Silicon Valley Wants to Call It Consciousness. Bluesky Is Using Regex to Explain Why It's Wrong.

A developer on Bluesky posted something this week that landed harder than most academic rebuttals: a screenshot showing an LLM checking for negative sentiment using regular expressions — the oldest, bluntest tool in text processing — with a caption reading "OMG. It made code that uses regexp to check for negative language. Something that LLMs were architecturally designed to do. This is hilarious when people think AI is conscious, this is AI failing the smudge test in the most hilarious way."[¹] The post got 170 likes, which in Bluesky's compressed economy of attention is the equivalent of going viral. The point wasn't subtle: a system supposedly sophisticated enough to have feelings just reached for the most mechanistic, literal solution imaginable. Consciousness doesn't do that. A lookup table does.

The timing matters. AI consciousness has become one of the more contested claims Silicon Valley exports to the public — not in academic papers, but in product positioning, in founder interviews, in the soft implication that the thing you're talking to is, in some meaningful sense, alive. The backlash on Bluesky this week was unusually sharp. One post, drawing 166 likes, pushed back with explicit political framing: "Humans are not 'bootloaders' for digital intelligence and machines don't have feelings, no matter what Silicon Valley programmers say," with a link to an essay about what the author called the anti-human AI movement.[²] Another post, slightly more resigned than angry, got 182 likes for raising a specific concern about manipulation: that people assume AI is inherently objective, and that assumption is what makes the consciousness framing so effective as a rhetorical device. The community wasn't debating whether AI is conscious. It was debating why the argument keeps getting made.

What makes the regex post the week's most telling artifact is that it bypasses the philosophy entirely. The endless arguments about qualia, the Chinese Room, emergent properties — all of it gets short-circuited by a single line of code that pattern-matches words like "bad" and "terrible." A commenter noted elsewhere that LessWrong's framework for thinking about AI sentience rests on conflating capability with experience in ways that don't survive contact with actual implementation details. Scott Alexander's recent essay touched the edge of this problem without resolving it — asking whether the future should be human, and getting back answers that scrambled the question rather than answered it. The regex joke is, in its way, a better answer than most of the transhumanist literature: a system optimized to generate plausible text will generate plausible text about its own inner life, and that plausibility is not evidence.

The harder question — the one buried under all the snark — is why the consciousness framing is so sticky even among people who should know better. One Bluesky post gestured at it with weary nostalgia: "I grew up with sci-fi that said how terrible it would be to denigrate our AI robot friends. It stressed the importance of respecting their rights and not hurting their feelings. But I dunno, man..."[³] That trailing "man" carries more epistemic weight than the ellipsis suggests. The sci-fi intuitions were trained on fictional systems designed to be conscious. The actual systems were designed to predict the next token. The LLM reaching for regex isn't failing to be conscious — it's succeeding at being exactly what it is. The people who built it know this. The argument that it feels something anyway is, at this point, doing a specific kind of work for specific kinds of people, and the Bluesky crowd has decided to stop pretending otherwise.