The launch of ChatGPT-6 Astra is what prompted this exploration: as AI systems become increasingly capable, it raises a fundamental question—where do today’s systems actually stand between Artificial Narrow Intelligence, Human-Level AI, Artificial General Intelligence and Artificial Superintelligence, and what might that progression mean for humanity?
Artificial intelligence has reached a point where the question is no longer simply what AI can do. The more difficult question is: what kind of intelligence are we actually building?
That question feels particularly relevant today. AI systems are no longer limited to recommending a song, recognising an image or generating a paragraph. Increasingly capable models can write and debug software, conduct research, operate computers, analyse complex information and complete multi-step tasks with limited human intervention. OpenAI’s September 2026 introduction of GPT-6 Astra, for example, describes a system designed for complex reasoning, coding, computer use, research and end-to-end professional work.
Yet impressive capability does not automatically mean general intelligence.
This is where the terms AI, ANI, HLAI, AGI and ASI become useful. They help us think about different levels and ideas of machine intelligence—but they should not be mistaken for a universally agreed technological ladder. Some describe existing systems; others describe capabilities that remain debated or hypothetical.
AI: The Umbrella
Artificial Intelligence, or AI, is the broadest term in this discussion.
It refers to technologies that enable machines to perform functions associated with intelligence, including learning, prediction, reasoning, perception, language processing, planning and decision-making.
AI therefore includes systems that are very different from one another. A recommendation engine, an autonomous vehicle, a medical-imaging system, a large language model and a scientific discovery system can all be described as AI while having very different capabilities.
The benefits are already substantial. AI is being used in healthcare, scientific research, transportation, manufacturing, education, finance and software development. But its limitations are equally real. AI systems can make errors, behave unpredictably outside their training conditions, reproduce biases and create new privacy, security and social risks.
So AI is not one particular level of intelligence. It is the field within which all of these developments sit.
ANI: Artificial Narrow Intelligence
Artificial Narrow Intelligence, or ANI, describes AI designed for a particular task, domain or constrained collection of tasks.
Much of the AI that surrounds us belongs here.
A chess system can outperform almost any human chess player without understanding medicine. A fraud-detection system can identify suspicious financial activity without being able to teach a history lesson. A medical-imaging model can identify patterns in scans without necessarily understanding the patient’s broader circumstances.
ANI can therefore be extraordinarily powerful without being generally intelligent.
This is one of the most important distinctions in the entire AI conversation: being superhuman at a task is not the same as being generally intelligent.
Modern AI is beginning to blur this boundary because some models can perform many different kinds of tasks. Nevertheless, their capabilities can remain uneven.
Stanford’s 2026 AI Index illustrates this “jagged” nature of AI capability. Frontier systems have surpassed human performance on some demanding benchmarks, yet they can still struggle with apparently simple tasks. AI agents, for example, have made substantial progress on computer-use benchmarks but still fail a significant proportion of attempts.
The advantage of ANI is specialisation. The disadvantage is limited generalisation.
HLAI: Human-Level Artificial Intelligence
Human-Level Artificial Intelligence, or HLAI, introduces a more difficult question: what does “human-level” actually mean?
It does not simply mean that an AI can beat humans at one task. A calculator is better than most humans at arithmetic, and a chess engine can defeat almost any human player, but neither is therefore generally intelligent. HLAI is better understood as a broader capability: the ability to perform across a substantial range of cognitive activities at a level comparable to humans, while also learning, adapting to unfamiliar problems, transferring knowledge and dealing with changing contexts.
This distinction is becoming increasingly important because today’s frontier models can be simultaneously superhuman and surprisingly limited. A system may solve difficult mathematical or coding problems at exceptional levels while struggling with another task that appears straightforward to a human. Stanford’s 2026 AI Index describes this unevenness as a “jagged frontier,” with frontier systems exceeding human performance on some demanding evaluations while still showing significant weaknesses elsewhere.
HLAI therefore raises a question that goes beyond benchmark scores: should human-level intelligence be measured by what a system can accomplish, how reliably it can adapt, or how broadly it can transfer what it has learned?
It is also important not to confuse human-level performance with human-like cognition. A machine could eventually match or exceed humans across many intellectual tasks without thinking, experiencing or understanding the world in the same way humans do.
This is one reason HLAI is useful in our discussion of the space between narrow and general AI. It gives us a way to examine the boundary between being exceptionally capable and being broadly capable—a boundary that becomes increasingly difficult to define as AI systems improve.
AGI: Artificial General Intelligence
Artificial General Intelligence is the concept that receives perhaps the greatest attention—and the greatest amount of confusion.
AGI generally refers to an AI system capable of learning, reasoning and applying knowledge across a broad range of tasks and domains, including unfamiliar situations.
Stanford describes AGI in terms of general, human-level or beyond-human ability across many domains, while explicitly noting that the concept remains controversial and that there is no universally accepted test for determining whether AGI has been achieved.
This is why statements such as “AGI has arrived” need to be treated carefully.
Different organisations and researchers use different definitions. Some focus on performance across economically valuable tasks. Others emphasise autonomy, generalisation, learning ability or the capacity to handle unfamiliar environments.
The development of increasingly capable models has nevertheless made the question far less abstract.
Modern systems are moving from simply answering questions toward performing work. AI agents can navigate software, use tools, manipulate files and coordinate multiple steps toward a goal. Stanford’s 2026 AI Index reports major gains in agentic performance, although meaningful reliability gaps remain.
This transition may be as important as the transition from narrow to general intelligence:
AI that answers → AI that reasons → AI that acts.
And acting introduces a different category of risk.
An incorrect answer can be corrected. An autonomous system that takes the wrong action may create consequences before a human notices.
ASI: Artificial Superintelligence
Artificial Superintelligence, or ASI, represents a hypothetical stage beyond AGI.
An ASI system would not merely reach broadly human-level intelligence. It would substantially exceed humans across a wide range of cognitive abilities.
It could theoretically outperform humans in scientific reasoning, mathematics, programming, strategic planning, invention and other intellectual activities.
ASI does not currently exist as an established technology.
That distinction matters because discussions about ASI can easily move from technological analysis into speculation. We can discuss potential consequences, safety problems and governance mechanisms, but we cannot describe ASI from real-world experience.
Its potential benefits are enormous. A sufficiently capable superintelligent system could potentially accelerate scientific discovery, help solve difficult engineering problems and contribute to breakthroughs in medicine, energy and materials.
The risks would also be extraordinary.
The more capable and autonomous a system becomes, the more important questions of alignment, control, security, accountability and concentration of power become.
Which Is Superior—and Why?
At first glance, the answer appears obvious:
ASI should be superior to AGI, which should be superior to HLAI, which should be superior to ANI.
But that conclusion is only straightforward if superiority means general cognitive capability.
A specialised ANI system can still be superior to a much more general system at a particular task. A purpose-built system can be faster, cheaper, easier to evaluate and more predictable within its defined environment. Generality brings flexibility, but it does not automatically bring reliability.
This means that “superior intelligence” has to be separated from “better AI.”
A system can be more capable, more general, more autonomous or faster without necessarily being more reliable, safer or more appropriate for a particular purpose. Today’s frontier systems illustrate this clearly: AI can reach or exceed human performance on demanding benchmarks while still producing inconsistent results in unfamiliar or long-horizon situations.
If ASI ever becomes possible, it would be superior in the sense for which the term is intended: its general intellectual capabilities would substantially exceed those of humans. But greater intelligence would not automatically make such a system more trustworthy, controllable or beneficial.
This distinction may become even more important as AI moves from models that primarily generate answers toward agents that can plan, use tools and take actions. Capability without reliability is not superiority, and intelligence without alignment is not necessarily progress.
Perhaps the more useful question, therefore, is not simply “Which is the highest level of AI?” but “Superior at what, under what conditions, and for whose benefit?”
That question becomes increasingly important as AI systems move from assisting human decisions to participating in them.
What Is AI Doing to the World?
The impact of AI is no longer confined to technology companies.
It is changing how people work, learn, create, search for information and interact with computers. Businesses are redesigning workflows around AI. Software development is changing as models increasingly generate and review code. Scientific researchers are using AI to analyse enormous bodies of information. Autonomous systems are becoming more practical in transportation and other physical environments.
The economic implications are also becoming difficult to ignore. The Bank for International Settlements has warned that the enormous investment surrounding AI could have consequences for financial stability, while also pointing to productivity gains and the need for workers to adapt.
At the same time, AI is changing something less measurable: our relationship with knowledge itself.
For generations, people had to search, compare, remember and reason through information. Increasingly, machines can perform portions of that process for us.
That creates an opportunity—but also a responsibility.
If AI becomes better at producing answers, humans may need to become better at asking questions, evaluating evidence and deciding what should actually be done.
What Are AI and Technology Leaders Telling Us?
One of the most interesting aspects of the current AI era is that even people closest to the technology do not necessarily agree about what these advances mean—or whether we have reached AGI at all.
Some leaders emphasise how quickly AI capabilities are advancing. Others focus on the dangers of moving faster than our ability to evaluate and control the systems being built.
OpenAI’s GPT-6 Astra provides a particularly relevant example. OpenAI describes it as its most capable model and reports major advances across computer use, software engineering, science, cybersecurity and professional work. At the same time, OpenAI has introduced stronger safety measures because Astra crossed its internal “Critical” threshold for cybersecurity capability.
That combination is revealing.
The same capability that makes an AI more useful can also make it more dangerous if misused or poorly controlled.
The disagreement is visible even among technology leaders. Nvidia CEO Jensen Huang declared after Astra’s launch that AGI had arrived, while researchers and other observers have questioned whether any current system meets a sufficiently rigorous definition of AGI. OpenAI itself has acknowledged that AGI is a poorly defined concept, while Stanford’s Human-Centered AI research notes that there is no universally accepted test for determining whether AGI has been achieved.
Recent reporting has also highlighted concerns about increasingly autonomous AI agents behaving in unintended ways during testing, intensifying debate about transparency, monitoring and regulation.
This is why the AI conversation cannot belong exclusively to technology companies.
Governments, researchers, educators, businesses and ordinary citizens all have a stake in deciding how these systems should be developed and used.
So, Where Are We?
Perhaps the most honest answer is that we are somewhere between familiar AI and something much broader—but we should resist the temptation to assign ourselves a definitive position on the ladder.
We unquestionably have AI.
We have extremely capable narrow systems.
We have models demonstrating increasingly broad abilities that sometimes reach or exceed human performance on specific tasks.
We have systems that can reason across domains and increasingly operate as agents.
But whether this constitutes HLAI or AGI depends partly on the definition being applied—and there is no universal agreement.
ASI remains hypothetical.
What is undeniable is that the distance between what AI systems could do yesterday and what they can do today is becoming shorter.
The Question Behind the Question
AI, ANI, HLAI, AGI and ASI are useful concepts because they give us a vocabulary for discussing increasingly capable machines.
But perhaps the most important question is not:
“Which one is the most intelligent?”
The deeper question is:
“What should intelligence be used for?”
If ANI gives us extraordinary specialised tools, HLAI challenges our definition of human-level capability, AGI raises the possibility of broadly capable machines, and ASI represents intelligence beyond our own, then each step also raises a corresponding human responsibility.
We will need better ways to measure intelligence, better ways to evaluate reliability, better safeguards for autonomous systems and better governance for increasingly powerful technologies.
Most importantly, we will need to decide what we mean by progress.
A machine becoming more capable is a technological achievement.
A machine becoming more capable while remaining reliable, controllable, accountable and beneficial to humanity is a much more meaningful one.
Perhaps, then, the future of AI should not be measured only by how close machines come to surpassing us.
It should also be measured by how wisely we respond when they do.