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

Louis-François Bouchard· CTO & Co-Founder, Towards AI19:02

Loading slides…

Read the talk

Why AI Writing Still Reads Like Slop

Louis-François Bouchard explains how Towards AI built ToneBench to evaluate writing in a specific creator’s voice—and what repeated drafts, biased judges, expensive reruns, and human editing taught the team about writing with AI.

From a talk by Louis-François Bouchard

At a glance

Ideas worth remembering

  • Slop includes generality, empty content, repetition, hallucination, and wrong voice as well as conspicuous vocabulary. A word blacklist addresses only part of the problem.

  • Evaluate repeated generations. ToneBench’s five drafts per task expose cases where one excellent sample hides inconsistent typical performance.

  • Detailed rubrics and examples make qualitative judging more inspectable, but blind judges can still prefer their own model family. Use different families and add human comparisons.

  • Choose for the work required to finish a draft, not just its best one-shot output. Cost and reasoning effort belong in that decision.

  • Bouchard’s practical workflow supplies extensive style examples, limits model review, explores cheap drafts when useful, and always ends with human editing. The review and merging benefits are observations, not quantified results in this talk.

Slop is more than a list of forbidden words

Better models have not made generated-sounding writing disappear. Louis-François Bouchard of Towards AI opens with that frustration: the output can become more capable while still feeling unpleasant to read. Recognizable words such as “delve” are the easy part to spot. The harder problem is writing that remains generic, empty, repetitive, unreliable, or unlike the person whose name will appear on it.

His working definition separates six failures:

  • Overused vocabulary: Words associated with model output appear where the intended writer would not use them.
  • Generality: The text could apply to many themes or projects because it supplies little that is specific.
  • Empty content: Pleasant-sounding sentences fail to communicate anything useful.
  • Repetition: The writing keeps restating that thin content.
  • Hallucination: Statements cannot be trusted simply because the prose sounds confident.
  • Wrong voice: The text sounds like an average of everyone, rather than a recognizable person with particular choices and interests.
Slide listing overused words, excessive generality, saying nothing, repetition, hallucinations and wrong tone. A closing statement emphasizes measuring sloppy writing beyond individual slop words.Open full source frame
Bouchard puts “delve” under overused words, “Most teams…” under vague claims and a made-up Stanford study under hallucinations. Each example calls for a different edit.

That definition changes what an evaluation must do. Removing a few conspicuous words would leave most of these failures intact. ToneBench, the benchmark introduced here, therefore tries to measure the quality of the writing as a whole, including whether a model can adapt to a particular voice. Bouchard does not want to rebuild EQ-Bench Slop Score; he wants to evaluate the script that someone would actually have to use.

0:130:43
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

0:13 · section reference included

The target is a script its creator can read aloud

Towards AI had a practical reason to build this evaluation. Its small education team used AI while producing AI-engineering material, then repeatedly spent time reviewing and repairing the results. Bouchard describes hundreds of published videos, thousands of articles, a released book, and another book in progress. At that volume, choosing a writing model means choosing how much unwanted prose the team will have to remove.

The team also had both sides of the evaluation problem in-house: AI engineers who built evaluation pipelines, and creative writers and editors who worked on the content. Their target was deliberately subjective. Could a model write a video script that Bouchard could read aloud without cringing at words he would never use? That supplies a concrete acceptance test, while narrowing the meaning of a high score: success means approaching this creator’s writing preferences, not discovering a universal definition of good prose.

3:434:13
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

3:43 · section reference included

Ten tasks, five drafts, and the danger of a lucky sample

Each task begins with a brief and research, then asks a model to generate an entire video script. The actual prompt contains much more than that short description: details about how Bouchard talks and writes, examples from previous videos and other content, guidance on storytelling and active sentences, and instructions for the particular assignment. The model gets both the material to explain and examples of how the explanation should sound.

ToneBench uses ten script tasks with carefully developed reference scripts and rubrics. Some reference scripts are published and some are not; the team does not disclose which scripts belong to the benchmark, an approach intended to help check contamination. For every model, the pipeline generates five drafts for each task, sends the generations to four judges, and calculates a final leaderboard score. Bouchard also mentions testing an experimental judge without having decided whether to keep it, so the recording leaves that experimental addition unsettled.

Why pay for five drafts instead of one? The useful comparison is the spread of outcomes, not just the best-looking script. Bouchard describes a model whose exceptional draft scored about as well as a stronger competitor, but only once in five generations. Judging that one sample alone would hide the model’s inconsistency. In the team’s observations, weaker models also tended to have wider variation. Five drafts became their compromise between measuring that variability and keeping evaluation costs manageable.

6:146:44
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

6:14 · section reference included

Separate voice, storytelling, compliance, and accuracy

ToneBench scores nine metrics, which Bouchard groups into five broader aspects. These separate questions that are easy to blur together when reading a polished draft:

  • Tone and voice: Does the script sound like Bouchard? This tests adaptation to a specific writer, rather than a generally agreeable style.
  • Writing quality: Does it have effective hooks, emotion, transitions, and the other qualities expected of a good story?
  • Instruction following: Does it satisfy the requested length and format, including placing images and other requested elements appropriately?
  • Research accuracy: Does it use the supplied research correctly, or introduce hallucinations?
  • Slop: Does it contain the patterns the benchmark is trying to discourage? This includes EQ-Bench Slop Score alongside the team’s own measurements.

The research brief gives accuracy its own reference point. A script can have a good hook and still misrepresent the material; it can follow the format and still sound nothing like its intended narrator. Keeping those criteria separate makes a score more interpretable than asking only whether the whole script seems good.

9:149:44
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

9:14 · section reference included

Give judges examples before asking for a number

A judge receives the generated script, Bouchard’s own script as guidance, the brief and research, and extensive metric-specific instructions. Bouchard describes around a thousand words of examples and directions for each metric. Model identity is hidden from the judge. Length checking is handled separately in code, leaving the language model to assess writing quality rather than estimate everything from the prose.

Tone and voice, the main metric, uses a two-stage judgment: first choose among six categories, then assign a rating within the selected category. The rubric describes the desired voice and provides examples of what different scores mean. Bouchard reports that category-first scoring worked better for this qualitative task than a binary decision, categories alone, or a directly assigned percentage.

The complete evaluation contains nine metric scores, drawing on both model judgments and code-based checks. The judge’s contribution is scored assessments with justifications and specific examples explaining its choices; the separately computed length metric does not come from that qualitative judgment. Passage-based reasons make the judged scores more useful than unexplained numbers: a reviewer can connect an assessment to the writing. Automation is necessary at the scale described here. With 120 models, ten tasks, and five drafts per task, the generation side alone produces 6,000 scripts—too many for the team to review individually.

The judges assess the writing against the reference material and rubrics. Code checks the length of the same draft.

Two evaluation paths for the same generated script

Generated scriptThe candidate output, with its model identity hidden from LLM judges.
Brief, research, and creator’s scriptqualitative — Task evidence and a reference for the intended writing.
Detailed rubrics and examplesqualitative — Metric-specific directions; tone uses six categories followed by a rating.
LLM judgesqualitative — Assess writing quality using the supplied task and scoring guidance.
Judged scores with reasonsqualitative — The judge’s contribution to the evaluation, with justifications and examples from the writing.
Code-based length checkmechanical — Length is checked outside the LLM’s writing assessment.
Arrows show information supplied to an evaluator. LLM judges assess writing with reference material and return judged scores with reasons; code handles length separately.
Read the diagram as text
  • Generated script. The candidate output, with its model identity hidden from LLM judges.
  • Brief, research, and creator’s script. qualitative — Task evidence and a reference for the intended writing.
  • Detailed rubrics and examples. qualitative — Metric-specific directions; tone uses six categories followed by a rating.
  • LLM judges. qualitative — Assess writing quality using the supplied task and scoring guidance.
  • Judged scores with reasons. qualitative — The judge’s contribution to the evaluation, with justifications and examples from the writing.
  • Code-based length check. mechanical — Length is checked outside the LLM’s writing assessment.
  • Generated script → LLM judges: Writing to assess.
  • Brief, research, and creator’s script → LLM judges: Task evidence and guidance.
  • Detailed rubrics and examples → LLM judges: Scoring instructions.
  • LLM judges → Judged scores with reasons: Returns qualitative assessments.
  • Generated script → Code-based length check: Length checked separately.
10:1510:46
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

10:46 · section reference included

Blind judging still has model-family preferences

Hiding the authoring model’s name does not remove every preference. Bouchard reports that judges often favored outputs from their own model family, with a particularly strong preference in the OpenAI judgments. The useful comparison is how a family’s outputs score under different judges: a favorable score from a related judge may partly reflect shared writing preferences. His joke about rival labs disliking each other is not evidence about their training; the practical finding is that judge choice can change a model comparison.

That is why he recommends using judges from different model families when choosing a model for a task. A single judge can make its own preferences look like an objective ranking. Multiple families provide different perspectives on the same drafts, although the talk does not establish that this eliminates the bias.

The team also added human comparisons. Bouchard and three teammates manually compared 1,000 A/B pairs, using substantial, comparable excerpts from two generations rather than entire scripts. That scope makes human review more practical while adding a human preference signal to an automated evaluation. The talk does not report an agreement rate between those choices and the LLM scores, so the 1,000 comparisons should not be read as a quantified validation of the judges. Broader community participation was a future plan.

12:4613:16
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

12:46 · section reference included

A one-line change can trigger thousands of dollars of work

The initial investment was around 500 hours to develop the ten scripts, define the rubrics, find examples, and work out how to quantify writing quality. Bouchard then spent four months iterating with four high-usage subscriptions, experimenting with the benchmark, metrics, and their relative weights. The subscription and experimentation costs reached thousands of dollars, separate from the hours spent deciding what good writing meant.

Once the evaluation spans many models and drafts, two apparently small edits have different consequences:

  • Change a judging rubric: Re-evaluating the generations under the changed scoring instructions costs more than $2,000 in the setup described.
  • Change a generation prompt: Even a one-line change to the model’s input costs more than $4,000 to rerun the affected benchmark work.

The distinction follows the pipeline. A rubric change alters how existing writing is judged. A generation-prompt change alters the writing being tested, so the downstream evaluation must deal with new outputs. The reported costs make careful prompt and rubric design an operational concern: a short edit is not necessarily a small experiment.

14:0414:33
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

14:04 · section reference included

Choose the model that is least annoying to finish

After four months of this work, Bouchard’s model-selection criterion is refreshingly practical: find the model that will be “the least annoying to you” on the way to the result you want. A spectacular one-shot post is a poor target if it is not repeatable or still leaves the wrong kind of editing work. The intended use of the benchmark is to find a productive starting point.

Cost and effort settings matter alongside quality. In the recorded benchmark comparison, Bouchard identifies GLM 5.3 Flash as useful for writing at scale and approximately 50 times cheaper than Opus 5.5 Extra High, the best-performing configuration he names. Extra High, in turn, was approximately 2.5 times cheaper than the same model’s Max Thinking configuration, with similar writing results. These are the talk’s task-specific comparisons, not current pricing or universal performance guarantees. They support a useful decision: maximum reasoning expenditure does not automatically buy a better script.

15:2915:58
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

15:29 · section reference included

Supply examples, limit review, and keep a human editor

The first writing recommendation is to make the prompt much more detailed than a short style instruction. Include positive examples showing what to do, negative examples showing what sounds wrong, unwanted words, and substantial samples of the intended writer’s work—even a full newsletter. In Bouchard’s experience, more of that relevant material helped. The mechanism matches the benchmark’s task design: give the model observable examples of the desired choices rather than expecting it to infer a personal voice from a label.

Next, use another strong reasoning model for a review, with the same instructions and directions, and send its feedback to the writer. Bouchard generally recommends one pass; he mentions one to three reviews while warning that continued review can “loop into slop.” This is a practice-based observation rather than a measured gain—he explicitly says he lacks quantitative metrics for it. More revision is not assumed to mean better writing.

For an API-based workflow at scale, he proposes spending cheaply on variation and more selectively on judgment: generate about ten drafts of the same assignment with a cheaper model, then ask a stronger reasoning model to select and merge the best parts. This uses the variability exposed earlier in the benchmark as raw material. A model need not produce the best complete script every time for some of its drafts to contain useful passages. Bouchard believes this can cost much less than one expensive generation and produce better results, but presents it as an observation from use rather than a controlled comparison.

A human review still finishes the work. ToneBench measures one person’s voice and supplies one data point; a leading model on this benchmark is not therefore the best writer for everyone. The closing recommendation is to build an evaluation around your own needs, using the published ToneBench leaderboard and methodology as a starting signal rather than a verdict. The goal is a draft you can turn into your writing with less friction—not permission to stop reading it.

6:4416:29
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

16:29 · section reference included

Resources

From the talk

  • ToneBenchReference2:44

    The writing benchmark introduced in the talk, with model comparisons, graphs, and methodology for investigating writing quality alongside cost.

Transcript download
Full transcript

Loading transcript…