A couple of times over the past week, I have opened presentation from colleagues that looked perfect at first glance – clean slides, polished language, concise bullets. But as you read deeper, it quickly becomes obvious that they were more AI than human. Lots of content but no real “so what”, nothing that really felt like the author had actually thought about the problem, content or message and a bit like “so what is this telling me”.
That’s AI slop – fast produced, polished output but with no perspective and no real substance – something I am sure we all see far too often – that email that you think “WTF what a load of dribble!”
I also want to credit and link to my friend and fellow MVP MVP Lisa Crosbie who has a great YouTube video on this topic - check it out.
This blog post discusses this and (because it made me think deep about what I create) and shares a 60:40 rule: 60% human thinking, 40% AI assistance that i genuinely use myself and with our customers. You might be thinking that is a lot of human input vs AI but think about it…
“If you want speed and substance, you must put human thinking at both ends of the process – before you ask Copilot to draft or create anything, and after it create the content for you.” “What is in this document that Copilot could never have written on its own?”
Why AI slop happens
Many people assume AI slop comes from AI models hallucinating.
In mid-2026, modern AI models draft well, reason well and understand well.
The primary source of AI-Slop today is (almost always) not model hallucination – it is human absence, poor input and steer.
I am not talking about simple prompts to review, summarise or search – this is about when we use AI to create content, reports, presentations.
The problem is (especially when we have lots to do) is we hand a vague brief to Copilot, accept the first draft, and pass it on as “our” finished work. This creates a two-part problem
- The content we produce looks good (that’s because AI tools are really good doing stuff and creating artifacts) but are not grounded on anything of substance and dont have the human’s true input, knowledge and steer
- Over time AI-Slop creates a feedback loop: Our enterprise AI (grounded in “our work”) learns from this AI‑generated content – language becomes repetitive, absent of our personal and organisational tone, recommendations become generic, and outputs sound like they were written by no one or anyone in particular and just feel like “AI”.
This is not a critique of Copilot or really of the human pilot – it is a critique of how “we” use AI.
From a work context in particular, tools like Copilot should be seen as an accelerator, a good first drafter, an assistant to work alongside to review, guide, suggest and help but not a replacement for judgment.
The difference between a useful draft and a deliverable that earns trust is the human work you add before and after the AI does its job.
The 60:40 rule
You will read this as think “what is so unique about this advice Rob?”. After all this is really how we would work with a human assistant, an intern, a trainee etc that is really really smart but that is not “in your head” and doesn’t “really” know you or what you want.
The rule (or guidelines) is simple. When creating important content (your boss, your team, your board, your clients) the mix of human and AI from a context perspective should be:
- 60% Human: define the need, add judgment, ground outputs in real work, own the result, read, review and refine
- 40% AI: accelerate drafting, structure, and critique
Put simply: humans set the brief and add the unique insight; Copilot does the heavy lifting.
Practical 6‑step workflow (to stop AI slop)
1. Human thinking (20%) – Start with a real brief
Before you open Copilot, capture your thinking. Use a one‑paragraph brief that answers: Who is the audience? What decision should this support? What’s the single takeaway? I use the Copilot mobile app to record quick voice memos before I ask for a draft — it forces me to articulate the problem in my own words. This step is about clarity and intent: the better your brief, the better the draft.
2. AI drafting (20%) – Ask Copilot to do the heavy lifting
Prompt Copilot to produce structure, alternatives, and a first draft. Example prompt: “Create three slide outlines for a 10‑minute executive update aimed at CFOs. Include one slide with a clear ask and one slide with three supporting metrics.” Let Copilot format, summarize, and produce options — but treat the output as a draft, not a deliverable.
3. AI critique loop (10%) – Force critical thinking from the model
Put Copilot into “critical mode.” Ask it to review the draft from multiple perspectives: strategist, risk manager, and CEO. Example prompts:
- “Act as a senior product strategist: identify the three weakest arguments and suggest improvements.”
- “Act as a risk manager: list hidden assumptions and dependencies.” Never accept the first pass without this critique. The critique loop surfaces gaps and forces the model to reason about trade‑offs.
4. Ground in evidence (10%) – Use your org’s data (Work IQ)
Generic content is easy to generate; contextual content is valuable. If you have a Microsoft 365 Copilot license (rather than just the basic Chat version), ask it to pull examples from OneDrive, SharePoint, meeting transcripts, or Teams chats from the last “x” months.
Example: “Include three examples of similar deliverables we completed in the last six months and cite the project names.” That turns a generic paragraph into a defensible, contextual narrative. You can point it at your customer success or proposals folder to ensure Copilot looks where you want it to!
5. Add what only you know (15%) – The human differentiator
This is the single biggest value add. Insert customer quotes, lessons learned, relationship context, or a local constraint that the model cannot know. Examples: a line from a customer call, a metric only your team tracks, or a decision trade‑off you’ve seen repeatedly. If you can’t point to something unique, you’re still producing AI slop.
Concrete example: Instead of “we improved onboarding,” write: “In our last engagement with Client X we reduced onboarding time by 28% – include the customer quote: ‘This cut our time to value in half.’” That single line changes the piece from generic marketing to evidence‑based insight.
6. Human edit + peer review (15%) – Polish and validate
Rewrite for voice and clarity, verify facts, and get a domain expert to peer review high‑stakes outputs. Use a short peer‑review checklist: logic, evidence, audience fit, and the final ask. High‑stakes work deserves high‑stakes review.
Do this before publishing your masterpiece
You have your final document, presentation, business case or proposal. Before you share it, publish it – ASK YOURSELF (Not Copilot) “What is in this document that Copilot could never have written on its own?“
If you can’t answer that in one sentence, you have a bit more work to do.
Practical prompts and quick wins
- Critique prompt: “Review this draft as a CEO. Is the ask clear? Would you approve funding? If not, what’s missing?”
- Grounding prompt: “Find three internal deliverables from the last six months that match this scope and summarise their outcomes.”
- Human insight prompt: “Insert one customer quote and one lesson learned that changes the recommendation.”
Use these as templates and adapt them to your context.
Human Peer Review
Kind of goes without saying but…
Think about if what you are producing needs a human peer review from human subject matter expert. This might be the legal team, a peer, someone not close to your subject than can review as a “does this make sense and do I understand what you are getting across”. AI loves to please and tell you your work is great. Challenge it.
Close and call to action
- AI helps you prepare, creates the draft, critiques your work (and it’s own).
- Humans create the value and uniqueness.
- Keep human thinking at both ends of the process and you’ll stop the slop.
Disclaimer: The % of human input will vary per task, based on the audience, the amount of time you have and if what you are doing is a draft or not. As an example, I often use Researcher in Copilot to find things, create me reports based on “stuff”, but anything I create to pass off as mine must have more me than machine 🙂
Leave a Reply