ai-slop-audit

AI Slop Audit

The detector. Given any social artefact, it decides how strongly it reads as AI slop, names exactly why, and says how to fix each finding; production-side prevention is the companion anti-ai-slop skill.

Use When

• Does this caption, carousel, email or ad copy look like a bot or ChatGPT wrote it? Tell us why it feels off and what gives it away.
• A campaign, content calendar, blog, email sequence or deck outline is finished; check it before we move to the next piece.
• Review an AI-generated image or video for the tells that make it look fake or generic.
• Grade it A to F with a genericness score, cite each finding, say whether it can ship, and keep the good parts.
• Run the final gate before anything from the engine is published or sent to the client.

Do Not Use When

• anti-ai-slop for guardrails applied while drafting and for humanising a draft.
• policy-ai-content-ethics for cultural bias, AI disclosure and copyright policy.
• premium-commercial-writing for general editorial standards and brochure copy.
• Stop the next asset or iteration when the verdict is F (blocked) until the blocking findings are fixed; never invent evidence for a finding.

Required Inputs

Artefact	Source/provider	Required?	If absent
The artefact itself: caption, post, thread, carousel, ad copy, campaign, calendar, blog, email, deck, profile/bio, image or video	Requester (pasted text, file, rendered post or URL)	Yes	Stop; no audit without the artefact, and never infer findings from a description.
Artefact type(s): written EN, written FR, image, video or multi-asset campaign	Requester, or read from the artefact	Yes	Classify from the artefact and state the classification in the report.
Client business name and industry	Brief or requester	Yes	Judge specificity only against what the artefact claims and mark on-brand fit not assessed.
Country/city of the audience	Brief	Yes	Default to Uganda / East Africa and state the default.
Intended channel and goal	Brief	Yes	Judge the CTA generically and mark channel fit not assessed.
Render, provenance or source evidence (C2PA data, rendered frames, links behind claims)	Requester or file metadata	Conditional (visuals, cited claims)	Mark the affected check NOT_ASSESSED; mark non-visual checks not_applicable.

Workflow

1. Confirm the artefact and inputs, then identify the artefact type(s) and load the matching checklists from the audit method; for a campaign or calendar, plan to audit each post and asset, then the set as a whole.
2. Run the automated gates (focal-word density, em-dash and rule-of-three patterns, opener clichés, mechanical formatting, citations and statistics, image and video checks); any [BLOCK] hit fails the artefact outright.
3. Compute the 0–100 genericness score from burstiness, focal-word density, duplication across the set and template similarity, and name its drivers.
4. Run the human-judgement review: substance, intent, specificity, hard parts, localisation, visuals and the artefact-specific checks.
5. Apply the ME1-ME7 machine-error checks at post, slide, caption and campaign-sequence level, and AS1-AS7 with evidence mode (cli, browser, llm_only, human_review) for visuals and rendered posts.
6. Grade A, B, C or F and write the report: blocking findings, slop findings by severity, what's good, recommended next step, each finding with quoted evidence and a concrete fix.
7. If the verdict is F, stop: the next asset or iteration does not start until the blocking findings are fixed.
8. After fixes, rerun the audit on the corrected artefact and log the new verdict; run it again at each checkpoint and as the final gate before publishing.

Yardstick and verified evidence

Low-quality content produced in quantity by AI and pushed at people who did not ask for it (Merriam-Webster 2025 Word of the Year, verified). Three diagnostic properties (Kommers et al., "Why Slop Matters", arXiv 2601.06060, verified): superficial competence, asymmetric effort, mass producibility. The human tell: absence of intent. You are measuring how strongly an artefact exhibits these.

The threat is documented, not rhetorical. Cite these where a client questions why the audit matters; do not embellish them or add unsourced figures:

• Spracklen et al., USENIX Security 2025 (verified): 19.7% of package references suggested by code-generating models were hallucinated — the "slopsquatting" supply-chain risk, relevant whenever AI output names a tool, plugin, or integration to install.
• Veracode (verified): 45% of AI-generated code samples introduced a known vulnerability; cross-site scripting failures in 86% of relevant cases; log-injection failures in 88%. Treat any AI-suggested embed, pixel, or script with the same suspicion.

These belong in the evidence column, not as decorative statistics. Use them only when on point.

Grades

Aggregate into a grade:

Grade	Meaning	Trigger
A — Clean	No blocking hits; genericness low; substance and intent present	ship
B — Minor slop	A few automated hits, no blockers; some genericness	fix listed items
C — Slopy	Multiple automated hits or weak substance/intent	rework before ship
F — Blocked	Any [BLOCK] blocker (fabricated stat/citation, garbled publishable image text, deceptive claim) OR no substance at all	do not ship

Outputs

Artefact	Consumer	Acceptance condition
AI slop audit report (verdict, genericness score, blocking findings, slop findings, what's good, next step)	Content owner and client reviewer	Every finding cites evidence from the artefact and carries a concrete fix; any [BLOCK] forces an F.
Verdict log entry for the checkpoint	Delivery lead and the next workflow	Records artefact, date, grade and whether progression is blocked.

Evidence Produced

Evidence	Format	Acceptance condition
Finding evidence	Quoted line, slide number, colour value, frame reference or URL per finding	No finding without evidence; inferences marked "(inference)".
Genericness score with drivers	Score 0–100 and named drivers in the report header	Drivers are measurable (for example banned-word density per 500 words).
Overlay record	ME1-ME7 and AS1-AS7 rows with evidence mode	Unavailable render evidence is NOT_ASSESSED; non-visual checks are not_applicable.

Capability and Permission Boundaries

Read and search only; analysis is read-only. Publishing, spend, live account changes, outreach and personal-data processing need explicit, action-specific client authority. The audit reports findings and fixes; it does not rewrite or publish the artefact.

Degraded Mode

Without the artefact or its render and source evidence, return the narrowest qualified result and mark the affected checks not assessed. The written-content gates can still be run on supplied text, with image, video and citation checks recorded as NOT_ASSESSED and no grade above the evidence seen.

Decision Rules

Condition	Action	Failure or risk avoided
Any [BLOCK] hit (fabricated statistic or citation, garbled publishable image text, deceptive claim) or no substance at all	Grade F and block progression to the next asset or submission.	Shipping fabricated or deceptive content.
The only signals are emojis, repeated openers, three-part phrasing or a familiar cadence	Do not call the post AI-written; cite the exact unit and the audience-value loss as a style finding.	False authorship accusations.
A repeated safety warning, accessibility label or approved campaign message recurs	Record it as a functional exception, not slop.	Stripping required repetition.
A visual uses purple gradients, glassmorphism, neon glow, AI-beige defaults, decorative editorial scaffolding or decorative motion	Report a blocking visual finding unless a functional state, accessibility need, data encoding or approved brand reason is recorded.	Generator-default visuals shipping.
SynthID is absent from an image	Do not treat absence as proof either way (Google-only).	Misreading provenance.
The artefact is a campaign or calendar	Audit each post and asset, then the set (do all twelve captions share one template?).	Missing set-level uniformity.
The artefact is clean	Report "This artefact is clean" honestly.	Invented flaws that erode trust in the audit.
A claim, date, link, testimonial, platform limit or performance number appears	Verify it and distinguish internal links from evidence citations; unverifiable items are NOT_ASSESSED or a [BLOCK] if fabricated.	Treating a plausible figure as verified.

Quality Standards

• Every finding is evidenced with a quoted line, slide number, colour value, frame reference or URL from the artefact.
• No fabricated flaws: nothing is raised that is not actually present, and a clean verdict is reported honestly when earned.
• A 0–100 genericness score is given and its main drivers named.
• Blocking and non-blocking findings are separated; any [BLOCK] blocker is called out distinctly and forces an F.
• Each finding has a concrete fix, a specific action, not "improve this".
• What's good is preserved: substantive, authored elements are named so a fix does not strip them.
• Localisation is judged: the report states whether the artefact fits the Uganda / East Africa (or named) market.
• Verified evidence only: the Merriam-Webster, Kommers, Spracklen, and Veracode figures are used verbatim and only when on point; no new statistics are invented.

Anti-Patterns

• Running the audit once at the end. Fix: run it after each major iteration and log the verdict each time, like a test suite at every checkpoint.
• Presenting a guess as a measured fact. Fix: mark inferences "(inference)".
• Padding the report with invented flaws. Fix: raise only what is present; "This artefact is clean" is a valid, wanted verdict.
• Stripping the strong parts while fixing slop. Fix: list what's good so the rewrite keeps it.
• Grading one caption from a set in isolation. Fix: audit each asset and the set as a whole for shared templates.
• Letting a C or F artefact move on because the deadline is close. Fix: an F blocks progression; a C needs rework before ship.
• Using the Spracklen or Veracode figures as decoration. Fix: cite them only where a client questions why the audit matters.

References

• Audit method: read when running the cadence, the layered checks by artefact type, the ME1-ME7 and AS1-AS7 overlays, the report template, the responsibility audit or the see-also routes.
• anti-ai-slop: read when fixing findings or humanising the draft (prevention companion and banned list).
• policy-ai-content-ethics: read when a finding involves disclosure, copyright or cultural bias.
• AI-slop responsible publishing standard: read when separating style findings from deception, cultural harm or provenance.
• Creative review gate: read when the audited asset is finished creative going to release.
• ai-readiness-diagnostic: read when the question is AI maturity rather than an artefact.
• Repository agent guide: read when checking the engine-wide cadence rule and anti-slop gates.