more efficient processing methods are developed.
- Common sense: Unlike humans, AI lacks real-world experience and intuitive understanding of physics, social interactions, and everyday reasoning—knowledge that people acquire naturally from childhood.
Does AGI exist yet?
Because the term AGI has been defined in different ways, what meets one person’s (or company’s) definition of AGI may already exist for them but not according to someone else. Using the definition from Google DeepMind’s paper that “an AI system that is at least as capable as a human at most tasks,” it makes sense to say AGI does not yet exist.
