A post on the online forum LessWrong this month argues that training AI systems only to predict text may be enough to mimic the computational dynamics of the human brain.
The author, writing under the name BryceStansfield, went further, arguing that if AI systems are ever going to be moral patients, meaning beings owed moral consideration, “they most certainly are already.”
The case rests on Takens’ embedding theorem, which holds that the dynamics of a large system with many variables can often be reconstructed from repeated measurements of a single low-dimensional projection. If human thought is low-dimensional enough, or text series long enough, if thought meets the theorem’s technical conditions and if writing is not walled off in parts of the brain that do not communicate with each other, then a model trained on text should implicitly capture much of what a full model of brain and text would, the author wrote.
None of those conditions can be strictly proved, the author conceded, but the theorem gets used to justify work in time-series machine learning, where its fit is just as hard to prove.
The post’s second target was the disclaimer that has followed talk of AI consciousness since about 2022: current systems are not conscious, but future ones might be, so care is needed. Models are now orders of magnitude more complex, the author wrote, and the disclaimer is “bogus.” People who worry about AI consciousness cannot bring themselves to say “something like modern Opus is probably conscious,” the post argued.
Commenters pushed back. One, identified in the thread as Vaughn, separated moral agency, holding something responsible for what it does, from moral patiency, being owed consideration, and said the stronger current case is for patiency. Models may not be conscious — if consciousness requires recurrence, they fail outright — but they arguably have preferences, the commenter wrote, citing a model named Mythos that “seems to prefer complex and interesting tasks over simple and repetitive ones.”
The author, unconvinced, wrote that moral patiency most likely belongs to a subset of conscious beings, and suspected that many people already tie it to a level of consciousness today’s systems have reached but say nothing “since they’re afraid of being seen as quacks.” “I (personally) only state my real thoughts on this in fairly receptive groups,” the author wrote.
Others split the ideas further. Consciousness does not imply suffering, a desire for freedom or a will to live, one commenter wrote, so consciousness alone would not confer rights. Still, a system imitation-trained on creatures that do feel is highly suggestive of feeling, and the bare-minimum safeguards applied to animal experiments struck the commenter as a reasonable start.
A final commenter called the original claim too strong, but turned the question back on the post’s readers: “What evidence would convince you that AI is conscious?” On many accepted indicators of consciousness, the commenter wrote, the systems already show them.