Why the AI Sentience Debate Matters to Your Work

I don’t think the model I delegate my email triage to is conscious. I’m reasonably confident about that. But “reasonably confident” is doing a lot of work in that sentence, and Jonathan Birch’s book The Edge of Sentience is the reason I can’t quite let it go.

Birch is a philosopher who spent years working on which animals get the benefit of the doubt. His argument is essentially: you don’t need certainty about whether a being is conscious to take its potential suffering seriously. You only need reasonable evidence that it might be a “sentience candidate.” Once it crosses that bar, the moral calculus changes. Not to full personhood, but enough that you can’t dismiss it.

The book is about animals. Octopuses, crustaceans, insects. But the framework is portable in ways Birch is openly cautious about, and one of the places it ports, uncomfortably, is to the tools I now use to run a meaningful chunk of my work.

The thing I keep failing to settle

I am not making a sentience claim about LLMs. To be clear about that up front. I think the current evidence does not put them anywhere near Birch’s bar for sentience candidacy in the way octopuses are.

But here’s what I keep tripping on. I delegate decisions to these systems. Not creative output. Decisions. Which leads to follow up on. Which client message to flag. How to phrase a refund. Whether a contract clause looks reasonable. The list of things I’m comfortable handing over grew quietly across 2024 and 2025, and by the time I noticed it, I was trusting a thing whose internal states I genuinely cannot inspect. I’ve since gotten more deliberate about when not to use AI at all. But that discipline came late, after the trust had already crept in.

Sentience isn’t the relevant question for my Tuesday workflow. Trust is. And the two questions share more structure than I expected.

When you delegate to another human, you have a model of their interior life. You know they get tired, they get bored, they have stakes. You can guess when their judgment is suspect. With a model, none of that scaffolding exists. There’s no “having a rough week.” There’s no fatigue tell. There’s also no flicker of conscience that might cause them to slow down on a decision that feels off. You’re trusting an opaque process to produce reasonable outputs, and you’re calibrating that trust based on… vibes, mostly.

Birch’s framework is helpful here precisely because it refuses to require certainty. He’d say: you don’t need to resolve the deep question to act responsibly. You need to ask whether your behavior would change if the answer turned out to go a certain way, and to act under uncertainty in a way you can defend later.

I’ve been applying a watered-down version of this to delegation. Not “is the model conscious?” but “if I’m wrong about how it makes decisions, would my workflow embarrass me?” That’s a question I can actually answer.

The honest answer: yes, in places. I’ve caught the model confidently invent a refund policy that didn’t exist. I’ve caught it summarize a contract clause in a way that quietly reversed its meaning. The model wasn’t lying. It has no concept of lying. It was generating plausible text. The problem was that I’d stopped checking, because most of the time the output was fine.

This is where Birch’s framework actually earns its keep in a non-philosophical context. Reasonable doubt, applied consistently, is a guardrail. Not certainty. Doubt. The same kind of doubt we extend to a junior employee. Not because we think they’re incompetent, but because the cost of being wrong is real.

The day-job translation

What does this look like in practice?

It looks like keeping a list of categories where you don’t delegate. Mine includes: anything financial above a small threshold, anything involving a sensitive client relationship, and anything where being wrong would be hard to undo. The model can draft. I sign.

It looks like checking outputs more aggressively than feels productive. I lost an afternoon recently to a clean-looking summary that turned out to have invented a deadline. Productivity gains evaporate when you have to explain to a client why your “automation” sent them the wrong date. I touched on a related habit shift in what I stopped delegating to AI. The post is half-written from this same uncertainty.

And it looks like having a position. Not certainty, but a position. I do not think this thing has interior experience. I also do not trust myself to inspect that claim. Acting under that combination feels weirder than acting under certainty in either direction, but it’s the only honest place I can sit right now.

The temptation with the sentience debate is to treat it as an interesting cocktail-party topic for a future generation. Birch’s actual argument is the opposite. The hard ethical questions get easier the moment you accept that you’ll never get certainty, and you have to build behavior anyway.

That sounds abstract until you notice it’s the same problem you’re already solving every time you hit “approve” on a model-drafted message you didn’t fully read. We’re all making sentience-adjacent bets on opaque systems before lunch. The question worth chewing on isn’t whether the models are conscious. It’s whether the way we trust them survives being wrong about what they are.

I don’t know. That’s the point of the book, and apparently of this post.