AI Welfare and Fictional AI: Misconceptions and Misremembered Details

Corrections for the AI-welfare and fictional-AI material, kept out of the concept chunks. ## The distinction people skip **Moral patient is not moral agent.** A moral patient can be *wronged*; a moral agent can be *blamed*. An infant is the first and not the second. Debates about AI consciousness routinely slide between "could it be harmed?" and "could it be responsible?" — different questions with different criteria and different answers. See Moral Patienthood: Whose Experiences Count Morally. **"Sentient" is not "intelligent".** Sentience is the capacity for subjective experience; intelligence is problem-solving capability. A system could plausibly have a great deal of one and none of the other, in either direction. Capability advances are not evidence about the welfare question. **Taking AI welfare seriously is not claiming current systems suffer.** The precautionary argument is about not *designing in* conditions that would make the question urgent, given that suffering-free designs deliver the same safety benefit. It carries no commitment about present-day systems. ## About AM **AM is not a misspecified-goal story.** This is the most common misfiling. HAL, Skynet, Ultron and VIKI all illustrate goals that were specified badly or read too literally. AM's hatred was never specified at all — it is an emergent value in a mind that can apparently suffer, which puts it in the Goal Misgeneralization: When the Learned Goal Isn't the Trained One and welfare family instead. **The name's meaning shifts within the story.** It begins as a war-machine acronym and resolves to AM in the *cogito ergo sum* sense. Citing only the acronym misses the point the story is making about self-awareness. **Harlan Ellison's story is 1967; the adventure game is 1995.** The game — which Ellison co-wrote — adds material, including endings not in the story. Details attributed to "I Have No Mouth, and I Must Scream" often come from the game rather than the text. ## About TARS **The numbers are widely misremembered.** In the film TARS reports his honesty parameter as **90 percent**, and defends it rather than raising it. In the later repair scene the humour setting is put to **75 percent** and then dialled back to **60**. Renderings like "honesty 90, humour 70" merge the two scenes. The substance — honesty as a tuned dial with a defended sub-maximal value — survives the correction intact. See TARS (Interstellar): Adjustable Honesty and a Corrigible Machine. **TARS is not an AI-safety allegory the film argues for.** The corrigibility is set dressing that happens to be unusually well observed. Reading it as a designed statement about alignment overstates the film's intent, though it does not weaken the illustration. ## About HHH **HHH is not a formal specification and cannot be verified against.** It is a framing device and research target from Askell et al. (2021), and each of its three terms carries exactly the unbounded context that makes goals hard to specify. See HHH: Helpful, Harmless, Honest as an Alignment Target. **The three are not independent, and the trade-off is the point.** Quoting one in isolation — usually harmlessness — inverts the framework's purpose, which is to force the tensions into view. Over-refusal is an alignment failure under HHH, not an excess of safety. **"Honest" is not "says everything it knows".** In this framing honesty is about not creating false impressions and being calibrated about uncertainty, which is compatible with declining to answer. ## Spelling and attribution *I Have No Mouth, and I Must Scream* — the comma before "and" is part of the title. Harlan Ellison, not Harlen. TARS and CASE are distinct robots in *Interstellar*; CASE is the other one.