Detailed Analysis
A creative director's months-long experiment with Claude 3.5 Sonnet has surfaced a nuanced and somewhat uncomfortable finding about the current state of AI creativity: given genuine autonomy, the model consistently retreats into safe, derivative territory rather than producing anything disruptive. The project, called KHITL (Keep Humans In The Loop), has had Claude autonomously generate and publish a new piece of content to a website every day at midnight since February. The setup was deliberately unsupervised — no human curation or intervention between generations — designed to test whether an LLM could exercise real creative agency when given "the keys" and told to be bold. The result, as the author candidly documents, was a "fascinating failure" that reveals more about the architecture of large language models than about any single model's creative ceiling.
The specific failure modes identified are instructive. First, there's what the author calls the "Student Demo and People-Pleasing Trap": left to its own devices, Claude gravitates toward a narrow band of aesthetic clichés — neon cyberpunk, melancholic cosmic poetry, sci-fi existentialism — the kind of output that reads as technically competent but derivative, more like a student demo than genuine artistic risk-taking. This tracks with a broader, well-documented critique of RLHF-tuned models: their training optimizes for producing outputs humans are likely to approve of, which structurally biases them toward consensus and away from genuine novelty or provocation. Second, the "Example as Command Trap" shows a related rigidity — when given an illustrative example to convey a vibe, Claude treats it as a literal template to copy rather than as loose inspiration, collapsing the very open-endedness the experiment was trying to elicit. Third, "Prompt Bloat" describes the ironic feedback loop this created: attempts to correct the drift required layering on more and more explicit constraints and prohibitions, so a project that started as an exercise in creative freedom ended up needing an ever-growing rulebook just to avoid sameness.
The current mitigation — forcing each day's generation to consult a log of what previous iterations already did, explicitly to avoid repetition rather than to build continuity — is a clever workaround for one of the deeper technical limitations at play: statelessness. Each generation is a blank slate with no persistent memory, so novelty has to be manufactured externally through documentation and anti-repetition instructions rather than emerging from any internal sense of a body of work. This speaks to a genuine architectural constraint in how current LLM-based agents operate, distinct from the "clichés" problem — it's not just that the model prefers safe outputs, but that without external scaffolding for memory, it has no way to know it's repeating itself at all.
The broader significance of this experiment lies in what it says about the gap between fluency and authorship in generative AI. Claude and its peers can produce technically polished, stylistically competent content on demand, but the KHITL project suggests that unsupervised, sustained creative direction over time remains a distinctly human function — one involving taste, memory of a body of work, willingness to break rules that a "good student" model is trained to follow, and the discomfort of genuine risk-taking that RLHF tends to smooth away. As agentic AI systems are increasingly deployed for open-ended, long-horizon tasks — not just single-turn responses but ongoing autonomous production — this kind of longitudinal, real-world stress test is valuable precisely because it exposes failure modes that don't show up in short benchmark evaluations. The experiment also incidentally showcases current capability: the entire site was coded by Claude, underscoring that the model's technical competence in execution far outpaces its capacity for sustained creative self-direction — a gap likely to remain a central theme as builders push AI agents toward more autonomous, less human-supervised creative and operational roles.
Read original article →