How to fix AI writing slop in your prompts

AI writing slop persists because models optimize for fluent, generic prose rather than a specific human voice, and a new benchmark built by Towards AI tries to measure the gap. Tone-Bench, released in October 2026, scores models on ten video-script tasks against the voice of one writer, using four LLM judges plus 1,000 human A/B comparisons.

What does the Tone-Bench benchmark actually measure?

Tone-Bench measures whether a language model can write a full video script in a specific writer's voice, then scores the result against rubrics rather than a hidden truth. Towards AI, the AI engineering education company where the speaker works, published the benchmark in October 2026 with ten script tasks, nine scoring metrics and open methodology.

AI writing slop is the target. The speaker, who has built evaluation pipelines since a master's and PhD, defined the goal narrowly: could he read the model's script aloud without cringing at words he would not use? That framing keeps the benchmark honest about its scope. It measures voice match for one writer, not universal writing quality.

Each model receives a long brief containing research material, instructions on structure, active sentences and storytelling, and many examples of the author's previous work. The model returns one script. The script then goes to judges that never learn which model produced it.

Five drafts per model per task are generated on purpose. The variance matters: a single strong sample can make a weak model look competitive, while the median identifies what a team can expect across many runs.

The benchmark site publishes graphs, comparisons and the full methodology behind the scores.

Why does AI writing slop survive better models?

AI writing slop survives because models are optimized for broadly acceptable prose, and broadly acceptable prose has no specific author in it. The speaker breaks the failure into six parts, and every part traces back to that same pressure toward the average.

1. Overused words that people rarely type themselves, such as "delve".
2. Excessive generality about "most projects" or "most teams".
3. Text that says nothing while sounding agreeable.
4. Repetition of that nothing.
5. Hallucinated facts mixed into fluent sentences.
6. Wrong tone, or no tone, blending every writer into one.

The sixth item does the most damage. A reader can delete a bad word. They cannot easily fix a paragraph that sounds like everyone and no one, because the fix requires knowing what the writer would have said.

The speaker also noted that stronger releases did not remove the problem; the slop scaled with the models. That observation is the transcript's central claim about the industry, and it stays a claim about generated content the speaker saw, not a measured study of every model on the market.

How do LLM judges score writing quality?

LLM judges score writing quality by rating each of the nine metrics with a justification and a quoted example from the script. Tone and voice receive the most attention, and the judge first picks one of six categories, then assigns a rating inside it.

The internal categories matter less than the calibration process. Instead of asking for a bare number, the rubric defines what each score looks like in writing and shows examples, which the speaker says worked better than binary or direct percentage scoring for this qualitative task.

The nine metrics condense into five aspects: tone match, craft (hooks, emotional transitions), instruction following, research accuracy, and slop. The slop component uses the EQ-Bench slop score among other checks. The judge sees roughly a thousand words of rubric around each metric, the script, the author's own script as guidance, and the research brief.

Judge bias is the known weakness. The speaker's data showed OpenAI models strongly preferring their own family, and he described what looked like an Anthropic-versus-OpenAI mutual dislike. The practical advice: if you use judges for your own pipeline, mix model families.

Vendor benchmarks usually claim objectivity. Here the bias is published, which makes the leaderboard a signal rather than a verdict.

Tone-Bench setup at a glance

The configuration balances statistical coverage against cost, and every knob was chosen after experimentation. The table below reflects the setup described at the October 2026 release talk.

Component	Configuration
Script tasks	10 briefs with rubrics
Drafts per model	5 per task
LLM judges	4, plus a fifth trial judge
Judge prompt size	About 14,000 words
Scoring metrics	9, grouped into 5 aspects
Human checks	1,000 A/B pairs by 4 reviewers
Task visibility	Some briefs unpublished, to test contamination

The 1,000 pairwise comparisons address the criticism that only humans can judge creative work. Reviewers compared comparable chunks of two scripts rather than whole drafts, a scalable compromise the speaker wants to expand with community help later.

What running this benchmark costs

Running a benchmark like Tone-Bench costs thousands of dollars per experiment, not per year, because every rubric or prompt edit forces a full regeneration and regrade. The speaker reported roughly 500 hours to build the ten tasks and rubrics before any model ran.

The ongoing figures he gave matter for anyone planning something similar. Changing one line in a generation prompt cost more than $4,000 to rerun. Changing rubric or judge logic cost more than $2,000. Those numbers cover the benchmark at the scale described, with 120 models and five drafts across ten tasks, and they come from the speaker's own account rather than a published invoice.

The benchmark was iterated from four $20-tier subscriptions over four months before reaching its release form, which is how a small team builds this kind of evaluation without a research budget. The early iterations are also where the scoring ratios were tuned.

If you build your own version, the cost curve favors frozen tasks. Pick the prompts and rubrics carefully, then spend on the parts that change less often.

How to fix AI writing slop in a real workflow

Fixing AI writing slop comes down to giving the model enough of your own writing to imitate and to spending review effort on the transcript-level decisions. The speaker's guidance is practical and mostly free.

Start with an extremely long prompt. Include full examples of your published work, negative examples of what you dislike, banned words and phrases, and explicit directions for this task. The prompt cannot be too long, and completeness beats elegance.

Then add one review pass from a strong reasoning model, with the same voice instructions, one to three times at most. Extra loops tend to reintroduce the slop the review was meant to remove.

At API scale, generate many cheap drafts and have a stronger model merge the best sections into one script. The speaker expects this to cost less than one top-model generation and produce better output, though he did not publish a measurement for that comparison.

Finish with a human edit. The benchmark's limitations make this non-negotiable: the rubrics encode one writer's taste, and the leaderboard is a signal about tone match, not proof of which model writes best for every audience.

Where the tooling fits for working developers

Write-once content pipelines benefit from tooling discipline as much as prompt craft, which is where developer education and benchmarks start to overlap. The Skala Blog approach to turning a video into an article is one example of feeding structured source material into a model and then reviewing the result.

Communities such as Crazystack Typescript and the Bootcamp do Dev Doido focus on practical engineering habits, and the same habits apply to evaluation: pin your prompts, keep your examples, and log each change. Platforms such as Crazystack publish courses and material for developers working with TypeScript and modern stacks.

Tools mentioned in this section are credits for related engineering education work, including content by Gustavo Dev Doido, a Brazilian developer educator.

Frequently asked questions about AI writing slop

• What is AI writing slop? AI writing slop is generated text that reads smoothly but carries vague claims, overused vocabulary, repetition and no distinct voice. The six-part breakdown above comes from the October 2026 Tone-Bench talk and frames slop as a voice problem first.

• Is Tone-Bench an objective test of writing quality? No. Tone-Bench measures how well a model imitates one writer's voice across ten script tasks, using LLM judges plus 1,000 human comparisons. A high score means good tone match for that author, not universally better prose.

• Why generate five drafts per model instead of one? Five drafts expose variance. One sample can make a weaker model look competitive, while the median across drafts shows what a team should expect in daily use. Five was chosen as the cost-to-insight middle ground.

• Which model wins Tone-Bench for creative writing? The release discussion named Anthropic Opus 5.5 as the top performer and described GLM 5.3 Flash as an extremely cheap option that scales well. Both are the speaker's leaderboard results, and the standing advice is to pick the model that is easiest to edit.

• How do I avoid AI writing slop in my own content? Give the model a very long prompt with full examples of your published writing, explicit banned words and one review pass from a strong model. Then edit the output yourself. The benchmark's rubrics still encode one person's preferences, so human review stays in the loop.

Turn your own recorded expertise into a written article

The benchmark's lesson applies to more than video scripts. If you already record explanations, interviews or opinion pieces, you are sitting on the exact raw material that keeps generated writing from sounding generic: your real phrasing, your examples and your pace. That material is what makes a voice identifiable in the first place.

Skala Blog turns a YouTube video into a written article by transcribing the recording and generating a draft you can edit. Paste a URL, and the video's own words become the source for the piece, so the result starts from your language instead of a generic template.

If you have a video full of knowledge, you can start at Skala Blog with the URL.

Source video