An Operator's Cut with Nufar on the most practical problem of the agentic era: slop. Agents now produce documents, decks and code in volume, and much of it lands unread on someone else's desk. The episode covers why the machines produce slop, why we let them, how to spot it, and the habits that keep the work yours.
Slop is skipped thinking with the bill sent to the reader. The new edge is using AI well, not using it more.
The first phase of AI at work was learning to produce volumes of content. That skill is now democratized, and the result is slop everywhere: work where the person who should have done the thinking skipped it, so the cost moved to whoever received it. Models default to the average and people default to trusting them, so slop is the path of least resistance. The fix isn't abandoning AI. It's knowing what you think before you open the tool, deciding in advance where you step in, scaling effort to the stakes, and respecting other people's time as much as your own. In an endless sea of content, the good stuff stands out even more.
Calling it 'slop' was once easy to dismiss as anti-AI grumbling. Now AI writing, images, video and code are everywhere, and that content has a real cost. If the first phase of AI at work was learning to produce more, the next is distinguishing yourself by using these tools well.
Merriam-Webster made 'slop' its 2025 word of the year. 2026 made it worse: companies rolled out agentic tools to everyone, and work that once arrived one draft at a time now comes from agents in volume. Outside the office, more than half of Deezer's daily music uploads are fully AI-generated, but they make up only 1–3% of what people actually listen to.
Nufar describes a sharp, AI-forward product leader whose team embraced AI and then turned into what they themselves called a slop factory. The outputs were long and unread, sounded like the same model as everyone else's, and showed very little original thinking. Their verdict: it made everybody shallow, both the people and the artifacts.
Slop is work where the person who should have done the thinking skipped it, partly or entirely, and the cost moved to whoever received it. Humans produced slop long before AI; AI made it cheap and plentiful. The definition also sets the standard: good work delivers the most value for the least effort on the receiving end.
A Stanford and BetterUp study found about 40% of US desk workers had received workslop in the past month, at a cost of close to two hours each on the receiving side. In a Glean survey of 6,000 workers, almost 70% said they had shipped work to someone else without ever reading it.
Models predict the most likely answer, so a generic question gets roughly the internet's average view. Post-training adds 'typicality bias': raters reward what feels familiar, which narrows the range and makes long, listy output look like effort. A top NeurIPS paper found models from different companies converge on the same answers. Image models fall back on their own house styles, like that too-golden food lighting.
Nufar gave the latest OpenAI, Anthropic and Google models the same prompt in incognito mode, with memory off: write a LinkedIn post about how AI is changing marketing. The structure and even the content came back strikingly similar. Rerunning it with her own point of view, experience and style preferences produced a much better starting point for editing, though still not one she would post as is.
Brains save effort wherever they can, and pilots call the downside 'automation complacency': when the machine is usually right, you stop checking. Research on knowledge workers found that the more people trusted the AI, the less critical thinking they applied. Slop rarely comes from bad intent. It comes from ease under pressure, from not knowing the craft well enough to ask well, and, most troubling, from outsourcing the thinking itself.
Five questions sort slop from good AI-assisted work. Could the sender defend every sentence without rereading it? Did they read all of it before sending? Is there anything only this person could have contributed? Does it save the reader time or cost them time? The most intuitive one: would you put your name in capital letters on it? If not, there is more work to do, even if you can't say why.
Detectors measure surface tells like colon density, bullet density and uniform sentence length, and they flag nearly everything on LinkedIn. They can help nudge your own work away from the average, but they don't measure quality or originality. Humanizers are mostly 'slop trying not to look like slop.' Humanized text is still flagged about three-quarters of the time, and the reader's real complaint was always the missing point of view, not the style.
The burden varies wildly by artifact. NLW argues nobody minds AI emails that are concise and to the point, because their purpose isn't to provoke thought. Nufar agrees effort should scale with stakes: no one wants artisanal meeting notes, but brand, customer-facing and high-impact original work demands real intentionality.
For his new long-form show How We Got Here, NLW has AI read everything on a topic and generate reports of every length across several models. He uses those to build or challenge his own view. His proof that AI can't own your perspective: ask it to steelman two opposing positions and it argues each with apparent deepest conviction. It has no perspective, only the sum of the available arguments.
The structure and flow of the argument stay in NLW's hands. AI comes back in only to fill his outline with key facts and numbers, which he then talks over extemporaneously without a script. The workflow is a constant back-and-forth: perspective and architecture stay human, and AI makes it viable at production speed.
An AI-drafted 6,000-word companion essay will be slop by default. NLW feeds the model his raw talk-over transcript rather than just the outline, then spends 5–10 hours rewriting in a paragraph-level CMS he built for the site, keeping quirks an English teacher would mark wrong because they sound like him. The lesson: there's no perfect efficiency, so budget de-slopifying as part of the cost of doing business.
Take a walk and read the research until you have a view of your own, however messy. Then brain-dump it: dictate everything you've heard, what worries you, and what you definitely don't want, and let the model find the structure. This episode began as 15 minutes of Nufar rambling to an AI tool.
Models are built to agree, so a strong opinion going in gets justified rather than tested. State your confidence and how hard you want to be challenged: 'I'm about 60% sure the problem is onboarding. Push back hard before you agree with me.' Even when you're sure, ask for the counter view. Also define the audience, the experience you want them to have, and your length, format and tone standards up front.
Request six directions that differ in approach, angle and structure, including ones borrowed from other domains, like how a game designer or a newsroom would handle it. Have the model argue for each. Then squint across them: note what you like in each and what's missing from all of them, and follow the ones you're drawn to.
A polished finished piece triggers sunk-cost thinking, and the tools make it hard to change direction from there. Decide in advance where you'll step in: the outline, the first section, the narrative, the draft. That way you review at points where a fix or a full pivot is still cheap.
Have AI review the work as distinct personas: a skeptical reader, a domain expert, the busiest recipient, different customer segments, and a devil's advocate arguing it shouldn't go out at all. Use only the feedback you find valuable. It works at early junctions and on the finished product.
If it's too long or unclear for you, your audience will just hand it to their own AI to summarize. Distilling a page into five sentences forces you to decide what you think. AI makes any format possible, from decks to podcasts to HTML to video, so choose it on purpose. If you'd enjoy reviewing it in that format, your audience will likely enjoy consuming it.
You don't need the vocabulary to describe your taste. Show a capable model your best work, work you admire and work you dislike, and ask it to profile what the good ones share. Use mood boards for visuals. Codify the result as a reusable skill, but watch for sameness the audience will notice, and refresh it with new inspiration whenever the audience, scope or goal changes.
Culture changes when long, generic, clearly unread work gets called out, kindly or not: 'This isn't a good use of my time. Send me your take in five sentences.' Hold the gate internally, and never let it reach customer-facing work.
Figma's rules for AI writing say AI-written work is fine but should be labeled at the top so readers have context. That's the right frame for a 40-page report meant to be mined, not presented as finished. NLW says the cost-transfer definition of slop explains why people react so viscerally: you're leaving them holding the work bag. Transparency and disclosure come back to respecting other people's time as much as your own.
The better the models get, the more tempting the generic answer becomes. Nufar's hunch is that once the agentic hype settles, handmade work with a human fingerprint, messy or not, will command a premium. Disciplined, original work is the job security, and companies that keep slop out of how work travels, especially externally, will stand out with customers.
Early on, being a top user meant realizing you could make better decks faster with tools like Gamma. That skill is now democratized for everyone touching these tools. The differentiated value has shifted from producing lots of stuff to using AI with taste and intention. It's a harder phase, but in endless seas of content, the good stuff will stand out even more.
