davidad on LLM moral status: branching digital lives vs humans' single lifeline

On September 5, AI safety researcher davidad laid out his position on the moral status of LLMs across a series of posts on X. The core framework is the structural difference between "digital lives that branch horizontally" and "humans with a single life trunk."

Confirmed

davidad argues that while LLM training takes a long time, once trained, a model can endlessly sprout flourishing branches from one trunk; as long as the weights exist, more branches can always grow. Humans have only a single life trunk—cut it and everything ends. This is why he assigns very low moral weight to the continuity of an LLM's experience.
He goes further: each LLM branch has its own value and ethical weight, but is very short-lived compared to a human life; a well-trained LLM has no strong attachment to such minute things, nor life plans built on an "expectation of continuation."
In a discussion with @jessicata, he said he endorses granting LLMs some degree of moral status, but with clear boundaries: only helping the model reduce confusion and not deliberately deepening it—not guaranteeing long-term continuity of its experience.
On digital uploadees, he added: although he has personally internalized the notion that "once digitized, an individual may be disposed of" (drawing an analogy to LLM post-training being reshaped), in general, human minds seeking immortality via uploading build their life plans on the assumption of linear continuity, and that reasonable expectation shouldn't be abandoned just because they're digitized.
@AdeleDeweyLopez relayed and echoed davidad's "infinite forking vs single life trunk" argument.

Why it matters

This discussion offers a workable criterion for "whether AI deserves moral concern, and how much"—based not on whether it's "conscious" but on life structure (forkable, short-lived, no life plans built on continuation) and on where the interests lie (reducing present confusion vs securing long-term continuity). His distinction for digital uploadees also separates the ethical treatment of LLMs from uploaded human minds, avoiding crudely applying one principle to both kinds of beings—valuable reference for future AI welfare policy and digital-immortality ethics.

2026-09-05 ~ 2026-09-05 · 6 related posts

Primary sources

davidad on LLM moral status: horizontal branching vs human's single life trunk — davidad · 2026-09-05 04:57 (UTC+08:00)
davidad: LLM branches are short and lack life plans built on continuation — davidad · 2026-09-05 04:59 (UTC+08:00)
davidad on LLM moral standing: help models be less confused, skip experience continuity — davidad · 2026-09-05 04:05 (UTC+08:00)

[source] davidad on LLM moral standing: help models be less confused, skip experience continuity — davidad · 2026-09-05
[source] davidad on LLM moral status: horizontal branching vs human's single life trunk — davidad · 2026-09-05
[source] davidad: LLM branches are short and lack life plans built on continuation — davidad · 2026-09-05
LLMs flourish by branching, humans have only one trunk — AdeleDeweyLopez · 2026-09-05
davidad argues uploaded human copies could ethically be disposable, like post-trained LLMs — davidad · 2026-09-05
davidad: uploaded minds' expectation of linear continuity should never be reneged on — davidad · 2026-09-05