On this page
The definition
Colander treats AI-generated content as slop when it is mass-produced with little human effort to capture attention or money, and gives you little in return. AI use alone never makes something slop.
The unit that matters most is the source: the channel, profile or page. Slop is a production pattern, and patterns show at the source level long before a single item gives itself away.
The three tests
An item or source is slop when it is AI-generated and meets at least two of three tests.
| Test | Question | Research root |
|---|---|---|
| Low effort | Is there little sign of human authorship, such as original footage, commentary, editing judgment or fact-checking? | Asymmetric effort, negligible exertion |
| Mass-produced | Does the source publish at a volume and sameness that points to an automated pipeline? | Mass producibility, volume, velocity |
| Hollow | Does it look competent while carrying little information, containing errors, or existing mainly to hold attention or push a link? | Superficial competence, information utility and quality |
What is not slop
- AI-assisted work with clear human authorship, such as AI used for editing, captions, dubbing or illustration.
- Openly artificial art, satire, parody and political expression, which have real expressive and dissent value.
- Low-quality human-made content. It may be bad, but it is out of scope.
- Deepfakes, fraud and intimate-image abuse. These are harms beyond slop and belong with platform and legal reporting. The extension offers a shortcut to the platform's own reporting.
Three types you can tag
| Type | What it is | Example |
|---|---|---|
| Filler | Generic content tuned for engagement, with no real subject and no next step | Surreal animal clips, endless AI "history" narration |
| Bait | Slop that routes the viewer to a link, product, install or scam | AI image posts with affiliate links in the comments |
| Deceptive | Synthetic content presented as real | Fabricated news events, staged rescue videos |
Deceptive slop is hidden first. Openly artificial, expressive work is labeled, not hidden.
The four evidence layers
No single signal hides anything. Each source and item is scored from four layers, and nothing is hidden by default until two independent layers agree.
- 01
Provenance
Is it AI-generated?
- The platform labels it AI-generated
- Content Credentials say it was made with AI
- The creator says it is AI-made
- A generator watermark is visible
- 02
Source behavior
Is it mass-produced?
- Posts at a volume no person could sustain
- Most recent items are AI-made
- One template across titles and thumbnails
- Many near-duplicate items
- Routes viewers off the platform
- Other pages share the same captions
- 03
Content rubric
Is it low effort and hollow?
- Taggers found little human effort
- Taggers found it hollow
- 04
Community consensus
Do people who usually disagree both call it slop?
- Tagged as slop by the community
- The community says it is not slop
Five verdicts
| Verdict | Evidence required | At Standard |
|---|---|---|
| Slop | AI evidence, plus a mass-produced source, plus community consensus or staff review | Hidden |
| Likely slop | AI evidence, plus behavior or rubric signals, with consensus still forming | Hidden |
| AI-made | AI evidence only | Labeled |
| Disputed | Tags and counter-tags split, or an appeal is open | Labeled |
| Clear | "Not slop" consensus or a successful appeal | Shown |
Labels describe content, never the people who made it.
Safeguards against wrong calls
- AI evidence is a gate. Without it, nothing can be rated slop, however low its quality.
- Two layers must agree before anything is hidden by default.
- Tags alone never make anything Slop. Slop needs AI evidence, a mass-produced source, and consensus or staff review.
- Mixed sources are never hidden as a whole. Their AI items get item-level labels.
- Sources with large audiences need staff review before a list-wide Slop verdict.
- Every hidden item states which signals agreed, and one click shows it again.
- Source verdicts expire and are re-scored every 90 days. An appeal triggers re-review at once.
A source counts as mostly AI when at least 80% of its recent items carry AI evidence, following Kagi's published rule. Other thresholds are starting values, calibrated against a hand-labeled set of at least 1,000 sources.
Strictness
You choose what each verdict does to your own feed. Standard is the default. Colander never decides what an adult may see.
| Level | Slop | Likely slop | AI-made | Disputed | Clear |
|---|---|---|---|---|---|
| Label | Labeled | Labeled | Labeled | Labeled | Shown |
| StandardDefault | Hidden | Hidden | Labeled | Labeled | Shown |
| No AI | Hidden | Hidden | Hidden | Labeled | Shown |
Label
- Slop
- Labeled
- Likely slop
- Labeled
- AI-made
- Labeled
- Disputed
- Labeled
- Clear
- Shown
StandardDefault
- Slop
- Hidden
- Likely slop
- Hidden
- AI-made
- Labeled
- Disputed
- Labeled
- Clear
- Shown
No AI
- Slop
- Hidden
- Likely slop
- Hidden
- AI-made
- Hidden
- Disputed
- Labeled
- Clear
- Shown
A hidden item leaves the page the way an ad does under an ad blocker: no gap and no placeholder, and grids close up so their rows stay full. In swipe feeds it is skipped. The toolbar badge still counts it, and the popup lists it with Show, Always allow, Not slop and Why, so every call can be checked and undone.
Disputed items always stay visible with their mark, and Clear items are always shown.
How signing works
Colander works like an ad blocker. Your extension downloads the whole core list, carrying no identifier at all, and matches every feed card against it on your device. Colander never asks a server about the page you are viewing.
Each copy of the list is signed by Colander. The extension checks the signature before it uses a list, refuses any copy that was changed on the way, and keeps working from its last good copy, offline too. It looks for updates about once an hour.
How consensus works
Each install gets a random pseudonymous ID. A tag applies on the tagger's own device at once and enters the shared pool with a weight based on that tagger's track record. New installs count for little, and bursts of tags on one source from new installs freeze its consensus until staff look.
Paying or giving never changes tag weight, review priority or any verdict. Bridging, where a verdict needs agreement between groups that usually tag differently, is added once tag volume supports it.
The tagging rubric
Tagging takes two clicks, on the content itself. You choose one of three tags.
- Slop. Optionally add a type (Filler, Bait or Deceptive) and tick which of the three tests apply.
- AI-made but fine. AI-generated, and fine. This counts toward the AI-made question only.
- Not slop. A counter-tag. Enough of them move a verdict to Disputed or Clear.
When you tick tests, ask three questions about the item.
- Useful. Would a viewer come away knowing or feeling something they came for?
- Accurate. Are the facts, dates, names and images right?
- Original. Is there footage, commentary or judgment you would not find on ten other sources?
AI voices, dubbing and translation are not evidence of slop on their own. Many creators who are non-native speakers or disabled rely on them.
Appeals
Every label links to a public page for its source, showing the verdict and the evidence behind it. A creator who disagrees can appeal from that page. Appeals are free.
- Start an appeal from the source page with an email address and a short statement.
- Prove control of the account by adding a short code to its description or bio, then choose Verify.
- Once verified, the verdict changes to Disputed and nothing from the source is hidden while staff review.
- The outcome and the reasoning are published in the decision log.
Unverified appeals expire after 14 days.
Why an AI detector cannot be the method
- Open-source detectors lose about half their accuracy on real social media. On one benchmark, accuracy fell 50% for video, 48% for audio and 45% for images.
- Text detectors punish the wrong people. Seven detectors misjudged 61.3% of essays by non-native English writers as AI-written.
- Detecting AI is not detecting slop. They are different tasks, and standard automatic measures did not reproduce editors' judgments.
- Labels and watermarks can be stripped, covered or never applied.
Detector models are therefore optional, run on your device, and can only add weight to the AI-made question. They can never mark something as slop.
Research
- Kommers et al., "Why Slop Matters", ACM AI Letters, March 2026Three family-resemblance features: superficial competence, asymmetric effort and mass producibility.
- Silbey and Hartzog, "AI Slop", September 2026Negligible exertion, asymmetrical imposition and domain degradation, scored along a spectrum.
- Shaib et al., "Measuring AI Slop in Text", 2025Information utility, information quality and style quality. Not all AI text is slop.
- Columbia IGP, "AI Slop and the Information Ecosystem", June 2026Slop as a high-volume subset of AI content, between benign and deliberately harmful.
- Madsen and Puyt, "The 7Vs of AI Slop", 2025Seven dimensions, including volume and velocity.
- Deepfake-Eval-2024 benchmarkOpen-source detectors lose about half their accuracy on real social media.
- Liang et al., 2023Seven detectors misjudged 61.3% of essays by non-native English writers as AI-written.
- Kagi SlopStopThe 80% rule for sources, and item labels for mixed sources.
- Community Notes bridgingAgreement between people who usually disagree.
- SponsorBlockRandom per-install IDs, votes and reputation instead of accounts.
Version 1, October 2026. Changes to this definition are published on this page with their date.