Beyond AGI: Measuring AI’s Capacity to Replace Human Labour

Under the most practically relevant definition, no, although looser definitions permit yes. That both answers seem reasonable exposes the problem with AGI as a binary threshold.

Microsoft researchers argued that GPT-4 showed “sparks” of AGI across mathematics, coding, medicine and law [1]. Under a breadth-based definition, stronger systems may qualify. OpenAI instead defines AGI as highly autonomous systems outperforming humans at most economically valuable work [2]—a standard not clearly reached.

Current AI is notably jagged: it can outperform an expert in one task and require supervision like an unreliable intern in another. GPT-5.4 matches or exceeds professionals in 83% of GDPval comparisons across 44 occupations [3], but GDPval tests specified work products, not complete jobs requiring context, responsibility and reliability [4]. METR likewise finds that success declines as tasks require longer sequences of autonomous work [5]. The OECD’s separate indicators for language, social interaction, problem-solving, creativity, vision, manipulation and robotic intelligence further reflect this unevenness [6].

The real crux of the issue: Intelligence is not directly observable. We infer human intelligence and capability from performance across tests and real-world tasks. AGI definitions likewise compare AI with humans across exams, professional outputs, coding, research and tool use. Until we can isolate core human competencies from the activities expressing them, human-level intelligence will remain proxied by task performance. The labour market is our largest structured and socially consequential catalogue of those tasks and capabilities.

Not entirely. “AGI” implies broad competence, autonomy and economic disruption, while “ASI” suggests systems broadly exceeding humanity. Their familiarity remains valuable in public discourse despite imprecision; DeepMind, for example, defines ASI relative to large human organisations [7].

But they are poor technical thresholds for people trying to make AI “go well.” Policy depends on specific capabilities—such as cyberattacks, pathogen design, labour replacement or AI research—which may arrive at different times. A system could transform the economy without universal competence or become dangerous before replacing most workers.

ASI has a different defence: it covers new activities, discoveries beyond the human frontier and hazards humans may be unable to contain, whereas labour-replacement measures are bounded by existing human work. Yet ASI is less precise than AGI. AGI has an approximate reference class—what humans can do—while ASI has only a lower boundary. Calling an eight-second, one-second and 0.0001-second 100-metre sprint “superhuman speed” is accurate but conceals enormous differences. Likewise, a slightly better scientist and a system performing centuries of research in a day imply different economies and safeguards. AGI and ASI can remain scenario terms, but decisions require magnitude, domain and consequence.

For the dominant labour-and-task meaning of AGI, I propose an AI Labour Coverage Index (ALCI): the proportion of current paid human labour that available AI and robotics could realistically perform. ALCI-50 would mean AI could replace 50% of current labour hours—not that it already does so or that 50% of people have lost their jobs.

ALCI matters because it makes explicit the human-task association embedded in AGI. AGI asks whether AI is as capable as humans but leaves unspecified which humans, tasks and share of the distribution. ALCI supplies an employment-weighted continuous measure instead of a disputed line.

The unit should be labour hours rather than job titles because jobs are bundles of tasks. AI might automate a nurse’s paperwork without examining a patient. Counting the occupation as automated or untouched would miss both realities. The International Labour Organization similarly finds that one quarter of global employment has some generative-AI exposure but expects transformation more often than replacement because most occupations retain tasks requiring human input [8].

Physical capability must be included. Labour involves mobility, dexterity, perception, strength, interaction and safe action in unpredictable environments. Work counts as replaceable only when the complete AI-and-robotics system performs it end-to-end at the required quality and reliability without equivalent human supervision. A model explaining plumbing does not replace a plumber; nor does a controlled robot demonstration replace safe work across homes, hospitals, farms or building sites.

ALCI should report three scores:

C75/E50/D12 would mean 75% technically replaceable, 50% economically replaceable and 12% automated. Anthropic’s 2026 analysis finds observed use far below theoretical exposure and little broad employment displacement so far [9]. The scores separate capability, affordability and adoption.

If humans invent new work that AI cannot perform, ALCI could fall without AI becoming less capable; automated jobs also disappear from its denominator. Acemoglu and Restrepo describe this creation of new labour-intensive activities as a “reinstatement effect” that can counterbalance automation’s displacement of labour [10]. Historical comparisons therefore require both a current labour-market score and a fixed base-year task basket. One shows present vulnerability; the other, technological progress.

The continuous index could also define a threshold state: Artificial Labour-Replacement Intelligence (ALRI). For example, ALRI-90/24 would be reached when ALCI-C and ALCI-E remain above 90% on a continually updated labour-task basket for 24 consecutive months. ALCI-D should not be required because regulation, worker preferences or institutional inertia may delay adoption despite replacement being feasible. The duration prevents a single favourable evaluation from becoming a historical announcement.

More importantly, sustained high current-basket ALCI would show that AI had not merely mastered a fixed catalogue of old jobs. Either AI and robotics had saturated nearly the whole human labour space, or they were also catching new occupations and tasks as humans created them. Keeping pace with the moving frontier of human comparative advantage would be meaningful evidence of economically relevant generality. The precise percentage and duration would be contestable, but unlike AGI they would be explicit and revisable.

Labour remains an imperfect proxy: employment incorporates embodiment, credentials, trust and market demand while excluding unpaid care and domestic work. ALCI’s virtue is not that labour exhausts human intelligence, but that the proxy is explicit. It replaces one operational use of AGI, not consciousness, autonomy, dangerous capability or every meaning of ASI.

Other uses need other measures. Autonomous Task Horizon better captures reliable agency, while biological, cyber, persuasion, replication and AI-research profiles better support safety governance. ALCI-100 cannot distinguish a system barely capable of all current work from one operating far beyond humanity. ASI should therefore be accompanied by a Superhuman Capability Profile comparing performance with the best individuals, organisations and, where meaningful, coordinated humanity across quality, speed, cost, autonomy, parallelism and novelty. The relevant safety question is not simply whether a system deserves the title ASI, but whether its capability in a domain exceeds the ability of human institutions to understand, supervise or resist it.

AGI has not become meaningless, but it has become insufficient. Most operational claims about human-level AI already depend on human tasks, so ALCI makes their labour component explicit, gradual and embodied; ALRI identifies sustained saturation rather than a momentary score. ASI retains a role beyond the existing human task space, but its magnitude and hazards require a profile rather than a single label. We can keep both umbrella terms while replacing their work as technical thresholds with measurements that say what has actually changed.