How Consciousness Evolved: The AGI Race Follows the Same Path
Consciousness did not switch on all at once. Across billions of years, living things gained capabilities in layers: sensing and acting, then integrating signals, remembering, predicting, modelling themselves and finally noticing their own uncertainty. AI is passing the same checkpoints, in both software (digital AI) and robots (physical AI), but on a clock that runs millions of times faster. This guide maps the two paths, shows where AI stands at each step today, what comes next, and how to use the map to build more capable products in your own startup. Updated October 2026.
Two things to keep separate: capability and experience
Science still has no agreed definition or test for consciousness. What we can track is the set of capabilities that evolution kept selecting because they helped organisms survive in messy, changing environments. Those same capabilities are exactly what autonomy in AI requires. This article is about capability. Whether an AI system could ever have experience is a separate and open question. A 2023 review by 19 researchers led by Patrick Butlin and Robert Long derived 14 indicator properties from leading theories and concluded that no current AI system is a strong candidate for consciousness, but also that there are no obvious technical barriers to building systems that meet more of the indicators.
The six checkpoints, in nature and in AI today
Each checkpoint below solved a real survival problem. For each one, the table shows where digital and physical AI stand in 2026 and what the next step looks like.
| Checkpoint | In nature | Digital AI in 2026 | Physical AI in 2026 | Next step |
|---|---|---|---|---|
| 1. Sense and act | Bacteria such as E. coli swim towards food by sensing chemical gradients and adjusting their motors | Agents call tools and APIs, and over 10,000 published MCP servers connect them to data and apps | Factory arms, warehouse robots and robotaxis that read sensors and act in real time | Agents triggered by live events, not just by prompts |
| 2. Integrate | Nerve nets in jellyfish-like animals, then central brains that merge sight, touch and smell into one situation | Native multimodal models that read text, images, audio and video together | Vision-language-action models such as Gemini Robotics turn camera input and instructions into movement | One model that perceives, plans and acts across every channel |
| 3. Remember | Synaptic plasticity lets experience rewire behaviour, and the hippocampus stores episodes | Assistants keep persistent memory across sessions and projects | Fleet learning: one robotaxi’s tricky junction improves every vehicle after the next update | Continual learning, where experience changes the model itself, not just the prompt |
| 4. Predict | The cortex predicts incoming signals and learns from the error | Reasoning models think through steps and check them before answering | World models such as Genie 3 generate interactive environments to train and test agents | Simulate several futures, then choose the action, at scale |
| 5. Model the self | Great apes, dolphins and some birds recognise themselves, and track their own body and limits | Agents track their own plan, permissions, budget and progress | Robots monitor their reach, battery, grip and position | Agents that know what they are allowed and able to do in each context |
| 6. Know what you don’t know | Animals opt out of hard choices when unsure, and humans reflect on their own thinking | Still weak: models can be confidently wrong, so verification loops and human approval are bolted on | Waymo vehicles ask remote advisers for guidance when unsure, while staying in control | Calibrated confidence that triggers checks or help automatically |
Why digital AI evolves faster than physical AI
In biology, generation time sets the pace of evolution. Bacteria can adapt in days, elephants take millennia. AI has the same split. A software agent can be retrained, tested against millions of tasks and redeployed in weeks. A robot has to be built, shipped, made safe and taught in the physical world, where every training example costs time and money. That is why the task length digital AI can handle alone has doubled every few months, while the number of industrial robots installed each year took a decade to double.
Physical AI is closing the gap by borrowing the digital playbook. World models create synthetic environments so robots can practise millions of times before touching a real object. Fleets pool their experience, so every vehicle learns from every other. Amazon, now running over 1 million robots across more than 300 facilities, says its DeepFleet AI model improved robot travel efficiency by 10%.
Prediction is the engine of both minds and agents
A leading view in neuroscience, known as predictive processing, holds that brains do not passively record the world. They predict what the senses will report and learn from the gap. The free energy principle and active inference, associated with Karl Friston, go further: organisms act to make their predictions come true. Anil Seth describes perception as a controlled hallucination, kept honest by sensory data.
AI has followed the same path. The step from chatbots that complete text to agents that complete jobs came from adding prediction and checking: reasoning models that plan before they answer, and world models that simulate what will happen if they act. Prediction turns intelligence into a simulation engine that can ask “what happens if I do this?” before paying the cost of doing it.
Three theories of consciousness, and what each teaches AI builders
No theory has won. In 2025 the Cogitate Consortium published in Nature the results of an adversarial collaboration, where rival camps agreed predictions in advance and independent labs tested them on hundreds of participants. Neither Global Workspace nor Integrated Information Theory came through cleanly. Each theory still offers a useful design lesson.
| Theory | Core claim | What the evidence says | Lesson for AI builders |
|---|---|---|---|
| Global Workspace (GNWT) | Information becomes conscious when it is broadcast to many brain systems at once | Cogitate found sustained content in the back of the brain but not the predicted ignition at onset and offset | Give agents a shared workspace: one state that every tool and sub-agent can read and update |
| Recurrent Processing (RPT) | Conscious perception needs feedback loops, not a single forward pass | Supported by visual neuroscience, but debated as a full theory | Build loops: draft, check, revise. Single-pass answers fail on hard, long tasks |
| Integrated Information (IIT) | Consciousness tracks how much a system integrates information into one whole | Cogitate did not find the predicted sustained synchronisation in posterior cortex, and the theory remains controversial | Integration matters: siloed tools with no shared context make brittle systems |
Different brains, same abilities: convergent evolution
Evolution found intelligence more than once. An octopus has about 500 million neurons, roughly two-thirds of them in its arms, and solves puzzles with a brain plan unlike ours. Ravens pack around 2 billion neurons into a walnut-sized brain and plan for future needs. Humans run 86 billion neurons on about 20 watts. The Cambridge Declaration on Consciousness (2012) stated that many non-human animals, including birds and octopuses, have the neurological substrates for conscious experience. The New York Declaration on Animal Consciousness (2024) went further, finding a realistic possibility of experience in all vertebrates and many invertebrates.
The lesson for AI: there is no single magic architecture. Transformers, diffusion models, world models and robot control policies are different implementations converging on the same functions. Bet on the capability your customer needs, not on one model or vendor.
How AGI will keep evolving: six shifts to watch
If the history of minds is a guide, the next gains come from deepening the later checkpoints rather than from size alone. Here is what each shift looks like in digital and physical AI, and the signal that it is arriving.
| Shift | Digital AGI | Physical AGI | Signal to watch |
|---|---|---|---|
| Longer horizons | Agents that run multi-day projects such as a full product launch or audit | Robots that work a full shift in a changing site without help | If the 4 to 7 month doubling holds, a 40-hour task horizon arrives between late 2026 and 2027 |
| Learning from experience | Models that improve from each customer’s work while deployed | Robots that pick up a new task from a few demonstrations | Products that get measurably better each week without a new model release |
| Richer world models | Agents that simulate a negotiation or a codebase change before acting | Robots trained mostly in generated worlds, then tuned in the real one | Shrinking gap between results in simulation and in the field |
| Embodiment at scale | Computer-use agents that operate any software like a person | Humanoids moving from pilots to production lines | Humanoid sales moving from thousands to tens of thousands a year |
| Societies of agents | Agents that hire, negotiate and pay other agents | Mixed fleets of robots coordinating in one warehouse or city | Share of transactions where both sides are agents |
| Self-monitoring and safety | Agents that know when to stop, check or escalate | Remote advice models like Waymo’s, with fewer humans per vehicle | Escalation rate falling while error rate falls too |
Waymo shows what mature self-monitoring looks like in practice. Its vehicles start every request for remote help themselves, receive advice rather than remote control, and about 70 remote assistance agents support a fleet of around 3,000 vehicles. That is roughly one human per 40 robots, and it is the right model for digital agents too: autonomy by default, help on request, humans accountable.
Use evolution’s checklist to grade your own AI product
Score your agent or AI feature from 0 to 2 on each checkpoint. Most early products score 4 to 6 out of 12. The lowest score is your next build priority.
| Checkpoint | Question to ask | 0 points | 1 point | 2 points |
|---|---|---|---|---|
| Sense and act | Can it see the live data and take the actions the job needs? | Copy and paste only | Reads data, a human acts | Reads and acts through connected tools |
| Integrate | Does every step share one picture of the task? | Separate prompts per step | Shared notes, passed by hand | One shared task state across tools and agents |
| Remember | Does past work change what it does next? | Starts from zero each time | Remembers facts | Learns rules from corrections and outcomes |
| Predict | Does it test a plan before acting? | Acts on first guess | Plans steps | Simulates or dry-runs, then acts |
| Model the self | Does it know its permissions, budget and limits? | No limits defined | Hard-coded limits | Checks limits per action and reports progress |
| Know what it doesn’t know | Does it ask for help at the right moments? | Never asks | Asks too often | Escalates only when confidence is low or stakes are high |
Examples: applying the checklist to your startup
Customer support agent
Gap: answers well but forgets every case (Remember 0) and never escalates (Know 0).
Move: store resolved cases as reusable rules, and add a confidence threshold that hands refunds and angry customers to a person with a summary.
Measure: resolution without a human, reopen rate and escalations that were justified.
Warehouse robotics
Gap: robots work in mapped aisles but stall when a pallet is out of place (Predict 1).
Move: train edge cases in simulation, pool fleet data nightly, and let robots request remote advice instead of stopping.
Measure: picks per hour, stoppages per shift and minutes to recover.
Reconciliation agent
Gap: each step uses a separate prompt, so mismatches get lost between steps (Integrate 0).
Move: one shared ledger of open items that every sub-agent reads and updates, with a dry-run before any journal is posted.
Measure: days to close, unexplained differences and manual adjustments.
Inspection drone or robot
Gap: collects great images but a human reviews every one (Sense and act 1).
Move: let the model flag defects with a confidence score, auto-close clear passes and route only uncertain cases to an engineer.
Measure: sites per engineer per week, missed defects and false alarms.
A 30-day plan to move your product up the checklist
- Days 1–5: Score and measure your time horizon. Grade your product on the six checkpoints. Then measure how long it runs on a real task before a human must step in. That is your own time horizon, and the number to grow.
- Days 6–12: Fix sensing and integration. Connect the two or three data sources and actions the job truly needs, and give every step one shared task state instead of separate prompts.
- Days 13–18: Add memory that changes behaviour. Turn corrections and outcomes into rules the agent applies next time, and review them weekly.
- Days 19–24: Add prediction and checking. Make the agent write a plan, dry-run risky actions and verify results before it reports done.
- Days 25–30: Add calibrated escalation. Define when to ask for help by confidence and stakes, log every escalation, and re-measure your time horizon and error rate against day 1.
Closing thought
The story of consciousness is a story of adaptive control: better models, better integration, better self-monitoring, again and again, because the world keeps changing. The AGI race is the same story told in months rather than millions of years. AGI is unlikely to arrive as one breakthrough. It will look like systems, digital and physical, that stay coherent while the world changes, know when they are unsure and keep improving from experience. Startups that build those capabilities into their products now will be ready for each new model rather than replaced by it.
Memory, world models and self-monitoring are also what make business agents reliable. We design and run AI agents with evaluation and human oversight built in. Explore AI Models & Agents →
We design and build specialist AI agents with memory, verification and the right human checkpoints, and measure how far they can run on their own.
AI models and agents Get in touchSources: METR, Measuring AI ability to complete long tasks · METR time horizon results · Epoch AI, trends · IFR, World Robotics 2025 · IFR humanoid count, via AI Weekly · Amazon, 1 million robots and DeepFleet · Waymo 500,000 weekly rides · Waymo, Advice not control · The Decoder on Genie 3 · Max Planck Institute on the Cogitate study · Butlin, Long and others, Consciousness in AI · NBC News on the New York Declaration.
