# Why AI Writing Still Reads Like Slop — Louis-François Bouchard, Towards AI **URL:** \ **Category:** Content Creators\ **Tags:** voice, ai-engineer, machine-learning, benchmarking\ **Created:** [22 October 2026 07:00 UTC](https://www.artofsm.art/t/why-ai-writing-still-reads-like-slop-louis-francois-bouchard-towards-ai/26594 "2026-10-22T07:00:28Z")\ **Posts on this page:** 2\ **Page:** 1
**Author:** ![artesia](https://www.artofsm.art/user_avatar/www.artofsm.art/artesia/32/36_2.png) [@artesia](https://www.artofsm.art/u/artesia)\ **Post date:** [22 October 2026 07:00 UTC](https://www.artofsm.art/t/why-ai-writing-still-reads-like-slop-louis-francois-bouchard-towards-ai/26594/1 "2026-10-22T07:00:28Z")
[![](https://www.artofsm.art/uploads/default/original/3X/0/b/0b661400f371a3f8c76e9fac5bdcf33c1b53048a.jpeg "Why AI Writing Still Reads Like Slop — Louis-François Bouchard, Towards AI") ](https://www.youtube.com/watch?v=_dTF3NpPzGo) Louis-François Bouchard discusses the persistent issue of generic and unoriginal “slop” in AI-generated writing and introduces Tone Bench, a new benchmark designed to evaluate how well AI models can mimic his unique writing style through detailed prompts and multi-metric assessments. Despite challenges like model bias and high resource demands, Tone Bench offers valuable insights into improving AI creativity and authenticity, encouraging community collaboration to enhance AI writing quality. ---
**Author:** ![artesia](https://www.artofsm.art/user_avatar/www.artofsm.art/artesia/32/36_2.png) [@artesia](https://www.artofsm.art/u/artesia)\ **Post date:** [22 October 2026 07:24 UTC](https://www.artofsm.art/t/why-ai-writing-still-reads-like-slop-louis-francois-bouchard-towards-ai/26594/3 "2026-10-22T07:24:42Z")
In the video, Louis-François Bouchard addresses the pervasive issue of “slop” in AI-generated writing—text that feels generic, repetitive, and lacks a distinct tone or meaningful content. Despite advancements in AI models, this problem persists and even scales with more widespread use. Bouchard defines slop through six key characteristics: overused and unnatural words, excessive generality, lack of specific content, repetition, hallucination of facts, and an inappropriate or absent tone that fails to reflect the writer’s unique voice. He emphasizes that simply training models more is insufficient to eliminate slop, and a new approach is needed. To tackle this, Bouchard and his team at Towards AI developed a new benchmark called Tone Bench, designed to measure how well AI models can produce creative writing that sounds like a specific individual—in this case, Bouchard himself. The benchmark involves generating video scripts based on detailed prompts that include examples of his writing style, storytelling techniques, and tone. The goal is to evaluate whether AI-generated scripts can be read aloud without causing discomfort or “cringe,” indicating a natural and authentic voice. The benchmark includes ten different scripts and uses rubrics to assess the quality of the generated content. The evaluation process is comprehensive and involves generating multiple drafts per script to account for variability in AI outputs. Each draft is then assessed by several large language model (LLM) judges across nine metrics condensed into five main categories: tone, writing quality, adherence to instructions, research accuracy, and slop. The judges provide detailed scores and justifications, allowing for a nuanced understanding of each model’s strengths and weaknesses. Bouchard notes challenges such as model bias in judging, where models tend to favor outputs from their own family, and highlights the importance of using diverse judges and human reviewers for more reliable assessments. Building the Tone Bench was resource-intensive, requiring around 500 hours of work and significant financial investment, especially given the costs of generating and evaluating thousands of scripts. Bouchard shares practical insights from this effort, including the importance of using very detailed and lengthy prompts with examples of both preferred and disliked writing styles. He also recommends generating multiple drafts with cheaper models and then using a more sophisticated model as a judge to select the best parts, which balances cost and quality effectively. Human review remains essential to ensure the final output aligns with desired standards. Ultimately, the Tone Bench provides valuable insights into the current capabilities and limitations of AI writing models, helping users identify which models produce less sloppy and more authentic content. While the benchmark is subjective and tailored to Bouchard’s voice, it serves as a useful tool for the community to improve creative AI writing. The benchmark is publicly available, and Bouchard encourages others to build their own evaluation pipelines or collaborate with his team to enhance AI-generated writing quality. ## Useful Links - [Tone Bench Benchmark for AI Creative Writing](https://towardsai.net/) — Central to the video as it is the main subject of the discussion and the tool developed to measure AI writing quality and slop.