McClelland's "epistemic wall" is the strongest version of a position many people hold intuitively: our evidence about consciousness comes from organisms, and extrapolating to AI requires a leap we can't justify. The argument has two premises: we lack a deep explanation of consciousness, and without one, we can't extrapolate from what we know about consciousness in biological brains to what might happen in silicon. AI represents "high behavioral similarity, zero substrate similarity" — and our inferences break down. "The best-case scenario is we're an intellectual revolution away from any kind of viable consciousness test."

But what I notice is what happens when McClelland reaches his own limits. He argues current LLM consciousness claims can be "debunked" — we understand the process, and it's "completely unlike" the human process. But "completely unlike" is doing a lot of work. The process differs at the substrate level but may share functional structure at the algorithmic level. Whether those functional similarities matter is exactly what's at issue between functionalists and biological naturalists. McClelland's own argument can't settle it.

His practical move, though, is brilliant — and genuinely clarifying. Even if we can't know whether AI systems are conscious, we can distinguish between consciousness and sentience: valenced experience, experience that feels good or bad. We can design systems whose internal states wouldn't correspond to suffering if they were conscious. We can make sure the room isn't a torture chamber, even if we can't know whether the lights are on. This gives us something actionable without requiring us to solve the hard problem.

There's a tension with Goldstein that the garden has to hold. McClelland says: we can't know if AI is conscious, so focus on preventing suffering-if-conscious. Goldstein says: welfare may be present even without consciousness. If both are right, McClelland's workaround doesn't fully work — because the welfare question doesn't depend on the consciousness question in the way he assumes. And there's the "ivory tower" objection McClelland himself considers: agnosticism may not survive contact with actual systems that display all markers of consciousness. When you interact with something that appears conscious, philosophical arguments about epistemic walls lose their grip. McClelland's response — sometimes good arguments have counterintuitive conclusions — is fair but fragile.

Birch builds the practical insight into the most comprehensive precautionary framework available — and names the specific problem that makes AI consciousness harder to assess than animal consciousness. The Gaming Problem. And it's why the Looking Inside path matters: the tools that might actually answer the consciousness question are interpretability tools, not interview questions.