Artificial intelligence has progressed so rapidly during the 2020s that the question of artificial general intelligence has shifted from the margins of computer science into mainstream technological debate.
Only a few years ago, artificial general intelligence — usually abbreviated to AGI — was largely hypothetical. Today, leading AI laboratories explicitly discuss AGI as an engineering objective, while frontier models can write software, analyse scientific papers, solve sophisticated mathematical problems, operate computers, conduct research and coordinate multi-step workflows.
Yet enormous uncertainty remains about what AGI actually means, whether current architectures can achieve it, and how far existing systems remain from genuinely general intelligence.
The central difficulty is that impressive performance on individual benchmarks does not necessarily imply intelligence comparable to a capable human across the enormous range of situations humans encounter. Modern AI remains strikingly uneven: extraordinarily capable in some domains while surprisingly unreliable in others.
The road to AGI therefore appears less like crossing a single technological threshold and more like solving a collection of interconnected problems involving reasoning, memory, learning, planning, generalisation, reliability, agency, computing infrastructure and safety.
As of October 2026, there is credible evidence that several of these barriers are falling rapidly. Others may require fundamental architectural innovations.
The resulting timeline could range from the late 2020s for systems possessing some characteristics commonly associated with AGI to several decades for systems satisfying the most demanding definitions.
What Would Actually Count as AGI?
Before examining the technological obstacles, there is an immediate conceptual problem: researchers do not agree on exactly what constitutes AGI.
Definitions range from relatively modest to extraordinarily demanding.
One definition might describe AGI as an AI capable of performing most economically valuable cognitive work at roughly human professional ability. Another might require a system capable of learning almost any intellectual task that a human can learn. Still stronger definitions require AI to outperform humans across essentially every cognitive domain while adapting autonomously to unfamiliar environments.
Google DeepMind has attempted to make the discussion more rigorous by breaking general intelligence into cognitive capabilities. Its 2026 framework identifies ten major faculties: perception, generation, attention, learning, memory, reasoning, metacognition, executive function, problem solving and social cognition. DeepMind argues that evaluating AGI requires measuring performance across all these dimensions rather than relying on a small collection of benchmarks.
That distinction matters.
An AI could be better than most humans at programming, mathematics and scientific literature analysis while remaining unreliable at planning an unfamiliar project lasting several weeks. Should that system qualify as AGI?
There is no universally accepted answer.
This definitional ambiguity will probably remain even after extraordinarily capable systems emerge. In that sense, one barrier to “achieving AGI” is not technological at all: humanity may never agree upon a single moment at which AGI has arrived.
Barrier One: Generalisation Beyond Training
Perhaps the most fundamental technical problem is generalisation.
Modern foundation models learn statistical structures from enormous training datasets. This allows them to perform tasks their developers did not explicitly program them to perform.
That is already a remarkable form of generalisation.
However, genuinely general intelligence requires something stronger: the ability to confront situations substantially different from previous experience and develop appropriate solutions.
Humans do this constantly.
A person transported into an unfamiliar building, organisation or country can combine prior knowledge, observation, analogy and experimentation to work out what is happening. Human intelligence is particularly powerful because knowledge learned in one context can often be transferred into another.
AI remains much more fragile.
One indication comes from ARC-AGI, a family of benchmarks intended to test adaptation to unfamiliar problems rather than knowledge retrieval. Frontier models have improved dramatically. By July 2026, Claude Opus 5 achieved approximately 97.5% on ARC-AGI-1 and 90.4% on ARC-AGI-2 under its highest reasoning setting.
But on the substantially more difficult ARC-AGI-3 environment, its verified score was only about 30.2%.
That contrast illustrates one of the defining characteristics of current AI development: yesterday’s seemingly impossible benchmark can become almost saturated while a newly designed test immediately exposes substantial weaknesses.
Solving generalisation may require better reasoning algorithms, richer world models, improved reinforcement learning, synthetic environments or architectures capable of constructing abstractions rather than primarily predicting patterns.
Significant progress seems likely during the late 2020s and early 2030s, although robust human-like generalisation across essentially arbitrary situations could take considerably longer.
Barrier Two: Long-Horizon Reasoning and Planning
Current AI systems are becoming excellent at solving problems lasting minutes or even hours.
AGI may need to pursue objectives lasting days, months or years.
This creates a very different computational problem.
Long projects require remembering earlier decisions, detecting mistakes, revising strategies, maintaining priorities, managing resources and dealing with unexpected developments.
Small errors also compound.
Imagine an AI system that performs each step in a complex task correctly 99% of the time. After hundreds or thousands of consequential decisions, that apparently impressive reliability may become inadequate.
Research by METR attempts to quantify this challenge by measuring the duration of tasks that AI agents can successfully complete. The organisation’s task-completion methodology evaluates frontier models on software engineering, machine learning and cybersecurity work that would take skilled humans varying amounts of time.
METR stresses that the resulting measurements should not be interpreted as evidence that AI can perform all human work of the equivalent duration. Its tasks are unusually well specified, predominantly technical and easier to evaluate automatically than messy real-world professional activity.
Nevertheless, the trend has been striking.
METR reports an approximately exponential increase in the length of tasks frontier agents can complete, while also warning that real employment involves tacit knowledge, interpersonal interaction and ambiguous success criteria that are poorly represented in current benchmarks.
Berkeley researchers reached a similar conclusion in their 2026 Agents’ Last Exam evaluation. Frontier agents could perform meaningful portions of professional work but remained far behind humans on difficult tasks requiring sustained reasoning, specialist knowledge and reliable long-horizon execution.
The trend suggests that the long-horizon problem may improve substantially during 2027–2032.
Achieving high reliability on projects lasting weeks or months, however, could take significantly longer.
Barrier Three: Continual Learning
Humans continually modify their internal models of the world.
If you learn something today, that information can affect decisions tomorrow.
Most contemporary language models operate differently.
The majority of their fundamental knowledge is established during training. Information can be added through context windows, retrieval systems, external databases and persistent memory, but the underlying model does not normally modify itself continuously in response to every experience.
This distinction is crucial.
An AGI operating for years should ideally accumulate expertise just as humans do. If an AI lawyer handles 10,000 matters, for example, the 10,001st matter should benefit from lessons extracted from those previous experiences.
The system would also need to distinguish genuine lessons from misleading information.
Naive continual learning creates several risks.
Models can forget previous skills when learning new ones, a phenomenon known as catastrophic forgetting. They can absorb incorrect information. Malicious actors could deliberately contaminate their learning. And uncontrolled self-modification creates substantial safety challenges.
Future systems may therefore separate different forms of memory: stable foundational knowledge, episodic experience, external databases and carefully validated model updates.
The AAAI Presidential Panel on AGI identified continual learning as one of the significant unresolved capabilities separating existing systems from more general intelligence.
This barrier appears technically tractable, and substantial progress is plausible within roughly two to seven years. Fully autonomous lifelong learning that remains stable, secure and aligned may take longer.
Barrier Four: Memory
Memory and continual learning are related but distinct problems.
A sufficiently capable AGI would probably require enormous working and long-term memory.
Human intelligence relies on multiple memory systems. We maintain temporary working memory, factual knowledge, memories of experiences, habits and contextual understanding accumulated over decades.
AI systems increasingly emulate these capabilities through long context windows, retrieval-augmented generation, vector databases, knowledge graphs and external storage.
But simply expanding context windows does not solve memory.
A system must determine what should be remembered, what can be discarded, which memories are reliable and which earlier experiences are relevant to the current problem.
There are also physical engineering constraints. A 2026 review of LLM inference found that as model sizes and context windows increase, memory movement and input/output bandwidth increasingly become major bottlenecks rather than raw arithmetic alone.
In other words, intelligence at scale is becoming partly a memory architecture problem.
This is one area where rapid engineering progress is likely. Hierarchical memory systems combining local context, compressed histories, databases and learned retrieval could provide increasingly human-like persistent memory before the end of the decade.
A reasonable window for major improvement is 2027–2030.
Barrier Five: Metacognition
Humans can think about their own thinking.
We recognise uncertainty. We sometimes realise that we do not understand something. We check our calculations. We seek expert advice when a problem exceeds our knowledge.
AI systems exhibit partial versions of these behaviours but remain poorly calibrated.
A model can produce an incorrect answer with exactly the same linguistic confidence it displays when producing a correct one.
This is one reason hallucinations are so problematic.
AGI needs more than factual accuracy. It requires reliable awareness of uncertainty.
An intelligent system should be able to say:
“I have insufficient information.”
“This conclusion depends upon an assumption.”
“I should verify this result.”
“This problem requires a different strategy.”
DeepMind’s AGI taxonomy explicitly treats metacognition as a core component of general intelligence and identifies it as one of the capabilities where better evaluations are particularly needed.
Reasoning models already use techniques resembling metacognitive checking, including verification loops, multiple solution attempts and tool-assisted validation.
This suggests substantial improvement could occur relatively quickly, perhaps during 2027–2031.
Perfect self-awareness, however, should not be expected. Humans are themselves frequently overconfident.
Barrier Six: Reliable Agency
Intelligence and agency are different.
A chatbot answering questions can be extremely intelligent without independently pursuing objectives.
An autonomous agent must interact with computers, websites, databases, robots and other people while keeping track of goals.
That enormously increases the consequences of mistakes.
An unreliable chatbot may generate an incorrect paragraph.
An unreliable autonomous agent controlling corporate infrastructure could delete information, send incorrect instructions or make expensive decisions.
The transition from conversational AI to agentic AI therefore requires a dramatic increase in reliability.
Experiments already show that frontier agents can use software tools, write and execute programs, inspect environments and recover from some failures. Yet research into long-horizon agents continues to find unstable performance, limited novelty and imperfect reuse of previous experience.
This is arguably one of the most important practical barriers to economically transformative AGI.
Systems do not merely need to achieve 90% or 95% success rates. High-consequence autonomous activity may require reliability closer to industrial or safety-critical engineering standards.
Useful agentic systems will continue improving rapidly through the late 2020s.
Highly dependable autonomous agents capable of operating for extended periods with minimal supervision may emerge during the early-to-mid 2030s, though some domains could mature earlier.
Barrier Seven: World Models and Physical Understanding
Language contains enormous information about reality, but language is not reality itself.
Human intelligence developed within physical environments. Humans learn intuitive physics by manipulating objects, moving through space and observing consequences.
AI researchers increasingly attempt to reproduce this capability through multimodal models and world models: systems that learn representations of how environments evolve.
Video generation, robotics and autonomous driving are particularly important because they force AI systems to predict physical reality rather than merely describe it.
Whether embodiment is necessary for AGI remains debated.
If AGI simply means superior performance across non-physical intellectual work, sophisticated robotics may not be necessary at all.
If AGI means intelligence possessing approximately the same breadth of capabilities as a human being, physical interaction becomes considerably more important.
Robotics also has a slower development cycle than software because physical systems must deal with hardware, manufacturing, energy, safety and unpredictable environments.
For this reason, digital AGI could plausibly precede broadly capable embodied intelligence by many years.
Strong world models may mature during the late 2020s and 2030s, while genuinely general-purpose robotics could remain a longer-term challenge.
Barrier Eight: Data
The spectacular development of foundation models was partly enabled by the enormous amount of digital information generated by civilisation.
But high-quality human-created data is finite.
Training increasingly sophisticated models simply by consuming larger collections of internet text eventually encounters diminishing returns.
This has encouraged several alternatives.
Synthetic data can be generated by existing models. AI systems can learn from simulations. Reinforcement learning can generate experience dynamically. Models can interact with software environments and receive feedback. Scientific AI can generate and test hypotheses.
The learning process is therefore gradually shifting from passively consuming human-generated data toward actively generating experience.
This transition may be critical to AGI.
If future AI can conduct millions of experiments within simulated environments, inspect the results and learn from them, the effective quantity of training experience could become enormous.
The data barrier therefore appears significant but probably not permanent.
It is likely to be progressively overcome during 2026–2032 through synthetic data, simulation, reinforcement learning and interactive environments.
Barrier Nine: Compute, Energy and Economics
Even if the algorithms required for AGI already existed, running them economically could remain difficult.
Frontier AI requires enormous data-centre infrastructure.
Training is computationally demanding, but inference at global scale may ultimately require even greater resources because models must serve billions of interactions continuously.
Energy supply, semiconductors, memory bandwidth, networking, cooling systems, transformers and grid connections can therefore constrain progress.
These problems are receiving extraordinary investment.
The industry is developing specialised accelerators, custom inference chips, improved memory systems, more efficient architectures, quantisation and increasingly sophisticated mixtures of large and small models.
The relevant question is therefore not whether AI demands large amounts of computing infrastructure — it clearly does — but whether capability improvements can continue faster than the economic cost of delivering those capabilities.
Some analyses warn that enormous data-centre investments will eventually confront financial and infrastructure constraints. Others expect continued improvements in hardware and algorithms to reduce the amount of computation required for any particular level of intelligence.
Both dynamics can occur simultaneously.
AGI may therefore be achieved in a research environment before it becomes inexpensive enough for mass deployment.
Hardware and energy are unlikely to constitute an absolute barrier, but they could shift timelines by several years.
Barrier Ten: Alignment and Security
There is another distinction that is frequently overlooked:
Creating AGI and safely deploying AGI are not necessarily the same event.
A laboratory might develop a system possessing extraordinarily broad capabilities while refusing to provide it with unrestricted autonomy because its behaviour cannot yet be adequately controlled.
The more capable an agent becomes, the more important issues such as goal specification, interpretability, cybersecurity and alignment become.
NIST research published in 2026 argues that there are fundamental information-theoretic limits to guaranteeing robustness and alignment under every possible condition, strengthening the case for continuous monitoring and updating rather than assuming perfect security can ever be achieved.
This means the safety problem may never be completely “solved”.
Instead, advanced AI may eventually resemble aviation, cybersecurity or medicine: sophisticated systems surrounded by layers of monitoring, testing, regulation, redundancy and incident response.
Safety could consequently become the factor determining when AGI moves from laboratory capability to widespread autonomy.
So When Will AGI Arrive?
Forecasting AGI requires extraordinary humility because definitions differ enormously.
Nevertheless, several useful data points exist.
The most recent large Expert Survey on Progress in AI was published by AI Impacts in September 2026 and reports responses from 1,580 researchers at leading AI venues. Its definition of high-level machine intelligence asks when unaided machines will be capable of accomplishing every task better and more cheaply than human workers.
The aggregate forecast places a 50% probability around 2042, although the survey was conducted in December 2024 and therefore does not incorporate the substantial advances of 2025 and 2026.
A separate 2026 Longitudinal Expert AI Panel used an even more economically demanding definition: commercially available AI outperforming the 90th-percentile professional across virtually all non-physical occupations while costing no more than five times equivalent human labour.
Under that definition, experts produced a conditional median year of approximately 2050, while superforecasters produced 2047. The same panel expected major improvements in long-horizon AI task completion much sooner, with the median expert forecast placing 80% performance on eight-hour technical tasks around 2030.
The difference is revealing.
AI capable of performing impressive multi-hour intellectual work could appear years or decades before AI satisfying an exceptionally demanding definition of general intelligence.
A Plausible Development Sequence
Rather than identifying one AGI date, it may therefore be more useful to think in stages.
2026–2029: AI agents become dramatically better at reasoning, coding, research, memory and computer use. Persistent memory and multimodal world models improve. Increasing amounts of routine professional work become partially autonomous.
2029–2033: Long-horizon agents become capable of completing projects that would require skilled humans many hours or days. Continual learning, self-verification and organisational memory become significantly more sophisticated.
2032–2040: Systems may begin satisfying weaker definitions of AGI — broadly capable digital agents able to perform a very large fraction of cognitive economic activity at human or superhuman levels.
2040–2055: More demanding versions of AGI could emerge, combining reliable generalisation, extensive autonomy, continual learning and performance comparable with highly skilled humans across almost every non-physical occupation.
These ranges should not be interpreted as precise predictions.
A fundamental algorithmic breakthrough could compress them dramatically.
Conversely, diminishing returns from scaling, infrastructure constraints, safety problems or previously undiscovered limitations of current architectures could push them much further into the future.
The Most Important Unknown: Do We Need a New Paradigm?
Ultimately, almost every AGI timeline depends upon one unresolved question:
Can current machine-learning architectures be extended into AGI, or is a fundamentally different approach required?
If transformers, reinforcement learning, tool use, multimodality, memory systems and agent architectures contain most of the necessary ingredients, the remaining barriers may primarily be engineering problems.
In that scenario, AGI could emerge surprisingly quickly.
If today’s systems lack essential mechanisms for abstraction, causality, continual learning or genuine generalisation, another major conceptual breakthrough may be required.
The history of AI provides reasons to consider both possibilities.
Scaling neural networks produced capabilities that many researchers had previously assumed would require entirely new theories of intelligence. But benchmark success continues to expose new limitations once evaluation becomes more demanding.
Even researchers remain divided. The AAAI Presidential Panel reported substantial scepticism that simply scaling existing approaches would be sufficient to produce AGI.
At the same time, organisations such as Google DeepMind now describe human-level AGI as a concrete target for the coming decade rather than distant speculation.
Conclusion
The greatest obstacles to artificial general intelligence are no longer simply making neural networks larger.
They are creating systems that can generalise, remember, learn continuously, plan over long periods, recognise their own uncertainty, interact reliably with the world and remain safe while doing so.
Progress across almost all these dimensions is already occurring.
That makes an important conclusion possible.
There is no obvious scientific principle demonstrating that AGI is impossible, nor is there convincing evidence that present systems have already achieved it.
Instead, humanity appears to be somewhere in the complicated middle ground.
Some of the necessary components are becoming extraordinarily powerful. Others remain immature.
The late 2020s are therefore likely to provide crucial evidence.
If long-horizon reasoning, continual learning and generalisation improve at anything resembling the pace seen in foundation models during the first half of the decade, the debate may rapidly shift from whether AGI can be created to determining precisely when increasingly general systems deserve the name.
A reasonable synthesis of the available evidence is that AGI-like digital systems could plausibly emerge during the 2030s, while systems satisfying the strongest definitions of economically competitive human-level general intelligence may be more likely to appear in the 2040s or beyond.
But the range of credible outcomes remains extremely wide.
AGI could arrive much sooner following an architectural breakthrough. It could also prove substantially harder than current scaling trends imply.
The uncertainty itself may be the most important fact.
For governments, businesses and society, preparing only after a consensus definition of AGI has been satisfied would probably be too late. The economic, legal and safety consequences will emerge gradually as AI systems acquire individual components of general intelligence — long before anyone can definitively point to a particular model and declare that AGI has arrived.
Leave a Reply