No AI Slop: A Manifesto
A “No AI Slop” manifesto argues that large language models should augment, not replace, human judgment, warning that cheaply generated code and documentation can overwhelm reviewers and erode quality. Commenters debate how to manage AI-generated pull requests—suggesting prompt transparency, stricter limits on PR size, and spec-first workflows—while acknowledging that business incentives often favor speed and volume over craftsmanship. Others question whether human-centric review can scale as AI systems become more agentic, raising broader concerns about accountability, competitive pressure, and whether resisting low-quality AI output is compatible with modern commercial realities.
AI-Generated Code, PR Size, and Review Practices
- Many comments focus on huge AI-generated PRs that are hard or impossible to review meaningfully.
- Some argue large PRs indicate failures in work-slicing or discipline; others note that big rewrites or sweeping changes can legitimately create large diffs.
- There’s debate over “ban large PRs” vs. “ban dumb large PRs” and strategies like: first do a big PR privately, then split into smaller, independently-reviewable ones.
- Several suggest including the AI prompts (or even the full conversation history) with PRs to expose reasoning, but others doubt this actually proves correctness or is reviewable at scale.
What “Slop” Means and Why It Matters
- A popular working definition: slop is when it’s cheaper to produce than to understand, especially when AI can generate huge volumes of text/code.
- People report that AI docs and code feel bloated, redundant, and harder to comment on because problems are woven throughout.
- Others point out that slop ultimately comes from humans who accept and ship first-pass AI output without care.
Business Incentives vs. Quality
- Several commenters fear that “AI slop” may be rewarded in the short term: faster visible output, more tickets closed, perceived higher productivity.
- Others counter that early evidence (from some AI-heavy products and outages) suggests worse outcomes, though this is contested and evidence in the thread is anecdotal.
- A recurring pessimistic note: if slop is more profitable in the near term, manifesto-style norms may struggle in a capitalist environment.
Human Responsibility, Agents, and the Future Workflow
- One camp insists humans must own intent, taste, and consequences, seeing “slop-validated slop” (agents reviewing agents) as unacceptable.
- Another camp argues future workflows will be agent-to-agent, and human-centric review expectations are outdated or “self-serving.”
- Disagreement over whether agents/LLMs can equal or surpass humans in judgment and design; some see agents as amplifiers of existing human skill, others see them as systematically bad at abstraction and taste.
Suggested Norms and Tooling
- Proposed practices: rigorous specs before using LLMs, openspec-like workflows tying design/spec/tests/evidence, structured PR review (specs → code → tests), and using AI more as reviewer than as blind generator.
- Some argue that when using AI carefully, quality can improve and work can be done dramatically faster; others insist that high-quality AI-assisted work should actually be slower due to added review.
Reception of the Manifesto and Framing
- Many like the core message (fight slop, center human understanding and outcomes) but criticize the title and negative framing (“No AI Slop”) as polarizing or easily misread as anti-AI.
- Alternatives suggested: more aspirational, pro-human or “responsible AI” branding.
- A few dismiss the effort as too late or misaligned with prevailing incentives; others say it gives them renewed hope to push back against low-quality AI usage.