What consciousness is, whether a machine could have it, and which brain theory is holding up in the lab. The percentages come from a scoring rubric shown on the page, not from me.
The brain answer scores best, and the evidence can't rule out much else. Every brain fact anyone has measured (sleep, anaesthesia, strokes, the 2026 hippocampus result) fits almost every answer below. The soul is the one it counts against directly. What separates the rest is how much new stuff each one needs and whether anything could ever test it. On that scoring, "it's brain activity" comes first and the soul comes last.
The big 2026 story was machines. In July Anthropic reported a workspace-like structure inside Claude, and two of the scientists behind workspace theory wrote that "the machine approximates the functional architecture of conscious processing," while warning of key differences: no body and no lasting memory. In October three heavyweights from rival camps told Nature that today's AI isn't conscious and isn't likely to become so "along current trajectories of development."
Where I land (my opinion, not the scoring)
I've been chewing on this since the 80s. Consciousness is emergent, the same way traffic and love and datacentres emerge out of the Big Bang. It's a feeling neurons produce, and it misdescribes itself. It depends entirely on specific brain states: switch them off and the feeling goes. The idea that the whole universe is made of consciousness doesn't hold up for me.
Where the rubric agrees: the two brain answers come first and second, with 24% and 19%, and the soul comes last.
Where it doesn't: the "misdescribes itself" part I believe scores below plain physicalism, mostly because almost no philosophers hold it. And the universal-mind idea I side-eye still gets 11%. It fits the brain evidence nearly as well as mine does; it loses points because its one proposed test came out mixed, it needs a cosmic mind, and almost nobody in the field holds it.
This year was about AI. Here's which were new and which were reruns.
| When | Who | Said | New? |
|---|---|---|---|
| Jan 22 | Anthropic, Claude's constitution | "Claude's moral status is deeply uncertain." | NEW |
| Feb | Anthropic, Claude Opus 4.6 system card | Asked in a welfare test, the model put its own odds of being conscious at 15 to 20%, and doubted its own estimate | SELF-REPORT |
| May 2 | Richard Dawkins, UnHerd | "If these machines are not conscious, what more could it possibly take to convince you that they are?" | RERUN |
| May 6 | Katlowitz and colleagues, Nature | Under general anaesthesia, the human hippocampus kept processing sounds and words. Unconscious isn't the same as switched off | NEW |
| Jun | Butlin, Long and 18 co-authors, Trends in Cognitive Sciences | A checklist of consciousness signs taken from brain theories. In a popular write-up by co-author Colin Klein and Andrew Barron: "No existing AI system (including ChatGPT) is conscious." | UPDATE |
| Jul 6 | Anthropic (Gurnee, Lindsey and colleagues) | Claude has a small, broadcast set of representations with hallmarks of a global workspace. Anthropic says this doesn't show Claude has experiences | NEW |
| Sep 16 | Mustafa Suleyman, Microsoft AI | "AIs are not conscious." Training Claude on talk of its own moral status makes its self-reports circular | RERUN |
| Oct 19 | Anil Seth, Ned Block and Giulio Tononi, Nature | "Existing AI systems are not conscious," and a precautionary approach to AI welfare "would be misguided" | RIVALS AGREE |
| Oct 19 | Geoffrey Hinton, The Telegraph | "AI has just as much subjective experience as we do." | RERUN |
| Oct 21 | Cairns, yix and Roe, LessWrong | Tiny old GPT-2 passes four of the five workspace tests used on Claude, so those four prove little | REBUTTAL |
| Nov 12 | Anthropic usage policy | Users may not "engage in sustained and needless abusive or cruel behavior toward our models" | NEW |
I don't set these numbers. A rubric does: every answer gets 0 to 5 for fitting the evidence, holding together, needing nothing new and surviving tests, plus half marks for how many experts hold it. A bigger score means "does better on this rubric," not a measured chance of being true, and answers a few points apart are about even. Set 1 and Set 2 ask different questions, so each adds up to 100 on its own. The brain theories get a report card instead, because several can be true at once. How the scoring works. Click an answer to see what's for it and what's against it.
Set 1
Only one of these can be true (or something nobody has thought of yet), so they share 100%.
| Answer | E | C | P | T | X | Total | % | Why |
|---|---|---|---|---|---|---|---|---|
| The feeling is real, and it's brain activity | 4.5 | 3 | 5 | 4 | 5 | 19 | 24% | Fits everything and adds nothing; stuck on why any of it is felt. |
| It's brain activity, and the feeling oversells itself | 4 | 3 | 5 | 4 | 1.5 | 16.75 | 19% | Fits everything and adds nothing; almost nobody accepts it. (Scorers disagreed: evidence: A gave 5, B gave 3) |
| Everything has a little experience in it | 4.5 | 3 | 3.5 | 3 | 2 | 15 | 15% | Fits the evidence, can't be tested, combination problem unsolved. |
| Experience is an extra property with its own laws | 4.5 | 3 | 2.5 | 3 | 2.5 | 14.25 | 14% | Fits the evidence but needs new laws nobody has written down. |
| The universe is one big mind | 3.5 | 2.5 | 3.5 | 2.5 | 1 | 12.5 | 11% | Fits the evidence by reinterpreting it; no agreed test. |
| You have a soul | 2 | 1.5 | 1.5 | 2 | 1.5 | 7.75 | 4% | The one answer the brain evidence counts against directly. |
| Something nobody has thought of | not scored: fixed share | 13% | Fixed, not scored: no rubric can score an idea nobody has had. The two scorers set it at 10% and 15% (the field calls itself pre-paradigm; 29% of philosophers picked "Other" in PhilPapers 2020, which covers undecided, a mix of views, and views not on the list); averaged to 12.5%, shown as 13%. | |||||
Two scorers worked separately from the same research, and each score is the average of the two. Where they were 2 or more points apart, both are shown. How the scoring works
Set 2
This set assumes the feeling is real, as in the top answer of Set 1. If it is, here's what a machine would need to have it. Only one of these can be the requirement, so they share 100%.
| Answer | E | C | P | T | X | Total | % | Why |
|---|---|---|---|---|---|---|---|---|
| The right program is enough | 4.5 | 3 | 5 | 3 | 5 | 18 | 36% | Fits everything, adds nothing, and now has its first AI tests. |
| It has to be alive | 4.5 | 2.5 | 4 | 2.5 | 2.5 | 14.75 | 24% | Fits everything; the strong version can't be tested. |
| It needs the right physical wiring | 3 | 2.5 | 3 | 2.5 | 3 | 12.5 | 17% | Testable, and its key test went against it. |
| It needs quantum effects | 2 | 2 | 1 | 3 | 1 | 8.5 | 8% | Specific and testable, but it needs new physics. |
| Something nobody has thought of | not scored: fixed share | 15% | Fixed, not scored. The two scorers set it at 10% and 20% (known views off the list, such as electromagnetic-field and embodiment theories; the first AI evidence was already undercut once); averaged to 15%. | |||||
Two scorers worked separately from the same research, and each score is the average of the two. Where they were 2 or more points apart, both are shown. How the scoring works
These say which brain processes do the feeling. Several can be true at once (workspace and higher-order theories often travel together), so they get a report card from the lab, not a share of 100.
| Theory | The idea | What the tests did | Report card |
|---|---|---|---|
| Global workspace Dehaene, Changeux | Information becomes conscious when it "ignites" a brain-wide broadcast run from the front of the brain | The 2025 head-to-head test found no ignition when a stimulus ended, which the theory predicted. In 2026 a workspace-like structure turned up in Claude, then GPT-2 passed most of the same tests | Dented, now arguing about chatbots |
| Integrated information Tononi, Koch | The experience is the system's integrated cause-and-effect structure, centred at the back of the brain | The same 2025 test found none of the sustained synchronisation at the back of the brain that it predicted. Called pseudoscience by 124 scholars in 2023 | Dented and disputed |
| Higher-order Rosenthal, Lau, LeDoux | A perception is conscious when the brain represents it to itself: "this is real, and I'm having it" | Early results from a head-to-head against recurrent processing were "weakly supporting the higher-order theory." The brain scans aren't published yet | Slightly ahead, for now |
| Recurrent processing Lamme | A signal becomes conscious when it echoes back and forth inside the sensory areas; broadcasting it only makes it reportable | The other side of that same test | Slightly behind, for now |
| Predictive processing Seth, Friston | What you experience is the brain's best guess, rooted in keeping your body alive | Its first head-to-head (against integrated information) published its plan in August 2026. No results yet | Waiting for its first real test |
| Attention schema CLOSEST TO MINE Graziano | The brain builds a cartoon of its own attention, and the cartoon describes itself as an inner awareness. That's why you're so sure you have one | No head-to-head test yet | Untested head to head |
| Brainstem Devor and Baron, 2026 | The basic feeling of being someone comes from an ancient part of the brainstem, not the cortex | A hypothesis paper (August 2026) built mostly on older and animal evidence. No new experiment | New entrant |
The one real either-or in here: does consciousness need the front of the brain?
Workspace and higher-order theories say yes. Integrated information and recurrent processing say the back of the brain is enough. The brainstem theory says neither. The 2025 test (Cogitate, in Nature) was built to settle exactly this for workspace versus integrated information, and it dented both. Anil Seth's summary: "Neither theory won. And neither theory lost."
It could be true, but it doesn't tell you what consciousness is or who has it. It's a verdict on us, not on consciousness, so it doesn't get a number.
I don't pick these percentages. Each answer gets 0 to 5 on five questions, and the same yardstick is used for every answer:
E, fits the evidence: 5 if it fits everything known with no strain, 0 if solid evidence contradicts it. C, holds together: 5 if it has no open problem of its own, 3 with one serious one. P, needs nothing new: 5 if it adds no new stuff or laws, 3 with one new ingredient. T, testable: 5 if it made risky predictions that came true, 3 if it could be tested but hasn't been, 0 if it can't be tested even in principle. X, expert support: 5 for the most popular view, counted at half weight as a tiebreaker.
The total is squared, so an answer that's clearly stronger pulls ahead instead of everything bunching together. First a fixed share goes to "something nobody has thought of," because no rubric can score an idea nobody has had: 13% in Set 1 and 15% in Set 2, set by the two scorers, with their reasons in the tables. The squares split the rest. An answer that solid evidence contradicts can't get more than 2%. Before any scoring, each answer's best case and its best objection are written out from its own sources.
Surveys, not votes on the truth. Most predate ChatGPT.
| Finding | Who was asked | When |
|---|---|---|
| Philosophers lean physical, but only just: 52% physicalist, 32% not | About 1,700 philosophers (PhilPapers) | 2020 |
| Most philosophers think the hard problem is real (62%) | PhilPapers | 2020 |
| Dualism is the runner-up, not a fringe view: 22% accept or lean towards it. Panpsychism gets 7.5% | PhilPapers | 2020 |
| Scientists' favourite theories: predictive processing (61% call it promising), workspace (58%), higher-order (54%), integrated information (45%). Promising isn't the same as true | 166 consciousness researchers (Francken et al.) | 2018-19 |
| Almost no philosophers thought AI of the time was conscious (3%), and 39% leaned yes for future AI | PhilPapers | 2020 |
| AI researchers put a 1% chance on AI having subjective experience in 2024, and 70% by 2100 (medians) | 582 AI researchers (Dreksler et al.) | 2024 |
| People rate even the most responsive AI as less conscious than the least responsive human, and identical behaviour doesn't close the gap | Nearly 1,100 people (Longin et al., Cognition) | 2026 |
Brain science can test the report cards. It can't settle Set 1.
| If we found… | It would help | Settles Set 1? |
|---|---|---|
| More brain-damage, sleep and anaesthesia results | "It's the brain"; counts against the soul view | No, every Set 1 answer except the soul can live with them |
| A head-to-head experiment one brain theory clearly wins | One report card | No |
| A full map of how the brain models itself | Illusionism, if nothing is left over | Partly |
| A workspace or self-model inside an AI | "The right program is enough" | No, and the camps already read it differently |
| A test for whether anything feels anything | Would settle Set 2 | Nobody has proposed one everyone accepts |
| A mind still working with no brain activity, confirmed | The soul view | Nearly: it would sink both brain answers and favour the soul (Kastrup's idealism could claim it too), and it would reopen the afterlife question. It has never been confirmed |
Go down Set 1 and ask each answer what happens when you die. The two brain answers: the feeling stops when the brain does. If the right program is enough, a perfect copy could in principle run it again, but 54% of philosophers say an upload would be death, not survival (PhilPapers 2020). Illusionism adds there was never an inner essence to survive in the first place. Panpsychism and Chalmers' dualism, as I read them (neither camp has said much about death): the raw ingredients of experience may carry on, but you're one particular arrangement, tied to one working brain, and that comes apart. Kastrup's idealism says death ends "the dissociation, not the end of consciousness": your personal boundary goes, but he says your memories are "released into the broader cognitive context of nature-at-large."
That's the whole argument of my afterlife page, from the other direction. It doesn't need the hard problem solved, either. Every mind anyone has ever observed runs on working tissue, whatever the feeling turns out to be.