Why AI Writing Still Reads Like Slop - Louis-François Bouchard, Towards AI

AI-generated writing quality ('slop') persists even as models improve, so the speaker built Tone-Bench, a new benchmark to measure sloppy vs. good creative w...

By Sean Weldon

Why AI Writing Still Reads Like Slop: A Benchmark-Based Analysis

Abstract

Despite continued improvements in large language model (LLM) capability, the phenomenon of slop - generated text that feels mechanical, generic, and unpleasant to read - has not diminished with scale. This synthesis documents Tone-Bench, a creative-writing benchmark developed by Towards AI to quantify slop and tonal fidelity in long-form generated text. The methodology issues detailed briefs across ten writing tasks, samples five drafts per model per task, and scores outputs using four blinded LLM judges across nine metrics condensed into five aspects, validated against 1,000 manual human comparisons. Findings reveal substantial intra-model variance, pronounced self-preference bias among same-family judges, and large cost-quality asymmetries - a low-cost model achieved comparable output at roughly 1/50th the price of the top-ranked configuration. A multi-draft generation plus judge-merge workflow is proposed as a cost-effective production pattern for practitioners.

1. Introduction

The widespread deployment of LLMs for content production has created a practical measurement gap: standard capability benchmarks - reasoning, coding, factual recall - correlate poorly with whether generated prose is pleasant and useful to read. Social platforms and publishing pipelines are increasingly saturated with text that is technically coherent but qualitatively unsatisfying. This synthesis addresses the specific claim that slop does not decay as models scale; as observed in the source material, "even as models get better... slop doesn't seem to go away. It just keeps on scaling." If slop is not merely a capability deficit resolvable through additional pre-training, it requires direct characterization and measurement.

Slop is operationalized as writing that "feels generated" and is unpleasant to consume, decomposed into six observable components: overused lexical markers (e.g., "delve"), excessive generality, semantic vacuity (text asserting nothing meaningful), repetition, hallucination, and wrong or absent tone relative to a requested voice. Notably, the benchmark's creator acknowledges being "guilty" of producing generic, LinkedIn-style writing, underscoring that slop is not confined to machine output alone but represents a broader degradation of distinctive voice in text production.

The central question motivating this work is whether these six properties can be measured with enough reliability to support model selection for production writing workflows, and what such measurement reveals about current model behavior. This synthesis covers the institutional motivation for benchmark construction, the technical design of Tone-Bench, validation against human judgment, the economics of benchmark development, and resulting recommendations for practitioners selecting models for creative-writing tasks.

2. Background and Related Work

This work sits at the intersection of LLM-as-judge evaluation methodology and creative-writing assessment, a domain that lacks the ground-truth structure available to factual or code benchmarks. Existing slop quantification is represented by the EQ-Bench slop score, which Tone-Bench incorporates as one signal among several custom measures rather than treating as sufficient on its own.

The institutional context is Towards AI, a content organization producing hundreds of videos, thousands of articles, and multiple books, where reviewing AI-assisted drafts for slop constitutes a material and recurring operational cost. The development team brought prior expertise in evaluation pipeline construction for external clients, doctoral-level evaluation research, and in-house editorial staff capable of adjudicating prose quality. Public engagement on social media confirmed community demand for a creative-writing benchmark distinct from existing capability leaderboards, reinforcing the premise that generic quality scores are insufficient - practitioners need a measure of fit to a specific voice under specific instructions, not an abstract quality score.

3. Core Analysis

3.1 Task Construction and Sampling Design

Tone-Bench evaluates models on generation of complete video scripts written to mimic a designated author's voice. Each prompt bundles three components: a content brief, supporting research material, and style exemplars. Ten distinct writing tasks are used, with a subset of source scripts published and another subset withheld specifically to test for training-data contamination - a design choice that addresses the risk that models may have memorized publicly available benchmark content rather than genuinely generalizing style transfer.

For each model and task, five drafts are generated to measure intra-model variability. This repeated sampling revealed that weaker models exhibit markedly wider variance across drafts, suggesting that single-sample evaluations of creative writing quality can be misleading and that variance itself is a meaningful quality signal independent of mean performance.

3.2 Judging Architecture

Evaluation relies on four LLM judges, including one experimental ("Jeff") judge, assessing nine metrics condensed into five aspects: tone/voice, writing quality, instruction-following, research accuracy, and slop (incorporating the EQ-Bench slop score plus custom measures). Each judge operates on approximately 1,000 words of instructions per metric, producing a cumulative judge prompt of roughly 14,000 tokens.

A notable methodological refinement concerns tone/voice scoring: rather than binary or percentage-based assessment, judges first categorize output into one of six tonal categories before assigning a numeric rating. This two-stage categorical-then-numeric approach was found more effective than direct scoring, suggesting that forcing an intermediate classification step reduces judge drift and improves discrimination between subtle tonal variations. Judges are blinded to model identity and instructed to disregard output length when assessing quality, controls intended to reduce confounding from verbosity bias and brand-association effects.

3.3 Judge Bias and Human Validation

A significant finding concerns self-preference bias: judges exhibit a tendency to rate outputs from their own model family more favorably, with OpenAI models showing this effect more strongly than other families when evaluating OpenAI-generated text. This is characterized informally as a rivalry effect between Anthropic and OpenAI models. The practical implication is that benchmark designers should deliberately diversify judge model families to avoid systematic score inflation correlated with provenance rather than quality.

To validate the automated pipeline, the team manually compared 1,000 A/B generation pairs using a team of four human reviewers, checking automated scores against human judgment. This validation step is treated as necessary rather than optional, with plans to scale human evaluation further through community participation - an acknowledgment that LLM-as-judge pipelines require continuous grounding against human preference rather than one-time calibration.

4. Technical Insights

Several implementation-relevant findings emerge from Tone-Bench's development and results:

A notable limitation concerns development cost: initial benchmark construction required approximately 500 hours across four parallel subscriptions over four months. Marginal changes are also costly - altering rubric or judging criteria requires over $2,000 to re-run, while changing a single line of a model's input prompt costs over $4,000 to test, implying that benchmark iteration carries meaningful ongoing financial overhead.

5. Discussion

These findings suggest that slop is better understood as a tonal and structural property of generated text rather than a pure capability gap, since even frontier models fail to achieve one-shot quality despite otherwise strong benchmark performance on reasoning and factual tasks. This reframes the practical question facing engineering teams from "which model is most capable" to "which model minimizes editorial friction at acceptable cost," a distinction with direct implications for production pipeline design.

The self-preference bias finding carries broader significance for the LLM-as-judge literature generally, indicating that single-family judge panels risk systematic distortion regardless of domain. Combined with the demonstrated cost-quality asymmetry between GLM 5.3 Flash and Opus-class models, this suggests that benchmark-driven model selection should weight deployment economics alongside raw quality scores, particularly for high-volume content operations.

A clear limitation acknowledged by the benchmark's creators is subjectivity: Tone-Bench reflects one individual's voice and tonal preferences, and its rankings should not be interpreted as universal "best model" conclusions. This raises an open question for future work - whether tone-matching benchmarks can be generalized across multiple target voices without losing the specificity that makes them useful in the first place.

6. Conclusion

Tone-Bench demonstrates that creative-writing quality can be measured with structured rigor - multi-draft sampling, blinded multi-judge scoring, and human validation - despite the inherent subjectivity of prose evaluation. Its central contribution is empirical evidence that slop persists independent of model scale, that judge bias is a material confound requiring cross-family diversification, and that cost-effective configurations (such as multi-draft generation with cheap models merged by a stronger judge) can rival premium single-pass generation. For practitioners, the actionable takeaway is to treat model selection for creative writing as an optimization over friction and cost rather than a search for one-shot perfection, while maintaining human review as a non-negotiable final step.

Sources

About the Author

Sean Weldon is an AI engineer and systems architect specializing in autonomous systems, agentic workflows, and applied machine learning. He builds production AI systems that automate complex business operations.

LinkedIn | Website | GitHub

Sean Weldon is an AI engineer and the founder of 22Atlas, a studio that builds AI operating systems for businesses, run by Always On Agents. He writes the Journal to keep track of what is working in AI engineering.

More about Sean