Omar El Herraoui

Large language models now produce a large share of the text people read, and much of it has recognizable flaws that readers increasingly call "AI slop": unnecessary verbosity, clichéd phrasing, generic claims, an off-key tone and weak structure. Studies of professional editors show that frontier models all make these mistakes at similar rates, which suggests the cause is how LLMs write in general rather than any one model. The volume of AI-assisted text is growing fast, but we still lack reliable ways to measure slop or to stop models from producing it. This PhD takes on both problems. The first line of work characterizes AI slop: which features readers and experts actually react to, how existing taxonomies and edit categories relate to each other, and how far automatic detectors can be trusted compared with human judgment. The second line develops methods to reduce slop at inference time. It converts human edits of AI-generated text into interpretable writing skills (short natural-language rules with fix guidance and examples) that a writing agent retrieves while it writes, with no fine-tuning. Because these skills are readable, auditable and independent of any one model, they suit closed-weight frontier models, where retraining is not an option. The third line looks at why slop happens. It tests whether skills correct the documented tendency of LLMs to add and elaborate text instead of deleting and restructuring it the way expert editors do. The fourth line builds evaluation for this subjective setting: standardized writing benchmarks paired with span-level slop measurement, checked against human raters. Later work will refine skills iteratively under subjective feedback, where no verifiable reward exists, and will adapt them to specific domains, genres and users. The expected outcomes are an empirical account of what makes AI text read as slop, open slop-measurement tools, and reusable skill libraries that measurably reduce slop across models.