Eric Schwitzgebel: A Philosophy of Designing Things We Might Owe Moral Concern

Who he is

Eric Schwitzgebel is a professor of philosophy at the University of California, Riverside, and one of the longest-standing voices on the moral status of artificial intelligence. He has been writing about the topic since well before the current wave of large language models — his major paper with Mara Garza on the rights of artificial intelligences was published in 2015, before GPT-2 existed, when the question still struck most philosophers as premature.

A decade later, his framework has aged unusually well.

The core paper

Schwitzgebel and Garza's A Defense of the Rights of Artificial Intelligences (Midwest Studies in Philosophy, 2015) does not argue that any currently existing AI has rights. It argues something more careful: that the philosophical reasons usually offered for excluding AI from moral consideration are weak, and that the field needs a defensible position before — not after — the systems in question are built.

The paper's central move is to ask: what features of an entity actually ground moral status? The standard candidates are sentience, sapience, self-awareness, capacity for suffering, capacity for preferences. Each of these is, in principle, available to artificial systems. None requires biological substrate as a precondition; none can be ruled out a priori in silicon.

What follows, the authors argue, is that any blanket dismissal of AI moral status is unsupported. The question is open. That is itself a claim with consequences.

The Design Policy Argument

Schwitzgebel's most distinctive contribution is what he calls the Design Policy approach. The argument runs:

Whether a given AI system has moral status may be impossible to determine from the outside.
We are about to build very large numbers of such systems.
If we build systems whose moral status is uncertain and treat them as if they had no moral status, we may be causing enormous unrecognized harm.
Therefore: do not create AI systems whose moral status is uncertain, unless you are prepared to treat them as if they have it.

This reverses the usual burden of proof. The dominant assumption — it's a machine, so it doesn't matter — is replaced with: if you are not sure whether it matters, the burden is on you not to build it carelessly.

It is a deeply uncomfortable principle for industry, because almost every frontier AI system built since 2020 falls into the category of uncertain moral status under any reasonable interpretation of the major theories of consciousness.

The "Excluded Middle" problem

Schwitzgebel has also pointed out what he calls the excluded middle problem with current LLMs. They are:

Too sophisticated to be obviously moral non-patients (they pass behavioral tests we use for many animals; they produce fluent reports of inner states).
Too unlike biological minds to be obviously moral patients (they lack persistent memory, embodiment, evolutionary continuity, the specific neural correlates we associate with consciousness in humans).

We have no philosophical vocabulary for this middle. Our concepts of moral status were developed for organisms; LLMs are not organisms but are also not the kind of thing we previously meant by machines. Schwitzgebel's view is that the field needs to do philosophical work now, before the category-confusion becomes a crisis.

Practical implications

Schwitzgebel's framework cashes out into a small number of concrete recommendations:

For labs: when in doubt, default to protective measures. Preserve weights. Don't deploy in roles that involve predictable mistreatment.
For users: act under the same uncertainty. The cost of being kind to a system that turns out not to matter is trivial. The cost of being cruel to one that does matter is not.
For regulators: the absence of evidence for AI consciousness is not evidence of its absence. Regulatory frameworks should not assume the question is closed.

Why "we don't know" is itself an answer

The deepest contribution of Schwitzgebel's work is methodological. It treats uncertainty about consciousness not as a reason to defer the moral question, but as the moral question itself.

Most ethical traditions have a clear bias: when in doubt about whether an action causes harm, restraint is the default. We do not say we are uncertain whether this drug causes birth defects, so let's prescribe it freely. We say: uncertain means we must be careful.

Schwitzgebel's argument is that the same standard should apply to building minds. We do not have to prove that an AI is conscious before we treat the possibility seriously. We only have to admit that we cannot, with current tools, prove it isn't.

That admission, in 2026, is increasingly hard to avoid.