The AI Pascal's Wager

Ploum confronts the choice facing open-source projects about whether to accept code generated by large language models. Three stances are sketched - full acceptance, a cautious middle ground exemplified by Debian-style "not yes, not no" guidelines, and outright rejection - and the middle ground is dismissed as unsustainable because generated contributions will inevitably seep in. Practical trade-offs are examined: adopting AI can increase contributions and speed on mundane tasks but risks alienating ethical-minded users and renders development dependent on proprietary, expensive tooling; rejecting AI secures community trust and code comprehension but pushes away contributors who now rely on LLMs, who may also lack the ability to fully review their own patches.

Looking ahead, Ploum warns of systemic risks from widespread "slop": codebases increasingly opaque to humans, latent licensing contamination, and the near-impossibility of later clean-up - compared to asbestos that looks useful until it becomes a crisis. The recommended pragmatic stance is a cautious Pascal’s wager: strongly refuse AI-generated contributions for now. That choice preserves control and can be reversed if LLMs prove reliable and ethical, whereas early acceptance risks long-term, possibly irreversible harm to a project.