Frequently asked questions about AGI

AGI is a system that can learn and reason across the full range of cognitive tasks at or above human level, transferring knowledge between domains without being retrained for each one.

No. Large language models are extraordinarily capable pattern learners with emergent reasoning. They are not yet general agents - they do not learn continually, maintain persistent goals over long horizons, or robustly transfer skills the way humans do.

There is no consensus. Forecasts from researchers and labs range from a few years to several decades. The honest answer is that we do not know - and any specific date should be treated with skepticism.

Advanced AI systems present real risks - misuse, accidents, misalignment with human values, concentration of power - alongside potential benefits. AI safety is a serious field working on these problems.

Unknown. Consciousness and general intelligence are conceptually distinct, and we lack a settled scientific theory of what produces subjective experience. A system could in principle be highly intelligent without being conscious, or vice versa.

Significant labour-market disruption is likely. The honest assessment is that some jobs change, some disappear, and new ones emerge - on timescales and in proportions that are hard to predict. Adaptation will matter more than prediction.

The research problem of ensuring advanced AI systems pursue goals and values that are genuinely beneficial to humans, robustly and at scale.

A hypothetical system whose general cognitive performance substantially exceeds the best human minds across essentially all domains. It is distinct from - and potentially follows - AGI.

Today's AI is mostly narrow - specialised systems trained for particular tasks or task families. AGI describes a single system that is general across cognitive domains and capable of fluid transfer between them.

A large model trained with explicit chain-of-thought and reinforcement learning on reasoning traces, so that it produces intermediate steps before its final answer. Recent reasoning models have shown large gains on math, code, and scientific problems.

Yes, mechanically - and that turns out to be a remarkably general capability. To predict tokens accurately at scale, models develop internal representations of grammar, facts, logic, and world structure.

A device that establishes a direct communication pathway between the nervous system and an external computer - from non-invasive EEG headbands to high-density cortical implants.

Learn the fundamentals (this site is a starting point), develop skills that complement rather than compete with AI, support good policy and research on AI safety, and stay open to revising your views as evidence accumulates.

AI safety is the broad research field focused on making current AI systems robust, honest, and controllable. AGI safety - sometimes called safe AGI or alignment research - focuses specifically on the harder problem of keeping highly capable, general systems aligned with human values as they begin to act autonomously over long horizons. Major dedicated groups include the alignment teams at Anthropic, OpenAI, Google DeepMind, the UK AI Safety Institute, and academic centres such as MIRI, CHAI at Berkeley, and the Center for AI Safety.

Responsible AI and trustworthy AI are practitioner-facing terms for the operational side of safety: fairness, AI transparency, AI evaluation, AI oversight, security, privacy, and accountability. The NIST AI Risk Management Framework (2023) and ISO/IEC 42001 (2023) are the most widely adopted reference standards.

AI governance covers the laws, standards, and institutions that shape how AI systems are built and deployed. AGI governance extends that to general-purpose AI and frontier models specifically. Concrete instruments in force in 2026 include the EU AI Act, the US Executive Order 14110 and its successor frameworks, the Bletchley and Seoul AI Safety Summit commitments, and the voluntary Frontier Model Forum commitments.

AI risk spans near-term harms (bias, misinformation, cyber misuse, labour disruption) and longer-term risks (loss of human oversight, misuse for biological or cyber weapons, concentration of power, and misalignment of highly capable systems). Serious researchers take both ends of the spectrum seriously and the field of AI ethics increasingly treats them as a continuum rather than rival camps.

The intelligence economy is shorthand for an economic regime in which reasoning, expertise, and judgment are increasingly produced by machines as well as people. It connects ideas like cognitive capital, the knowledge economy, AGI economics, and the future of work, and it is the framing used by groups such as the Stanford Digital Economy Lab and the Brookings AI Equity Lab.

Collective intelligence is the problem-solving capacity that emerges when many agents - human, artificial, or mixed - coordinate. With modern AI it becomes a design discipline: how to combine people, models, and tools into hybrid teams that outperform any component alone.

Regulation (EU) 2024/1689 takes a risk-based approach. It bans a short list of practices (e.g. social scoring, certain biometric categorisation), imposes detailed obligations on high-risk uses (such as critical infrastructure, employment, and law enforcement), sets transparency rules for generative AI, and creates a separate tier of obligations for general-purpose AI models - including stronger requirements for models above a compute threshold considered systemic. Provisions phase in between 2025 and 2027.

Frontier AI is the working term for highly capable general-purpose models at the cutting edge of capability - the models that AI Safety Institutes evaluate pre-deployment and that the Frontier Model Forum was set up to coordinate around. The category is defined more by capability and risk profile than by any single technical threshold.

A specific kind of AI evaluation that probes a model for capabilities relevant to large-scale harm - for example uplift on biological-weapon synthesis, advanced cyber operations, autonomous replication, or deceptive behaviour. The UK and US AI Safety Institutes run these jointly on frontier models before public release.

A capability-threshold framework, pioneered by Anthropic in 2023 and now adopted in some form by other major labs, that commits a developer to specific safety measures when a model crosses defined capability levels - and to pausing training or deployment if the corresponding safeguards are not yet in place.

Mechanisms that keep a person meaningfully in control of consequential AI decisions - including human-in-the-loop review, audit logs, the right to explanation, and pre-deployment access for regulators. The EU AI Act, NIST AI RMF, and ISO/IEC 42001 each formalise pieces of this.

The technical effort to understand what neural networks have actually learned - identifying internal features, circuits, and representations so behaviour can be predicted and audited rather than only observed. Mechanistic interpretability is the most active strand and is treated as a core pillar of alignment.

If general-purpose machine reasoning becomes cheap and abundant, the parties that control the largest compute, the best models, and the most data could capture an outsized share of the resulting cognitive output. AI governance and AGI governance frameworks specifically try to keep this from collapsing into a single point of failure - economically, politically, or militarily.

Cognitive capital extends the older idea of human capital to include the stock of machine reasoning a society can deploy. It is the central unit of account in the intelligence economy: a society's wealth is increasingly a function of how much useful cognition - human plus machine - it can mobilise.

Early studies (Brynjolfsson et al. 2023, Noy and Zhang 2023, GitHub Copilot trials) consistently show large productivity gains for knowledge workers using AI, with the largest gains concentrated among less experienced workers. The pattern so far looks more like augmentation than wholesale replacement, but task composition is shifting fast and some occupations will compress.

Open-weight models lower the cost of misuse but also enable independent safety research, scrutiny, and competition. Most serious analyses - including the International AI Safety Report - treat this as a tradeoff that depends on the capability level of the model and the maturity of mitigations, not a one-sided question.

Useful starting points: the International AI Safety Report (annual), the Stanford AI Index (annual), publications from the UK and US AI Safety Institutes, the alignment posts published by Anthropic, OpenAI, and Google DeepMind, and the curated reading list on this site.