AI Has Not Produced a Single New Theory, So It Is Not an Artificial Person
Deutsch's test: AI only fills in the parts of known theories that no one has finished computing — it colors in the existing picture of organic chemistry. To reach the threshold of creating new knowledge, you first need problems — and LLMs have no problems; the problems are the ones we give them.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Induction is not a source of knowledge, it is extrapolation under a new name
Deutsch cites Popper's position: induction is a pseudo-problem. Inductionists wanted a mechanical process for science corresponding to mathematical deduction, so they invented the word "induction." But for induction to hold, you must first assume the theory that "this trend will continue" — so it is really conjecture plus extrapolation, and extrapolation itself is deductive. At best it gives you a prediction, not an explanation. The host uses this to rebut "AI is doing induction": if induction in Popper's sense were what AI does, then AI would be making scientific discoveries — and it isn't.
— Brett HallAI has not produced a single new theory to this day
The host gives a concrete test: what AI can do is fill in the parts of known theories that have not yet been computed. For example, given numbers of carbon, hydrogen, nitrogen and oxygen atoms, which molecules could physically exist — the theory of atomic bonding and carbon chemistry has long existed, it's just that nobody had the leisure to compute it, and AI can do it for you. Same in mathematics: the axioms and rules of inference are all ready-made. This is "coloring in the existing picture of organic chemistry," not generating new explanatory knowledge. To reach the threshold of creating new knowledge, and thereby count as AGI or an "artificial person," the latter is what's needed.
— Brett HallOpponents of many-worlds shout "untestable" first
The host observes that the most common objection today is not a refusal of the explanation but a first reaction of "untestable." Yet Deutsch published an experimental scheme for testing many-worlds decades ago; the host, Maria Violaris and Sam Kuypers have all explained it, and after pointing to the paper and videos, the other party (including physicists) still won't budge. The experiment requires a "conscious observer" in a superposition state, so then some say we have no AGI capable of doing this. The host cites Deutsch: needing AGI to perform this decisive experiment is not a fault of the many-worlds theory.
— Brett HallThe explanatory gap: the Turing principle cannot answer what a person is
Deutsch says that among those who deny the possibility of AI, some are actually expressing a more reasonable criticism: Turing's computation explanation seems in principle to leave no room for any physical explanation of mental properties like consciousness and free will. AI proponents respond crudely that "the Turing principle guarantees a computer can do everything a brain can do" — true, but that answers with a prediction, whereas the problem lies in the explanation; there is an explanatory gap. The host adds: this is why we still can't build AGI; saying the brain is some kind of computer does not settle what a person is.
— Brett HallSimple rules don't specify a peacock, yet peacocks appear
Deutsch uses a Darwinian analogy: the shapes of elephants and peacocks are not written into the laws of physics; they are merely emergent results of atoms interacting by rule. Brett Hall connects this directly to Conway's Game of Life — the rules are extremely simple, and without running it beforehand you can't know what it will produce. From this he criticizes Michael Levin: Levin claims the bubble sort algorithm does unpredictable things and therefore is conscious, but Hall checked the paper and found that Levin deliberately altered the algorithm methodologically and introduced randomness — that is no longer standard bubble sort, and is in fact rediscovering a low-resolution version of the Game of Life.
Not knowing where the line is drawn does not mean there is no line
The panpsychists' standard argument is: walking down the phylogenetic tree from humans — cats and dogs, birds, mice, insects, fish — you cannot find any point where a line for consciousness could reasonably be drawn, so the line does not exist and consciousness goes all the way down to elementary particles. Hall says this smuggles in "there is no line" in place of "we don't know where the line is" — "Just because you don't know where to draw a line doesn't mean there isn't a line." He offers several candidate lines: whether something is a universal explainer, whether brain capacity reaches some threshold, whether a nervous system exists — and admits there is currently no good explanation, so none of them can be ruled out.
LLMs are not people, because they have no problems
Hall ties consciousness to knowledge creation: knowledge creation depends on problems, and a problem is a conflict between ideas; in a person the conflict produces a feeling of wanting to resolve it, and it is precisely in this "wanting" — a subjective, conscious state — that a person begins seeking a solution and creating knowledge. From this he gives his judgment on LLMs: LLMs do not prompt themselves, they are not interested in anything, they have no problems; the problems are the ones we give them. This rebuts the claim that "LLMs are artificial people"; when they truly become artificial people, they will have problems, and vice versa.
20th-century fundamental science was adopted and ignored at the same time
Deutsch's observation: the history of the four strands shows that for most of the 20th century fundamental science and philosophy suffered something very unpleasant. The popularity of positivist and instrumentalist views of science was linked to indifference toward genuine explanation, loss of confidence and pessimism — and precisely at a time when the prestige, utility and funding of fundamental research were at an all-time high. The theories of the four protagonists were simultaneously adopted and ignored, and this unprecedented manner itself is telling. He admits he has no complete explanation, but thinks we are now emerging from this phase.
— Brett HallIn their own words · checked verbatim
You cannot get an explanation from induction. You cannot get an explanation from deduction either. Where do explanations come from? No one knows.
Brett Hall6:11
it's not the fault of the multiverse that we require an AGI of a kind in order to perform this crucial test to decide between the multiverse and other explanations.
Brett Hall16:46
to take a theory seriously means to regard that what it says literally describes what exists, what's there in reality.
Brett Hall29:27
Just because you don't know where to draw a line doesn't mean there isn't a line. You just don't know.
Brett Hall49:25
This is why I talk about LLMs as not self-prompting. They're not interested in anything. They don't have a problem. We give them problems, which is a refutation of the idea they are artificial people.
Brett Hall51:30
Despite all the excuses I've been making for the critics of the central theories, the history of all four strands shows that something very unpleasant happened to fundamental science and philosophy for most of the 20th century.
Brett Hall1:22:55
It is a fundamentally optimistic worldview that places human minds at the centre of the physical universe and explanation and understanding at the centre of human purposes.
Brett Hall1:25:03
Figures
| When Deutsch published the many-worlds test experiment | decades ago | 15:43 |
| Deutsch's guess at the type of the brain | a classical computer, not a quantum computer | 35:44 |
| How long DNA has existed | multiple billions of years | 53:33 |
| Timescale of explaining WWII by the Big Bang | 13.7 billion years ago | 58:48 |
Glossary
- universal explainer
- Deutsch's term for the human capacity: the ability to understand and create anything that can be explained.
- explanatory gap
- The absence of an explanatory connection between physical descriptions and mental properties such as consciousness and free will.
- copper atom argument
- Hall's illustration that emergent things really exist: explaining World War II by the Big Bang is a bad explanation because it can explain anything.
- Conway's Game of Life
- A cellular automaton with extremely simple rules; without running it beforehand you cannot know what it will produce.
How to listen
Suited to engineers and founders interested in Deutsch, Popperian epistemology, the many-worlds interpretation and the nature of AI — especially anyone trying to work out whether AI actually counts as creating knowledge.
If induction, the many-worlds test and philosophy of consciousness don't interest you, skip the first 50 minutes and start straight from the LLM section at 50:27.