Combating AI Slop: A Framework for Effective AI Use
🚨 Combating AI Slop: A Framework for Effective AI Use
TL;DR: AI-generated content is flooding our workflows with diminishing returns. This page defines AI Slop, identifies where it happens, and provides actionable tools to filter signal from noise.
1. AI Slop is flooding orgs — 95% aren't seeing ROI, 40% more content YoY, and 53% of recipients are annoyed by low-quality AI output that shifts work downstream
2. Two main sources — Confluence pages (published without review) and PRs (AI-generated code without self-review)
3. Fix it with machine-first review — Use AI agent review, to review content before publishing/requesting human reviewers
🎯 The Decision
Adopt a "machine review before human review" workflow — AI catches the obvious slop so humans can focus on substance.
✅ The Action
| Confluence | Run pages through AI for summary before publishing. If it can't be summarized clearly → revise first. |
| PRs | Require AI Agent review before requesting human reviewers. |
Longer version
Slop Metrics of this article
| Word Count | 807 words |
| Reading Time | 5 minutes |
Slop Score of this article: 87/100 ✅ PASS
| Verbosity Ratio | 82/100 | 30% | 24.6 |
| Action Item Presence | 95/100 | 35% | 33.3 |
| Decision Clarity | 90/100 | 35% | 31.5 |
| TOTAL | 89.4 |
Scorecard Summary
───────────────────────────────────────────────
│ SLOP SCORE: 87/100 │
│ ████████████████████████████████████████░░░░░░ 87% ───────────────────────────────────────────────
│ ✅ PASS - Ready to publish
│ Thresholds:
│ • 80-100: ✅ Publish
│ • 60-79: ⚠️ Revise recommended
│ • 0-59: 🚫 Requires revision before publish
└───────────────────────────────────────────────
📉 The Problem
| 95% | of organizations are not seeing ROI on AI investments |
| 40% YoY | increase in volume of code and Confluence pages |
| 53% | of recipients report being annoyed by AI-generated work |
| 38% | report being confused |
| 22% | report feeling offended |
What is AI Slop?
AI Slop (noun): Low-quality, AI-generated content that appears productive but adds cognitive burden rather than value. It shifts work from the creator to the recipient—who must now spend time and mental energy deciphering, validating, or diplomatically addressing subpar output.
"The cost of slop isn't just bad content—it's the hidden tax on everyone who has to deal with it."
🎯 The Real Cost
| Recipients | Time spent figuring out what the author actually meant |
| Reviewers | Mental energy diplomatically addressing low-quality PRs |
| Teams | Meeting time re-explaining poorly documented decisions |
| Organizations | Eroded trust in AI tools, slower adoption of effective patterns |
The paradox: AI was supposed to save time. Instead, poorly used AI shifts the time burden downstream—often to more senior or specialized people whose time is more valuable.
🔍 Where AI Slop Happens (Scope)
| Confluence Pages | ✅ Yes | High volume, often unreviewed before publishing |
| Pull Requests | ✅ Yes | AI-generated code submitted without self-review |
| Email/slack | ❌ No | Microsoft Copilot/slack AI summary already provides summarization |
| Blogs | ❌ No | Built-in summarize buttons available |
🎭 Slop Scenarios
Scenario 1: The Wall of Text Confluence Page
What happens: Author uses AI to generate a 10-page design doc. Readers can't find the decision, the tradeoffs, or the action items.
Symptoms:
- No clear structure or TL;DR
- Repeats the same point in different words
- Sounds authoritative but says nothing concrete
Impact: Readers skim, miss critical info, or skip entirely. Decisions get made in Slack instead.
Scenario 2: The Drive-By PR
What happens: Author uses AI to generate code, runs it once, and opens a PR without reviewing their own changes.
Symptoms:
- Obvious bugs that a single read-through would catch
- Inconsistent style with the rest of the codebase
- Comments that describe what the code does (redundant) rather than why
Impact: Human reviewers become unpaid QA. Review fatigue leads to rubber-stamping.
Scenario 3: The "Looks Complete" Artifact
What happens: A doc or PR appears thorough—proper headings, good formatting—but lacks substance.
Symptoms:
- Generic statements that apply to any project
- No specific decisions, numbers, or tradeoffs
- Reads like a template that was never filled in
Impact: False confidence. Teams proceed thinking alignment exists when it doesn't.
🛠️ Effective Tools & Mitigations
For Confluence Pages
| AI Summary Gate | Before publishing pages >1 screen, run through AI: "Summarize this in 3 bullets. What's the decision? What's the action?" If you can't answer, the page isn't ready. |
| Slop Score (Proposed) | Automated scoring on publish: verbosity ratio, action item presence, decision clarity. Flag pages that score poorly for author revision. |
| Reading Time Indicator | Display estimated reading time. If >5 min, require a TL;DR section. |
For Pull Requests
| Machine Review First | Require AI review before requesting human reviewers. AI catches the obvious issues; humans focus on architecture and logic. |
| Self-Review Checklist | PR template requiring author to confirm: "I have read my own diff," "I have run the tests locally," "I can explain why each change exists." |
| Slop Detection Bot | Flag PRs with: no test changes, >500 lines with no description, or AI-generated commit messages ("Update file," "Fix bug"). |
📋 Recommended Workflow
|
|
🎯 Success Metrics
| PR Review Cycles | Reduce average rounds from 3 → 2 |
| Confluence Page Engagement | Increase read completion rate by 20% |
| Reviewer Satisfaction | Reduce "annoyed" responses from 53% → <25% |
| Time to Decision | Reduce time from doc publish to team alignment |
💡 Key Principles
- AI is a draft tool, not a publish tool. Every AI output needs human review before sharing.
- The author owns the quality. Using AI doesn't transfer responsibility to the recipient.
- Volume ≠ Value. A 40% increase in output with declining quality is a net negative.
- Machine review before human review. Don't waste human attention on problems AI can catch.
- If you can't summarize it, you don't understand it. And neither will your readers.
🔗 Related Resources
- Stats are from WSJ, HBR articles.
Comments