Claude finds religion while social science becomes a paper factory

Anthropic is literally taking Claude to church. In a move that sounds like the setup for a high-concept sci-fi novel, the AI lab has been hosting dialogues with clergy, philosophers, and "wisdom traditions" to figure out the "moral formation" of its models.

It’s a fascinating attempt to move beyond just training on the chaotic mess of the open web. Anthropic wants Claude to have a "character"—one that doesn't just fold under pressure or turn into a sycophant the moment a user gets pushy. They’ve even experimented with giving Claude a literal "internal conscience"—a mid-task tool that reminds the AI of its ethical commitments.

It’s all very noble on paper, but while the philosophers are debating virtue, the actual users are busy turning the academic world into a high-speed assembly line. According to a massive new survey of 1,260 social scientists, we’re seeing the birth of the "AI paper factory."

The stats are pretty wild: 81% of researchers have messed around with chatbots, but the real power players—the 20% using autonomous coding agents like Claude Code—are operating on a different level. These early adopters are starting more projects and posting up to 75% more working papers than their peers.

But here is where the "pragmatic realism" kicks in. This isn't some democratic revolution in science. The survey found that if you’re a junior researcher at a top-25 university with a typically male name, you’re significantly more likely to be riding the agentic wave.

The gender gap is particularly depressing. Twice as many researchers with typically male names are using coding agents compared to those with female names. High-status private universities are also seeing 40% higher adoption rates. We aren’t just automating research; we’re automating the existing hierarchy of the Ivy League.

And for what? While the "early pipeline" is flooded with new working papers and grant proposals, the number of actual journal submissions hasn't budged yet. We are effectively creating a massive bottleneck of "academic AI slop" that the peer-review system is nowhere near ready to handle.

Anthropic is right to ask what it means for an AI to be "good," but the reality on the ground is messier. Researchers are optimistic about their own productivity, but they’re deeply skeptical about whether this makes the social sciences any "better."

More papers usually just means more noise and more competition for an already microscopic amount of attention. You can give Claude a moral constitution and an internal monologue, but if the end result is just a faster way to clog the scholarly record with incremental "daring" discovery, we might be solving the wrong problem.

Let's plan for the long game: will a "virtuous" AI matter if it’s mostly used to drown out human researchers who don't have access to the latest $20-a-month "Antigravity" tools? The philosophers might need to stay for a few more sessions.

Sources: Widening the conversation on frontier AI, Coding agents in the social sciences.