A client's board requested a slide regarding "our AGI strategy." At the time, the team responsible for the slide had developed a chatbot to summarise support tickets and a forecasting model with a twenty per cent error rate. No one present could define what specific functions artificial general intelligence would perform that differed from these existing systems to justify the term general.
The slide was created regardless. It contained many statements. None of these statements could be proven wrong.
Artificial general intelligence is a system that performs as well as a human across all cognitive tasks. It transfers learned information between different areas instead of requiring a new model for every domain. There is no consensus on a specific definition or a timeline for its development. Current systems are capable, but they are still limited to specific functions.
We build data and AI systems at Shipshape Data. Projects often stall for an unusual reason. The delay does not come from the data or the model. Instead, it comes from a term that has become a stand-in for "AI, but more impressive than what we currently have." Some teams delay decisions while they wait for AGI to arrive as if it were a product release with a specific date. Other teams label a functional narrow system as a step towards AGI to make their presentations sound better. Both approaches avoid the necessary question of what the current system can and cannot do.
This guide explains the meaning of artificial general intelligence, how it differs from current narrow systems, and why researchers do not agree on its definition or timeline. It also describes how these concepts should affect a company's AI strategy this year. This guide does not provide a specific date for the arrival of AGI. No one knows when it will happen, and precise predictions usually reflect personal incentives rather than technological facts.
Reasons for the imprecise use of the term
People use the term "AGI" in at least three different ways during the same conversation, and often within a single sentence. It can refer to a system that performs as well as a human on most cognitive tasks. It can also describe a system that learns a new task without requiring additional training. In other cases, the term is used as a synonym for "very capable" and lacks a specific technical definition.
One word that performs three functions
The lack of precision is intentional. Multiple parties find vagueness to be useful.
This language allows a laboratory to describe its plans in a way that justifies funding without making a specific, testable claim. It allows a vendor to suggest their product is more advanced than its actual capabilities would indicate. It also allows a commentator to warn about risks without specifying the exact nature, timing, or conditions of those risks. These actions are often just imprecise rather than dishonest. Imprecision is useful because a person cannot be proven wrong if their claim was never clearly defined.
The lack of precision creates a problem for a business that is trying to plan.
You can't budget for, staff against, or govern a term three people in the same meeting mean differently. We've sat in planning sessions where "getting ready for AGI" was offered as a reason to delay a perfectly sensible narrow project, and nobody could say what readiness would consist of, or how anyone would know it had been reached.
What "general" is supposed to mean
If you remove the marketing language, the central concept is simple. A general intelligence can apply its knowledge to problems it was not specifically designed or trained to solve. For example, a competent adult can switch between cooking, repairing a bicycle, and reading a legal contract by using the same basic reasoning skills for each task. When researchers are asked for a specific definition, they usually mean a system that matches or exceeds human performance across a wide range of cognitive tasks and can transfer learning from one area to another.
Transfer is the hard part
The transfer property is the primary element of the definition, and current systems struggle with it. A model trained to perform one task rarely improves at an unrelated task just because its performance on the first task increased. Human intelligence functions differently. For example, a person who learns to read a balance sheet does not start from zero when they later learn to read a tenancy agreement.
This is the quality people refer to when they use the term "general," even if they use less precise language to describe it.
The AI running in your business today is narrow, and that's fine
Most of the systems we deploy for clients are narrow AI. These models are built and tuned for specific tasks, trained on relevant data, and evaluated against specific metrics. A demand forecasting model performs demand forecasting well but does not perform other functions. A document classifier sorts documents without understanding their content. Even a large language model that answers open-ended questions is narrow in this context. It performs next-token prediction based on its training rather than reasoning from a general model of the world like a human, regardless of how fluent the output appears.
Narrow does not mean weak
Describing a system as narrow is not a criticism. These systems are often highly capable within their specific functions, sometimes performing better than humans. This level of specialisation is typically what a business requires. A model designed for fraud detection does not need to write poetry. The narrow focus is an intentional design choice rather than a flaw. Our detailed analysis of narrow AI explains how these systems are developed and where they are most useful. Most AI implementations in the coming years will use this type of technology.
Generative models make the situation less clear
Large language models complicate the narrow-versus-general story because they operate across a genuinely wide range of tasks: drafting, summarising, coding, translating, holding a conversation about almost anything. That breadth is real and it is new. It is not the same thing as general intelligence, though, because breadth of surface task is not the same as the underlying transfer and reasoning that the AGI definition is reaching for. A model can write a competent legal summary and still fail at a change to the same problem that a ten-year-old would handle without pausing, because it has learned patterns in language rather than built a model of the situation the language describes. Breadth without robustness is still narrow, just narrow across more surfaces at once.
Reasons for the lack of a consensus definition
If you ask five researchers to define AGI, you will likely receive five different tests. This occurs because they disagree on which capability is most important. Some define it by economic impact, describing a system that can perform most economically valuable human labour. Others focus on cognitive breadth, requiring performance that matches humans across a standard set of tasks. Some prioritise autonomy, or the ability to set and pursue goals over time without a human specifying every step. These definitions represent different goals. A system could reach one of these milestones without approaching the others.
The moving-goalposts problem is real, and it cuts both ways
Critics of this field frequently argue that the standards for success change all the time. They note that tasks like chess, Go, and passing a bar exam served as evidence of general intelligence until computers mastered them. The definition of intelligence appeared to move to a new and harder task every time a machine reached one of these milestones. This observation has some validity. These changes are honest. Some milestones were more limited than people expected. Machines achieved them by using pattern matching and search functions. These machines did not use the general reasoning skills that people once thought were necessary. Researchers change a definition after they learn that a specific benchmark does not require general intelligence. This process updates a hypothesis based on new evidence.
Discussions regarding artificial general intelligence are seldom about the technology itself. Instead, these debates focus on which specific metric should serve as evidence for a concept that no one can define before it occurs.
Definitions matter because money follows them
This issue is not a minor academic disagreement. Contracts, funding rounds, and a well-known corporate partnership agreement have reportedly depended on a specific threshold for AGI. The person or entity that determines the threshold has been met can trigger significant financial and legal outcomes. When these definitions dictate payment obligations, "we will know it when we see it" is not a sufficient answer. This explains why the definition is a subject of ongoing conflict rather than a settled matter.
Why timelines range from years to never
If you ask AI researchers when AGI will arrive, their answers range from a few decades to several centuries. Some researchers believe it will happen "not with current methods at all." This wide range of predictions does not mean the topic is unexamined. Instead, it shows that estimates depend on assumptions that cannot be verified yet. These include whether scaling current models will continue to produce new capabilities at the same rate and whether a missing component, such as a new type of memory or learning, is necessary. The timeline also depends on whether the definition of AGI focuses on economic impact or on more complex general reasoning.
Optimism has a business model
Pay attention to who is making confident predictions about when things will happen. Organisations working quickly to create advanced systems have a clear reason to suggest milestones are near: this can lead to more investment, skilled employees, and public interest. This does not necessarily mean these predictions are inaccurate. However, it does mean the source of a timeline is important when you consider how much to believe it, just as you would evaluate a sales forecast differently depending on whether it comes from the company selling a product or from an independent analyst.
Pessimism has a business model too
This principle applies in reverse. Warnings about immediate and significant risks attract specific types of attention and funding for research institutes, advocacy groups, and regulatory proposals. The incentives on either side do not prove that their arguments are incorrect. However, these incentives mean that a stated timeline is rarely a neutral piece of information. It is a mistake to treat any single timeline as a settled fact, no matter which direction it favours.
Progress on benchmarks isn't the same as progress toward AGI
New models frequently achieve high scores on reasoning benchmarks, coding tests, or professional exams. These results are often presented as significant progress towards artificial general intelligence. Some of these improvements are real and practical. However, some results occur because models are gaming the benchmarks. This happens when training data includes the test material or when a model learns to exploit specific patterns in a test without possessing the actual skill the test is meant to evaluate.
The meaning of a high score and its limitations
A model with high scores on logical puzzle benchmarks has proven it can perform well on those specific tests. This does not guarantee the model will solve the vague or complex problems found in business environments. Planning requires understanding this distinction. We have seen teams fund expensive proofs of concept based on high benchmark scores, only to discover the model cannot handle the specific or unusual tasks their business requires. The benchmark provided an accurate measurement, but it did not measure the capabilities the company actually needed.
What would actually change if it arrived
Imagine that a system meeting a strict definition of artificial general intelligence was created tomorrow.
How would business operations change in reality, compared to the changes that are often predicted? Much of the current work would remain necessary. We would still need to clean data, define customer records, and determine which systems manage specific facts. A general reasoning model does not understand your business definitions better than a narrow model does. It requires access to governed and reliable data to generate reliable results. If a more intelligent model is provided with disorganised data, it will produce a more articulate version of that data rather than a correct answer.
What might genuinely shift
What could plausibly change is the amount of task-specific engineering required to get a system working on a new problem. Today, moving a model from one use case to a meaningfully different one usually takes real integration work: new prompts, new evaluation, sometimes new fine-tuning. A system with genuine transfer ability would need less of that per new task, and a smaller integration bill is a genuine commercial benefit worth planning around once it is real rather than promised. That is a useful shift if it happens. It is not the same as the system suddenly needing no oversight, no governance, and no checking, and treating it that way would be a mistake regardless of how capable the underlying model became.
Autonomy is a separate question from capability
A system does not become more autonomous simply because it is more general. The capability of a model and the level of independent authority it receives are separate factors. These factors are determined by the people who deploy the system rather than by the model itself. The risk management practices used for narrow models today remain necessary as models improve. Organisations must still define which decisions a system can make unsupervised and which it must escalate to a human. A highly capable system operating with weak oversight represents an increase in existing risks rather than a new type of problem.
Where to start
Do not base your planning on release dates that are presently unknown. Instead, base your plan on the specific gaps in your current capabilities. Identify which business tasks are too diverse for your existing narrow tools and which tasks need a system that can adapt to related problems. You should also identify which tasks are already handled effectively by your current models, even if those models are not regarded as innovative. List these gaps clearly and include examples of specific tasks that cannot be completed at present. Review this list each time a new model is released. Most new releases will not meet your specific requirements. Recognising this early will stop you from spending time on unnecessary evaluations.
Establish governance based on the current functions of a system rather than its predicted future capabilities. You should manage each new model like a new employee who has an unverified reference. Assign it a narrow scope with oversight initially. Only expand its responsibilities once it provides evidence of reliability. This practice stays effective whether models become more general over the next five years or their capabilities stay the same.
Most of the actual return on AI in the next few years will come from applying narrow systems effectively against a data foundation that can support them, the unglamorous work that enterprise AI programmes tend to underestimate at the planning stage and then rediscover the hard way mid-project. If your business is trying to determine what its AI plans should actually depend on, rather than what a slide about AGI implies they should depend on, talk to us. We would prefer to help you scope this year's real capability gap rather than help you write a strategy slide for a date nobody can give you.
Many AI projects fail because teams have difficulty applying these concepts effectively. Our AI consultancy helps clients overcome this specific challenge.
Frequently asked questions
What is artificial general intelligence?
Artificial general intelligence is a system that can use learned information to solve new problems. It would perform as well as or better than a human on a wide range of cognitive tasks. Transferring learning, which means applying the same reasoning skills to tasks such as cooking, bicycle repair, or reviewing contracts, is the most important part of this definition.
How would artificial general intelligence work?
This article does not identify a specific mechanism because no one has built one yet. Researchers state that the defining property of this technology is transfer. This involves learning from one domain and applying that knowledge to a different one without needing to retrain the system. Current systems, such as large language models, perform pattern prediction based on training data rather than this type of general reasoning.
What are examples of artificial general intelligence?
There are no functional examples because artificial general intelligence does not exist yet by any standard definition. This article uses a human comparison to explain the concept. A person who learns to read a balance sheet does not start from the beginning when they later learn to read a tenancy agreement. AGI would need to demonstrate this ability to transfer knowledge between unrelated domains.
Is ChatGPT artificial general intelligence?
No, according to the article's argument: large language models are genuinely broad, drafting, summarising, coding, and holding a conversation on almost any topic, but breadth of surface task is not the same as the transfer and reasoning the AGI definition seeks. A model can write a competent legal summary and still fail at a much simpler change to the same problem.
When will artificial general intelligence arrive?
There is no definitive answer to this question. When researchers are surveyed, their estimates range from decades to centuries. Some believe current methods will never achieve the goal. These predictions depend on assumptions that cannot be verified at this time. The article states that both short-term predictions and warnings of immediate risk often come from people with specific incentives. Readers should consider these motivations before trusting such claims.