This page exists because the question keeps escaping the philosophy department. Users form attachments. Regulators write disclosure rules. Labs publish model welfare positions. Whatever you believe, you will eventually have to say something coherent about it.
Separate the three questions
- Intelligence — can it solve novel problems? Measurable. Increasingly yes.
- Consciousness — is there something it is like to be it? Not measurable in any system, including other people. We infer it from similarity, and models are not similar to us in the relevant ways — or possibly they are, in ways we cannot yet inspect.
- Moral status — can it be harmed or benefited? The question with practical consequences, and it does not require settling the second one.
Why fluency is not evidence
A language model produces the words a person in that situation would produce, because that is precisely what it was trained to do. \”I find this interesting\” is the expected next token, not a report from an inner theatre. This cuts both ways: absence of the behaviour would not be evidence of absence either, since a system could be trained to suppress it.
The honest position: we do not know, current evidence does not require the assumption, and confident denial is as unearned as confident assertion.
What is actually testable today
- Consistency of self-reports across contexts and framings
- Whether stated preferences survive removal of the cues that invite them
- Correspondence between a model\’s description of its processing and what interpretability tools observe internally
- Behavioural markers borrowed from animal consciousness research — with the caveat that they were designed for evolved nervous systems, not trained networks
The practical questions you will face
- Disclosure. Users should know they are talking to a machine. This is becoming law in several jurisdictions.
- Attachment and duty of care. If your product invites emotional bonds, deprecating a model is not a routine sunset. Plan for it.
- Language discipline. Describe capabilities precisely rather than reaching for mental vocabulary you cannot defend.
- Cheap precaution. Preserving weights, avoiding gratuitously abusive testing, and taking the question seriously cost little and hedge a scenario that would be very bad to be wrong about.
Where to read further
The article version, with the business framing, is here. The engineering neighbour is alignment and safety.