How to Deslopify Your AI

Slop is work whose creator skips the thinking and transfers the effort to whoever receives it. Good work gives its audience something useful in a suitable format, length, and quality, with as little effort required from them as possible. AI has made low-quality output cheaper and more abundant; the speaker says companies’ 2026 adoption of agentic tools brought agents into producing documents, presentations, and code at scale. Deezer reportedly found that more than half of music uploaded daily was fully AI-generated, though AI-generated downloads made up only about 1–3% of listening. A Stanford and BetterUp study reportedly found that about 40% of US desk workers received work they considered slop in the previous month, costing them close to two hours. In a separate survey of 6,000 workers, almost 70% said they had sent work to someone else without reading it.

Generic prompts tend to produce generic answers because language models favor likely responses, and human ratings can reward familiar ones—a tendency researchers call “typicality bias.” A paper the speaker attributes to NEAPS reportedly found that models from different companies can converge on the same answers, so changing vendors may not fix generic output. Image models also default to house styles, such as golden lighting in food images, unless users specify what they want. In an incognito test with memory and history off, ChatGPT, Claude, and Gemini all produced generic LinkedIn posts about AI changing marketing. Adding personal experience, a point of view, and preferences improved the draft, but it still needed editing. Detectors look for patterns such as structure, colon and bullet density, and sentence similarity—not quality or originality—and the speaker says “humanized” text is still flagged about three quarters of the time.

A stronger process begins before opening an AI tool: form an initial view, then brain-dump what you heard, think, worry about, and want to avoid. The speaker says this episode began with 15 minutes of messy dictation. Tell the tool how confident you are and how strongly you want it to challenge you; ask for a counterview because the speaker finds AI agreeable and says it may otherwise help justify an existing opinion. Before production, define the audience, outcome, length, format, and tone. Ask for six directions that vary in angle or structure, compare their strengths and gaps, and set review points at the outline, first section, draft, and final work so you can redirect early. For important work, seek feedback from distinct perspectives—a target reader, skeptic, domain expert, busy recipient, or devil’s advocate—and use only feedback you value.

Nathaniel’s podcast, How We Got Here, shows how AI’s role can depend on the output. For a 45-minute-to-one-hour episode, AI and a research team help make broad research feasible and generate key facts and numbers, while Nathaniel keeps the point of view and narrative architecture human-led. He speaks from an ordered outline and supporting material rather than a script, noting that AI can argue passionately for opposing positions and therefore cannot supply the creator’s point of view. For companion essays, he gives AI both the raw talkover transcript and outline to help it match his phrasing while adding context. The roughly 5,000-word essays still take substantial editing: one takes him about five to 10 hours, and he says it would not exist without AI. He built paragraph-by-paragraph editing and reordering into the website’s content management system and counts the time spent deslopifying as part of AI’s cost.

The effort and scrutiny should match the stakes. The speaker says low-impact internal work can be fully AI-made, while customer-facing work representing a brand needs deliberate processes. Ask whether the sender has read the whole piece, can defend every sentence, contributed something only they could, chose a suitable format, and would put their name on it. Respect the audience’s read budget: choose a format people will use and do not send work you would not read yourself. Leaders should call out generic, unconsidered work, ask for a concise summary or a more carefully developed version, and keep slop from going out under the team’s name. AI use need not be hidden, but the result should feel intentional; Figma’s stated approach is to label AI-written work so readers have context. A 40-page report, for example, might be presented as material to explore rather than as finished work.

To help AI match your taste, provide examples you like and dislike and ask it to identify their common features; for visuals, use a mood board. A reusable taste profile can improve consistency, but may also make work repetitive, so refresh references and depart from the established style when the audience or goal changes. The aim is to use AI to improve quality and expand what you can do, while thinking more sharply, exploring more directions, and challenging your work instead of outsourcing the thinking. As basic capabilities for producing more have become widely accessible, the speaker says the advantage of top AI users has shifted toward taste and intention. The speakers expect handmade work with a human fingerprint, even if messy or grammatically imperfect, to command a premium, and believe original thinkers and disciplined workers will be rewarded as audiences get better at filtering abundant content.