Source: raw/How_to_Deslopify_Your_AI.md
Creator: Nathaniel Whittemore (NLW) with Nufar Gaspar, The AI Daily Brief, “Operator’s Cut” · URL: https://www.youtube.com/watch?v=QSZKZHeGZr4 · Duration: not recorded · Platform: YouTube (auto-captions, fetched 2026-11-05). Sponsors: KPMG, Robots and Pencils, CData, Blitzy. A companion page is on the show’s website and was not reviewed.
Gaspar defines slop as work “where the person who should have done the thinking basically skipped it… and the cost moved to whoever received the work.” She explains why models default to the average answer and why people let it through, gives five questions for spotting slop, and lays out a five-stage process: think before opening the tool, steer the generation, review with AI personas, teach the model your taste, and set team norms. Whittemore describes how he splits the work between himself and AI on his long-form history show. Both say the effort should match the stakes: a short internal email can be fully AI-made, while customer-facing work cannot.
Key Takeaways
Why slop happens
- The machine. A model predicts the most likely answer, so a generic question gets “roughly the average of what the internet thinks”. Memory helps, but cannot supply a view you have not shared. Raters in post-training reward what feels familiar (she cites Stanford researchers’ term “typicality bias”), and length reads as effort, so defaults come out long and full of lists. She cites a NeurIPS best paper finding that models from different companies converge on the same answers: switching vendor will not buy originality. Image models have house styles too (the too-golden light on food).
- Her test. In incognito mode (no memory) she gave three current frontier models from OpenAI, Anthropic and Google the prompt “write a LinkedIn post about how AI is changing marketing” and got near-identical structure and content. Re-run with her own view, experience, preferred structure and words she never uses, the draft was “a much better starting point for editing”, though still not postable.
- The people. Brains save effort (“cognitive miser”); pilots call it automation complacency; research she cites found that the more people trusted AI, the less critically they checked it. Slop comes three ways, rarely from bad intent: it is easy and you are rushed; you do not know the craft well enough to ask well; or you ask the tool what it thinks before deciding what you think. She calls fighting the third “recovering the lost art of thinking”.
- The numbers she relays: about 40% of US desk workers received work they considered slop in the past month, costing close to two hours to deal with (Stanford and BetterUp); almost 70% of 6,000 surveyed workers had shipped work to someone else without reading it (survey name unclear in the captions); more than half of the music uploaded to Deezer daily is fully AI-generated, but such tracks are about 1% to 3% of listening.
Five questions for spotting slop (yours or anyone’s)
- Could the sender defend every sentence without rereading it?
- Did the sender read all of it before sending?
- Is there anything only this person could have contributed: a view, a decision, real data?
- Does it save the reader time, or is there a better format?
- Would you put your name on it in capital letters? Her favourite, because you often know before you can say why.
Detectors and humanizers. Detectors measure structure (bullet density, sentences of equal length) and flag almost everything on LinkedIn; use them on your own work only to push away from the average, since they do not measure quality. Humanizers are “slop trying not to look like slop”, and their output is still flagged about three-quarters of the time.
Whittemore’s split for a researched long-form show
- Decide up front which parts are yours and which are AI’s. For his history show How We Got Here:
- research goes to AI, as reports of several lengths from several models, because models “emphasize” different things;
- his point of view is never outsourced. He shows why by having a model steelman two opposite positions; it argues either with equal conviction, so it cannot supply yours;
- the narrative structure stays his (“I almost never agree with how AI wants to construct an argument”);
- AI then turns his outline into bullets of key facts and numbers, and he talks unscripted over them.
- For the written companion essay (about 6,000 words): he gives the model the raw transcript of his talk, not just the outline, so it keeps his phrasing. He still spends five to ten hours rewriting, paragraph by paragraph, in an editor he built into the show’s website.
- His three rules: know which parts are yours, never outsource your point of view, and accept there is no perfect efficiency. Deslopping time is a cost of doing business. A short AI email that “just does its damn job” is not slop.
Gaspar’s five stages
- Before you open the tool.
- Know what you think first: walk, read the research, then dictate a messy brain-dump (hers for this episode was 15 minutes) and let the model find the structure.
- Say how sure you are and how hard to push: “I’m about 60% sure the problem is onboarding. Push back hard before you agree with me.” Ask for the counter-view even when you are sure.
- Define the audience, what you want them to feel or do, and the standards for length, format and tone.
- While it generates.
- Go wide: ask for six directions that differ in approach, angle, structure or visuals, including ones borrowed from other fields (how a game designer would handle it), have the model argue for each, then note what you like and what all of them miss.
- Pick your intervention points in advance (outline, first section, narrative, draft, final). A single end-to-end run leaves you with a polished piece that sunk cost tempts you to ship.
- Review.
- Convene a board of AI reviewers: a skeptic, the target reader, a domain expert, the busiest recipient, customer segments, and a devil’s advocate who argues it should not go out. Keep only the feedback you value.
- Respect the read budget: never send what you would not read yourself. If it is long, the recipient’s AI summarises it and “two models… are talking while two people pretend to communicate.”
- Choose the format on purpose (five sentences, a deck, audio, an HTML page, video). If you would enjoy reviewing it in that format, the audience probably will too.
- Teach your taste. Show a capable model your best work, work you admire and work you dislike, and ask what the good examples share and to profile your voice; for visuals, build a mood board. Turn the profile into a skill. Then break it on purpose when the audience, scope or goal changes, or enough time passes, so your work does not all look the same.
- Change the culture.
- Name slop when you receive it, especially as a leader: “send me your take in five sentences.”
- Customer-facing slop is never acceptable.
- Do not try to hide AI completely; make the work feel like yours. Whittemore notes Figma’s published rules for AI writing include labelling AI-written content at the top, so readers know how to read it (for example, a 40-page report meant to be skimmed for interesting parts).
- The common thread (Whittemore): slop moves the cost of work from producer to reader, so the fix is “respecting other people’s time as much as you respect your own.” Gaspar expects handmade work to become the premium product, and Whittemore says being a top AI user now means using it with taste and intention, not producing volume.
Try It
- Add a confidence line to your next analytical prompt: “I’m about N% sure that X. Push back hard before you agree with me.”
- Brain-dump by voice for 10 to 15 minutes before you ask for any draft, then ask the model to structure it and play back what it heard.
- Ask for six directions, not one, and include at least one borrowed from another field.
- Write down two or three checkpoints (outline, first section, draft) before starting an agent run on anything customer-facing.
- Save a reviewer-board prompt with five personas (skeptic, target reader, expert, busiest recipient, devil’s advocate) and run it before sending.
- Build a taste profile: give the model three pieces you are proud of, three you admire and three you dislike, ask for a description of your taste, and save it as a skill.
- Adopt a team rule: label AI-written documents at the top, and answer unread-looking documents with a request for the five-sentence version.
Open Questions
- The studies are relayed without links: the Stanford/BetterUp workslop figures, the 6,000-worker survey (captioned “Glenn”), the typicality-bias work and the NeurIPS convergence paper were not checked.
- The exact model versions in Gaspar’s incognito test are unclear in the captions.
- The companion page with the full playbook was not reviewed.
Related
- Anti-AI Slop Guide — banned patterns for UI and a de-slop review skill for writing; this article covers the thinking that comes before it
- Building an AI Champions Network (Nufar Gaspar) — Gaspar’s playbook for spreading these habits inside a company
- Intent Engineering (Grace Clark) — a voice-guide skill built by having Claude study you
- Friction-Maxxing (Nate B Jones) — keeping your own judgment by seeking disagreement
- Productivity Theater (Nate B Jones) — counting the human work an agent’s output creates
- Marketing Voice Prompt Patterns — prompt patterns for writing in a brand’s voice
- Tulsee Doshi (Google DeepMind) — “the model provides the data, you provide the insight”