← narwal.one/Second Brain
SecondBrain
Ask the Brain
Index/Sourceupdated Wed Nov 04 2026 08:00:00 GMT+0800 (Philippine Standard Time)

How to Deslopify Your AI (AI Daily Brief)

ai-daily-briefwhittemorenufar-gasparoperators-cutai-slopworksloptastepoint-of-viewreview-habitsai-writingcognitive-offloadingread-budgetnovember-2026

How to Deslopify Your AI (AI Daily Brief)

An Operator's Cut episode of The AI Daily Brief (aired Monday 2026-11-02 while Nathaniel Whittemore was on the road; captured 2026-11-04 via Web Clipper). Nufar Gaspar (Super Intelligent) leads, in her second Operator's Cut after How to Go from AI Power User to AI Leader (AI Daily Brief). The episode is "about as applied as it gets": why AI Slop happens in the machine and in us, how to detect it in your own and other people's work, and the habits that keep the work yours, scaled to the stakes. Whittemore contributes the full de-slop workflow behind his new show How We Got Here (Podcast). Gaspar's habits are filed as the Deslopifying Playbook; the training-side mechanism as Typicality Bias.

The line that organises the episode is Gaspar's definition: slop is work where the person who should have done the thinking skipped it, partly or altogether, and the cost moved to whoever received the work. Whittemore's gloss: that "almost wholly explains" the visceral reaction, because "you're leaving them holding the work bag."

The definition and the numbers

  • Merriam-Webster's 2025 word of the year was slop: "digital content of low quality that is produced usually in quantity by means of artificial intelligence." Before that it meant the scraps you feed to pigs.
  • 2026 made it worse: companies rolled out agentic tools to everyone, so work that used to arrive one draft at a time now arrives as documents, decks and code in volume.
  • Scale outside the office: Deezer reports that more than half of the music uploaded to the platform each day is fully AI-generated (verified: >50% and ~90,000 tracks a day as of June 2026, up from 44% in April), while AI-generated tracks make up about 1–3% of streams, most of those flagged as fraudulent.
  • The trigger: a training session with managers and a product leader Gaspar respects, AI-savvy and sharp, who had pushed their employees hard to use AI. The team became "a slop factory": long outputs nobody read before sending, copy that sounds like the same model as every other company, very little original thinking. "It made everybody shallow, the people and the artifact."
  • Humans produced slop long before AI; AI made it cheap and in volume. The standard that follows from the definition: non-slop work is the most effective output for the least effort on the receiving side, and compacting it to the format, length and quality the audience needs is on the responsible AI user.
  • Stanford and BetterUp's "workslop" study (verified 2026-11-04: Stanford Social Media Lab + BetterUp Labs, published in HBR, Sept 2025, n=1,150 US desk workers): 40% received workslop in the past month, at 1 hour 56 minutes of receiver cost per instance, about $186 per employee per month; roughly 15% of content received qualifies as workslop.
  • Glean's survey of 6,000 workers (the transcript renders it "Glenn"; it is the Work AI Index 2026 already in the vault under Botsitting). Verified 2026-11-04 against the report: the figure is 69% of AI users admit to "botshitting", defined as shipping AI-generated work they haven't verified; Gaspar's "without reading it" is a paraphrase. The report adds that 41% sometimes deliver AI work they could not explain if asked, and that the rate falls from 74% where only productivity is measured to 64% where quality is measured too. Gaspar expects listeners to say the real number in their company is higher.

Why the machine produces slop

  • A language model predicts the most likely answer. Give it a generic question and you get roughly the average of what the internet thinks. Memory and context help, but they cannot tell the tool what you think today if you have not shared it.
  • Post-training narrows it further: people rate the answers, and raters reward what feels familiar. Researchers at Northeastern and Stanford named this Typicality Bias in the Verbalized Sampling paper (arXiv 2510.01171, Oct 2025); it is folded into the model, the range of answers narrows, and the default comes out long and listy because length looks like effort and that is what people reward.
  • Switching vendors does not help: last year's best paper at NeurIPS (Artificial Hivemind, Jiang, Choi et al., University of Washington; NeurIPS 2025 Best Paper; 70+ models) found that models from different companies converge on the same answers. Image models are the same story, each with a house style it falls back to (the too-golden light on the food ad).
  • The incognito experiment. Gaspar opened ChatGPT, Claude Opus and Gemini, latest versions, in incognito mode with memory and history off, and gave each the same generic prompt: write a LinkedIn post about how AI is changing marketing. Even she was surprised how similar the results were, in structure and in content. Re-running the same task with her own point of view, her experience, her structural preferences and the words she never uses produced a much better draft. Still not postable as-is, but a real starting point for editing. She recommends everyone run the experiment themselves.

Why humans tolerate it

  • Psychologists have called the brain a cognitive miser since the early 1980s: it saves thinking and energy wherever it can. Pilots call the downside automation complacency: when the machine is usually right, you stop checking. Knowledge-worker research finds the more people trusted the AI, the less critical thinking they applied to its output, and the models make it very easy to feel confident.
  • Three ways slop gets made, almost none with bad intent:
    1. It's easy. It looks fine, you're under pressure, you forward it.
    2. Not knowing the craft. Someone who vibe-codes without knowing what they're doing asks naively and gets naive results; a senior engineer asking for the same thing gets better work because they know what to ask.
    3. Outsourcing the thinking. Asking the tool what it thinks before deciding what you think, because it is easy and the tools are opinionated. "The most troubling part"; preventing it is her "crusade" of the last few weeks, which she calls recovering the lost art of thinking.

Five questions that tell slop from good AI-assisted work

  1. Could the sender defend each and every sentence without rereading it? In honesty, often not.
  2. Did the sender read all of it before it went out? The 40-page report looks fine after a page or two; then you tire and offload the reading to the other side.
  3. Is there anything only this person could have contributed: a point of view, a decision, real data? If anybody could put their name to it, it is not high-quality work.
  4. Is there a more effective way to convey this? Does it save the reader time or cost them time?
  5. Would you feel comfortable putting your name in capital letters on it? The most intuitive test and Gaspar's favourite: sometimes you cannot explain why you are uncomfortable, but you know.

Two technological fixes, both judged weak. AI writing detectors measure structure, column and bullet density, same-length sentences and similar features; turn one on in LinkedIn and nearly everything lights up. You can run one on your own work to move out of the average, but they do not detect quality or originality, which is what matters. Humanizers (skills or processes that make text sound more human) are "slop trying not to look like slop": humanized text is still flagged about three-quarters of the time, and the reader's problem was always the missing human point of view, not the style. Use them to polish something already interesting, never to skip the thinking.

Whittemore's workflow: de-slopping How We Got Here

Whittemore's new show, How We Got Here (Podcast), is a 45-to-60-minute deep dive on the cultural, technological, social and economic history behind one topic (first episode: the fifteen-year history of America falling out of love with tech, which shapes the data-centre debate). It would be impossible for him to make without AI and has the potential to fall into every AI trap. His starting positions:

  • Decide up front which parts are yours and which are AI's. That plan is the foundation.
  • Research is AI's. There is no way to do all his own research. He has models read everything that exists on a topic and turn it into whatever medium gets him up to speed: a five-minute overview or a fifty-page report, from several different models because they emphasise different things ("not so much better or worse; I just want all of it").
  • The point of view is never AI's. His demonstration of why: he routinely has a model steelman two wildly different perspectives and argue each "as though it is deepest conviction." It will argue whichever side you ask, because it is not a person with a perspective, it has "the sum total of all the arguments." So the point of view has to be yours. Everything builds from there.
  • Narrative architecture stays his. "I almost never agree with how AI wants to construct an argument." He does not expect everyone to hold this line; it comes from years of long-form work.
  • Then AI comes back in to produce bullets of key facts and numbers in the sequence of his outline (the same kind of dossier a human researcher produces for the Daily Brief), over which he talks extemporaneously. Nothing is scripted.
  • The companion essay is the concession. Each episode gets a 5,000–6,000-word companion essay, which he cannot write in the same week. Two moves de-slop it: the model drafts from the raw transcript of his talk-over, not just the outline, so it keeps his phrasing and tone even when allowed to expand; and then an unavoidable five-to-ten-hour rewrite, paragraph by paragraph, in a content-management system he built into the show's website so he can reorder and edit at scale. "My talking doesn't sound like the average AI writing. There are things an English teacher would say I do poorly, but they sound like me."

His three takeaways: have a position on which parts are yours versus AI's; never outsource the point of view; accept that there is no perfect efficiency, and treat the de-slop hours as a cost of doing business. One asymmetry he insists on: nobody objects to an AI-written email that is concise and does its job. Its purpose is not to provoke thought, so sameness does not matter. The slop burden scales with what the artifact is for.

Gaspar's five-stage playbook

Filed in full as Deslopifying Playbook. The spine:

  1. Before you open the tool. Know what you think: take a walk, read until you have a view, then brain-dump it to the model in a messy monologue and let it find the structure (this episode began as fifteen minutes of her talking at the tool). Skipping this is "the one thing that has the highest downstream impact." Tell the tool how confident you are, because it is built to be agreeable: "I'm about 60% sure the problem is onboarding. Push back hard before you agree with me." Define the audience, the experience you want them to have, and the standards of length, format and tone.
  2. Shepherding the execution. Go wide: ask for six directions that differ in approach, angle, structure or visuals, borrow from other domains (how would a game designer or a news editor handle it), make the tool argue for each, then squint. Decide your interception points in advance (outline, first section, first narrative, draft, finished good) at the moments where a pivot is cheapest; "here's what I need, go do it end to end" does the most damage because a polished finished piece triggers sunk cost. Convene a board of reviewers: the skeptic, the target reader, the domain expert, the busiest person who will receive it, and a devil's advocate arguing it should not go out at all.
  3. Review habits. Respect the read budget: never send what you would not read yourself. If it is too long for you, your audience will hand it to their own AI to summarise, "and now we have two models talking while two people pretend to communicate." Distil a page into five sentences; it forces you to decide what you think. Choose the format on purpose (deck, podcast, HTML page, image, video); if you would enjoy reviewing it in that format, the audience will probably enjoy consuming it.
  4. Teach your taste. Most people lack the vocabulary of a linguist or an art director, so show a capable model your best work, work by others you admire, and work you dislike, and ask what the good ones have in common. For visuals, mood boards beat adjectives. Turn the profile into a skill that opens any work meant to sound like you. The downside is sameness; the audience notices, so break the mould deliberately when the audience, scope or goal changes.
  5. Culture. Name slop every time it arrives, especially as a leader: "This isn't a good use of my time. Send me your take in five sentences." Hold the gate. Customer-facing work is never allowed to be slop. And do not hide the AI: Whittemore's earlier episode on the Druckenmiller op-ed argued that removing em-dashes and AI tells is a proxy for pretending you invested time; the right move is to make it feel originally produced. Figma has published its rules for AI writing, one of which is that AI-written content is fine but must be labelled as such at the top so readers have context ("let's rip through this and find the interesting parts" is a legitimate mode when declared).

The closing argument

Gaspar: the models will keep getting better, and the better they get the more tempting the generic becomes. Her hunch is that once the agentic hype passes, handmade work becomes the premium and people pay more for a human fingerprint, even messy and not quite grammatical, if it sounds like you. Disciplined high-quality work is job security "regardless of what it is."

Whittemore's self-interested take for an audience that wants to be in the top 1% of AI users: what that means has shifted. At the beginning it meant realising you could use Gamma for better decks faster and get more comprehensive answers. "All of that has been completely democratised as a skill now." The differentiated value moves from producing a lot of stuff to using the tools well, "and well, it turns out, means with taste and with intention." The first phase gave points for showing up to the party early; this one is harder, and the good stuff stands out more in an endless sea of content. Incidental tell: he wanted to use the phrase "load-bearing" in the episode and "literally just couldn't anymore" because of how much AI has used it.

Caveats

  • Operator testimony from two people who sell AI training (Super Intelligent's "intentional AI user" courses are plugged at the end). The study citations were checked on 2026-11-04 and hold, with one wording fix: the workslop figures (Stanford/BetterUp, n=1,150), the Deezer shares, the NeurIPS 2025 Best Paper (Artificial Hivemind) and the typicality-bias coinage (Verbalized Sampling, Northeastern + Stanford) are all real; Glean's 69% is "shipped without verifying", not "without reading". Still unverified: the "three-quarters still flagged" humanizer figure and the Figma AI-writing rules.
  • Sponsor reads (KPMG's "AI amplifiers" study of 500+ early-career professionals, Robots and Pencils, CData, Blitzy) are ads and excluded from the claims above.
  • Auto-captions render Gaspar as "Newfar" and Whittemore as "Nataniel"/"Danielle"; the "Glenn" survey is Glean's.
  • The companion page with the slides is on aidailybrief.ai.

Connects to your work

  • For Sunil's own public writing. The five detection questions and the "name in capital letters" test are a pre-publish checklist for Medium and LinkedIn; the incognito experiment is worth running on a draft topic before writing. The vault's House style section in CLAUDE.md is already a "teach your taste" skill in Gaspar's sense; her sameness warning (break the mould when the audience or goal changes) is the caveat it lacks.
  • For leading a team. "Name it when you receive it" and Figma's label-on-top rule are two team norms that cost nothing; they extend AI Communication Etiquette's declare-the-role rule. The two-hour receiver cost per workslop instance makes slop a budget line alongside Verification Tax and Botsitting.
  • For the 10X IT programme. Of the three sources of slop, the craft gap (naive ask versus senior ask) is the trainable one and maps to Capability Gap; the outsourced-thinking source is the one Cognitive Offloading and Productive Resistance already track. Interception points decided in advance are the same idea as review budgets and verification insertion points in the agentic playbooks.
  • Thought-leadership hooks. "Slop is the transfer of thinking cost from producer to receiver"; "two models talking while two people pretend to communicate"; "we all got points for showing up early; now the product has to matter."

Cross-links

Source

  • source — YouTube, The AI Daily Brief, captured 2026-11-04