On Google DeepMind’s levels, today’s generative AI is at most “emerging AGI”, and embodied AI sits only between Levels 1 and 2 of a separate embodied-AGI scale. On Google DeepMind’s 2023 levels framework, large language models count as “emerging AGI”, comparable to unskilled humans. Whether a body is needed for general intelligence at all is still argued: the DeepMind authors say physical ability adds generality but is not a prerequisite, while embodied cognition theory holds that human embodiment is necessary to ground meaning. The measured gap is still wide: on ARC-AGI-3, humans solve every environment and frontier models score below 1%. In practice, the healthcare AI that works on human body data today is narrow, bounded and regulated.
AGI vs generative AI: how does AGI differ from today’s models?
AGI is not one agreed threshold but a family of related concepts with different implications. One operational version comes from Morris and colleagues at Google DeepMind, who define AGI by capabilities rather than mechanisms and score generality and performance separately. Their five performance levels run from emerging through competent, expert and virtuoso to superhuman; a competent AGI outperforms 50% of skilled adults across a wide range of non-physical tasks.
As of September 2023, the authors placed frontier language models such as ChatGPT, Bard and Llama 2 at “Competent” for some tasks (short essays, simple coding) but “Emerging” for most, including mathematics and factuality, which makes them Level 1 General AI overall. No public system had reached Competent AGI at the time of writing.
A 2023 GPT-4 evaluation made the case for generative AI as AGI, calling the model “an early (yet still incomplete) version” of an AGI system on the strength of its breadth across mathematics, coding, vision, medicine and law. The same paper is candid about what is missing: GPT-4 generates only the next word, with no mechanism to revise earlier output, so it produces arguments “linearly”. The authors raise the possible need for a paradigm beyond next-word prediction.
So the useful distinction is this: generative AI is broad but unevenly reliable, and AGI, on DeepMind’s definition, requires breadth plus reliable performance plus the ability to learn new skills. A psychometric framework based on Cattell-Horn-Carroll theory scored GPT-4 at 27% and GPT-5 at 57%, with critical deficits in areas such as long-term memory. For the wider split between systems built for one job and systems meant for any, our piece on narrow AI vs general AI covers the vocabulary, and what engineering teams actually ship as applied AI covers the deployment side.
Does artificial general intelligence need a physical body?
Whether AGI requires robotic embodiment is still a matter of debate.
Embodied cognition, as the Stanford Encyclopedia of Philosophy summarises it, rejects or reformulates the view that the brain is a computer and the seat of cognition, holding instead that the body and its interactions with the environment constitute or contribute to cognition. In its strongest form, the “constitution” thesis, the body is literally part of the cognitive system, which then consists of more than the nervous system and sense organs. Applied to AI, the implication is direct: if the theory is right, a fully functional brain model would need more than neurons, for example a robotic body. Rodney Brooks argued in 1991 that for simple intelligence, explicit world models get in the way and it is better “to use the world as its own model”.
The other side is the DeepMind framing. Most AGI definitions focus on non-physical tasks, and the levels paper suggests the ability to do physical tasks increases generality but should not be treated as a necessary prerequisite. The authors leave a door open: embodiment may be needed to build world knowledge for some cognitive tasks, in which case it could be critical to some paths to AGI.
Embodied cognition has its own weak spot. Some tasks are “representation-hungry”, such as thinking about unicorns or counterfactuals, and the world cannot serve as its own model for things it does not contain. Critics also doubt that Brooks-style architectures scale up to the advanced behaviour cognitive scientists study.
Our reading: the body is probably not required to reach high scores on cognitive benchmarks, but it is very likely required for an AI to act competently in the physical world, which is what most people picture when they hear “human-like”. Those are different targets. The question of whether a general system would design its own embodiment is taken further in would AGI make its own body.
Why walking and grasping are still harder for AI than reasoning
Moravec’s paradox is the observation that tasks biological organisms find effortless, such as walking and recognising faces, are very hard for AI, while abstract computation and logic come comparatively easily. The paradox arises from evolution: sensorimotor skills evolved over millions of years, whereas abstract reasoning emerged recently. Brooks made the same point in 1991, arguing that moving through a dynamic environment and sensing it well enough to stay alive is where evolution concentrated its time, and that this part of intelligence “is much harder”.
The paradox remains largely unresolved: deep learning has narrowed the gap in constrained settings, but robust, adaptive sensorimotor intelligence comparable to even simple animals remains elusive. The authors of the embodied Turing test proposal write that AI systems easily defeat any human at chess and Go, yet no effective system can walk to the shelf, set up the pieces and play, nor match a four-year-old’s sensorimotor capabilities. Their proposed fix is an embodied Turing test aimed at the capabilities shared with all animals, inherited from over 500 million years of evolution. The DeepMind authors note the same asymmetry from the other direction: physical capabilities seem to be lagging non-physical ones.
What LLM-driven robots still lack
Bolting a language model onto a robot produces what one 2025 critique calls “weakly embodied AI”, which inherits some of the problems of symbolic, rule-based AI. Three gaps stand out in that analysis.
- Symbol grounding. An LLM can talk about riding a bicycle, but grounding drawn only from text gives little ability to do it when embodied in a robot.
- Real-time control. Closing the control loop at 3 to 10 Hz is not fast enough for real robotic tasks.
- Touch. There is “no touch on the Internet”, tactile data depends on the specific hand and sensor, and haptic data is gathered actively through exploratory movement.
The same paper points out that cross-embodiment models work because they reduce every robot to the coordinates of its gripper or centre of mass, which is the shallowness of the embodiment, not its depth. A separate neuroscience-inspired review finds that current agents, including vision-language-action and navigation models, have only reached the first of four cognition stages, memory-based pattern recognition.
ELLMER, from the University of Edinburgh, is one measured example of an LLM-driven robot. It combined an LLM, retrieval over a curated knowledge base, and vision plus force feedback on a seven-degree-of-freedom Kinova arm to make coffee in an uncertain kitchen. It interpreted a high-level spoken command and completed a long-horizon task while handling disturbances. The limits are in the same paper: identification of a white cup was about 90% at 20 to 30% occlusion but fell to about 20% at 80 to 90% occlusion. The system adjusted to real-time changes but struggled with proactive adaptation such as switching task midway without prior programming. Hard-coded maximum velocity and force limits and workspace boundaries sat underneath the LLM, and generated code ran only against predefined functions.
How far are current AI systems from human-like general intelligence?
Far, on the measures below that test generality rather than knowledge. The numbers come from different methods and should not be averaged.
| Measure | What it tests | Current result |
|---|---|---|
| ARC-AGI-3 (ARC Prize, March 2026) | Learning unfamiliar interactive environments with no instructions | Humans solve 100%; frontier models score below 1% |
| ARC-AGI-3 official leaderboard at release | Same, through a general API with no task-specific harness | 0.10% to 0.50% across the four frontier models listed |
| DeepMind Levels of AGI (2023) | Breadth and reliability on non-physical tasks | Frontier LLMs at Level 1, Emerging AGI |
| Embodied AGI levels (2025) | Omnimodal sensing, human-like cognition, real-time response, generalisation | Embodied AI between L1 and L2 of five |
| Neural Brain cognition stages (2025) | Depth of cognition in the perception-action loop | Agents at the first, memory-based stage of four |
ARC-AGI-3 is worth singling out because it defines AGI as the ability to acquire any skill a human can, as efficiently as a human can. It scores action efficiency against a human baseline and squares the ratio, so needing ten times as many actions as a person earns 1% credit for a level. That definition matches Chollet’s 2019 argument that skill on a fixed task can be “bought” with priors or training data, so intelligence should be measured as skill-acquisition efficiency.
Benchmarks carry their own caveat. A 2026 forecasting review notes that leading benchmarks often saturate within two years and that training-data contamination has been documented across widely used suites.
Artificial general intelligence in 2018 and the forecasts since
In 2018, OpenAI’s charter framed AGI around systems doing most economically valuable work at or above human level. The charter also stated that AI would have broad societal impact before AGI arrives.
The expert survey of that era, by Grace and colleagues and published in JAIR in 2018, polled researchers who had published at NIPS and ICML in 2015; 352 responded. It defined high-level machine intelligence (HLMI) as unaided machines doing every task better and more cheaply than human workers. The aggregate forecast gave a 50% chance of HLMI within 45 years and a 10% chance within 9 years. Full automation of labour came much later, at 50% in 122 years. Asian respondents expected HLMI in 30 years, North Americans in 74. Earlier, four polls in 2012 and 2013 had put the median 50%-confidence date at 2040 to 2050.
Over the past five years, expert forecasts have moved toward the near term. A 2023 survey of AI researchers put the median for a 50% chance of HLMI at 2047, a 13-year shift from 2060 in a similar survey a year earlier. Prediction markets and compute-centric models now place central estimates in the 2030s, and some forecasters, including leaders at frontier labs, put AGI within one to four years. The same review warns that forecasting infrastructure is immature, with no resolved forecasts to calibrate against.
Treat any single date with suspicion. The review calls whether scaling current architectures is enough the central disagreement in AGI forecasting. The deeper treatment of where general AI may be heading lives in our guide to the future of artificial general intelligence.
Would copying the nervous, endocrine or immune system get us closer?
For the nervous system, researchers have proposed concrete architectures. One 2025 proposal for embodied agents borrows its architecture from neuroscience, combining multimodal active sensing, a closed perception-cognition-action loop, neuroplasticity-driven memory and neuromorphic hardware. Efficiency is part of the motivation: the human brain runs on about 20 W through sparse, event-driven processing. The NeuroAI paper by Zador and colleagues adds that biological neurons rarely fire above 100 spikes per second, with typical cortical rates possibly below one. Even so, the Neural Brain authors say that matching the brain’s flexibility and generalisation remains an open research problem. Whole brain emulation, scanning and faithfully simulating a biological brain, faces the embodied-cognition objection that neurons alone may not be enough.
One argument brings in the wider body through homeostasis. Its authors hold that “core affect” tracks endocrine, visceral and muscular changes in the body, signalling what in the environment matters to the organism. Because the authors treat emotion and cognition as inseparable, they conclude that cognition depends on the whole living body engaged with its environment. That gives a principled reason to think bodily regulation matters for intelligence, but it is an argument about human cognition, not a recipe for building AI.
What AI can do with human body data without AGI
Narrow, well-specified tasks, each authorised as its own medical device. The US FDA had authorised over 1,600 AI-enabled medical devices for marketing as of September 2026. Its examples include skin-cancer imaging diagnostics, deep-learning image sharpening, diabetic retinopathy detection from retinal images, heart-attack risk estimation from a smart sensor, and automated insulin dosing from continuous glucose monitor readings. The FDA does not regulate AI as such; it regulates medical devices, including those that use AI. Its drug centre reports experience with over 500 submissions containing AI components between 2016 and 2023, and it published draft guidance in 2025 on using AI to support regulatory decisions for drugs and biologics.
On molecular data, AlphaFold 3 predicts the joint structure of complexes of proteins, nucleic acids, small molecules and ions, with far greater accuracy on protein-ligand interactions than leading docking tools.
The pattern across these examples: each system does one well-specified job on one kind of body data. Where those tools already ship in imaging and drug discovery, and where they remain experimental, is covered in our article on generative AI in medical imaging. The computer vision layer behind cleared tools is covered in AI-enabled medical devices, and language-model use across the sector in large language models in biotech and life sciences.
A feasibility check before scoping body-data AI
- Is the task bounded: one input type, one output, a known reference standard?
- Would a skilled human doing the same task repeatedly agree with themselves?
- Does success depend on physical manipulation in unstructured settings? If so, expect the occlusion and adaptation limits seen in ELLMER.
- Does the product count as a device under the FD&C Act, meaning it is intended for diagnosis, treatment or prevention of disease or to affect the structure or function of the body?
- Is the expected outcome possible at today’s capability levels, or does it quietly assume the system will be smarter than its users?
Scoping to the capability that exists
The last question on that list is the one we establish first in any generative AI engagement. “We want an AI that is smarter than us” and “we want AI to do this repeatable task instead of us” are different propositions, and the second is far more likely to return measurable value. TechnoLynx scopes work to problems that pay off at realistic intelligence levels and does not commit to outcomes that would need super-human capability.
AGI may come sooner than the researchers in Grace and colleagues’ survey expected, or it may stall on the problems Moravec described. What you can decide now is whether your project needs general intelligence or one reliable narrow capability.
