AGI has no universal definition.
Different frameworks propose different thresholds; there is no single recognised test that can declare a system to be AGI.
Technological frontier · Critical perspective
AI safety and AGI.
“Artificial general intelligence”, benchmark performance, autonomy and reliability are not interchangeable. This page separates the concepts and explains why safety measures must grow with a system’s capabilities and access.
Primary sourcesLive benchmark linksWork-centred analysis
Destruction · 1836
In the fourth painting of The Course of Empire, a city that appeared all-powerful collapses into chaos. The work is not used here as a catastrophic prediction about AI, but as a reminder: when technical power advances faster than our ability to control it, failures can be amplified rather than contained. AI safety is about identifying those weaknesses, limiting their consequences and preserving the ability to intervene before a crisis. View the work ↗
Scaling laws are observations.
They describe measured trends as data, model size and compute increase. They do not guarantee general intelligence.
A high score remains local.
Each benchmark measures particular tasks under particular conditions. It does not summarise reliability, usefulness or safety.
Control must precede autonomy.
The more a system can act, the more explicit its access limits, approvals, monitoring, shutdown and accountability must be.
01 / Define without oversimplifying
What do people mean by artificial general intelligence?
AGI is a research and forecasting concept, not a settled technical category. Definitions differ on the human reference level, the range of tasks and the degree of autonomy required.
Outperform humans at most economically valuable work.
OpenAI’s Charter defines AGI as highly autonomous systems that outperform humans at most economically valuable work. This framing emphasises broad usefulness and autonomy.
Read the OpenAI Charter ↗Separate breadth, performance and autonomy.
Google DeepMind researchers propose levels based on the range of tasks a system can perform and its level of performance. Autonomy is considered separately because it changes the risk of deployment.
Read “Levels of AGI” ↗What AGI does not automatically mean
A fluent conversation, an excellent mathematics score or success on one autonomous task does not establish general intelligence, consciousness or human-like understanding of the world.
02 / Scaling laws
Why more compute often improves model performance.
Scaling laws describe empirical relationships: on some training metrics, error decreases fairly regularly when model size, data and compute increase together.
Training computeMore operations generally allow larger models or more data, while increasing financial, material and energy costs.
Model size and dataCompute-optimal training research shows that increasing model size alone is inefficient: data must scale with it.
Data quality and post-trainingData selection, instruction tuning and human or automated feedback strongly shape the final behaviour.
Inference-time computeReasoning systems may spend more computation on a problem, try several approaches and use tools, at a higher cost and latency.
03 / Capabilities
What powerful general-purpose models can do — and what still needs checking.
There is no universally “best model”. Suitability depends on the task, allowed tools, cost, latency, error tolerance and the degree of human control required.
Complex problems
Decompose questions, compare hypotheses, calculate and explain a process across many domains.
Code and documents
Draft, modify and test code; summarise long records; prepare analyses and professional documents.
Several media
Depending on the model, analyse text, images, documents, audio or video within one task.
Tools and interfaces
Search, run code, call software or operate an interface when an application grants access.
Longer tasks
Plan several steps, maintain context and correct some mistakes without becoming consistently reliable.
Uneven reliability
Produce excellent work and then fail on a simple step, invent a source or pursue the wrong interpretation of a goal.
04 / Current benchmark families
Measure a specific capability, not a single score for intelligence.
The links below point to maintained project pages and dashboards, so that results can be read with their current protocols rather than frozen into a ranking that will quickly become obsolete.
01ARC-AGIAbstract generalisationSolve unfamiliar visual problemsTests whether a system can infer a rule from a few examples. Useful for generalisation, but too narrow to establish AGI by itself.Official leaderboard ↗
02Humanity’s Last ExamExpert reasoningDifficult questions across disciplinesMeasures high-level knowledge and reasoning, including multimodal questions; it does not measure the whole of real work or autonomy.Project page ↗
03SWE-benchSoftware engineeringResolve real issues in code repositoriesPerformance depends heavily on the agent, tools and protocol surrounding the model.Official results ↗
04METR time horizonTask durationEstimate the length of tasks an agent can completeA longer time horizon does not by itself demonstrate general autonomy.Current research ↗
05HELMStructured evaluationCompare several capabilities under visible protocolsProvides more context on scenarios, prompts and metrics than a single aggregate score.Dashboard ↗
05 / Why AI safety matters
The more capable a system becomes, the farther an error or misuse can travel.
AI safety covers work to prevent, detect and reduce harm from AI systems, from routine reliability failures to severe risks enabled by greater capability and autonomy.
Dual-use capabilities
Code, scientific assistance and automation can support legitimate work or make attacks easier. Access and high-risk use therefore need specific controls.
A convincing but wrong action
A system may misunderstand an instruction, fabricate information or use a tool incorrectly. Automation can propagate a small error before it is noticed.
More steps, less direct oversight
When an agent plans and acts for longer, it becomes harder to anticipate each action and detect when the objective has been misread.
Information and persuasion
Personalised content at scale can amplify fraud, manipulation and disinformation, while polished language may invite excessive trust.
Concentration and skill loss
Provider dependence, version changes, outages and gradual deskilling can create systemic organisational risk.
Safety includes occupational health
A technically safe model can still intensify work, reduce autonomy, shift responsibility or impose continuous monitoring. These effects belong in AI safety.
A simple principle
Safety is not about predicting the date of AGI. It is about checking, before each increase in capability or autonomy, that dangerous behaviour can be detected, consequences limited and decisions challenged.
06 / From power to control
Five layers of control before a system is allowed to act.
Model evaluations are necessary but insufficient. They must be combined with technical, organisational and human measures suited to the real context of use.
Test what could go wrong
Capabilities, severe errors, misuse, robustness, cybersecurity and effects on work.
Grant minimum access
Bound data, tools, budget, duration, number of actions and operational scope.
Keep human review meaningful
An identified person approves sensitive, irreversible or consequential decisions.
Make incidents visible
Logs, alerts, worker feedback, version tracking and an accessible reporting channel.
Be able to stop and learn
A shutdown process, fallback route, incident analysis and reassessment after major changes.
NIST AI Risk Management Framework
A voluntary framework for governing, mapping, measuring and managing AI risks throughout the system lifecycle.
Open the framework ↗Assess what changes before and after.
Use the short questionnaire to structure a discussion about psychosocial risks, opacity, oversight burden and skill retention before a pilot.
Open the short questionnaire →