SniffSlop

Hover a paragraph, find out how likely it is written by AI, or define whatever questions you want to ask with Jev.

demo.webm

Install

No build step — it's plain JS, load the folder as-is.

Chrome

1. Go to chrome://extensions
2. Turn on Developer mode (top right)
3. Load unpacked → pick this folder

Firefox

1. Go to about:debugging#/runtime/this-firefox
2. Load Temporary Add-on… → pick manifest.json inside this folder

Temporary add-ons are cleared when Firefox restarts; a permanent install needs a signed build.

API key

Obtain your Jev API key at typesafe console. Paste your API key into the extension by clicking the extension toolbar icon and hit Save. Use it at your own risk.

The key is stored with storage.local, which keeps it on this machine and this browser profile — it is not browser-synced, so it does not travel to your other devices. Nothing is collected by me, no telemetry, no analytics, and no server belonging to this extension. The only host it can reach at all is api.typesafe.ai, because that is the sole entry in host_permissions; any request anywhere else would be blocked by the browser.

The key is read only by the background script, which attaches it as a Authorization: Bearer header on calls to TypeSafe. The content script running inside web pages never sees it — it hands text to the background script and gets an answer back, so page JavaScript has no path to it.

Use

1. Add your API key as above.

2. Press the shortcut (default Cmd+Shift+X / Ctrl+Shift+X) to arm sniff mode on a page.

3. Hover any block of text with 25+ words. It outlines grey while the request is in flight, then turns green → amber → red and shows a card with one section per configured question.

   The outline colour comes from the first question that maps onto a scale — a score (position across its levels) or a noul (its probability). Choice answers have no inherent ordering, so if every question is a choice the outline stays neutral blue and the card carries the detail.

   Hue is the answer, saturation is the certainty. An uncertain answer looks washed out instead of confidently green or red. Confidence gets no row of its own: it is computed from the probability distribution the card already draws, so showing it would restate the bars. Nouls have no confidence field, so certainty is derived from the distance to 0.5.

4. Shortcut again or Esc to disarm.

Results are cached per page by text hash, so re-hovering a block you already scored is instant and costs nothing until you reload. Saving anything in the popup clears that cache — answers are only valid for the settings that produced them — and immediately re-scores whatever is under the cursor.

Settings

• Questions — any number, all three primitives. Every hover sends them in a single request: the API scores each question independently against the same document, and the document is only transmitted once, so N questions cost far less than N calls.

  type	criteria	answer
  score	ordered array, low → high, 2–10 levels	score (weighted position across levels) + confidence + probabilities
  choice	object of option_name → description	choice + confidence + probabilities
  noul	optional true/false descriptions	noul, a single 0–1 probability of yes

  The default is one score question: How likely is this text AI slop, with levels Likely human / 50/50 slop / AI Slop. A score of 1.43 there means "leaning slop".

• Minimum words per block — default 25, raise it to ignore short blurbs.

• Shortcut — on Firefox, Change records a new binding directly in the popup (commands.update()). Chrome has no equivalent API, so there the button opens chrome://extensions/shortcuts instead. Either way the browser can refuse a combination it reserves for itself.

Define your own questions

Slop detection is just the default. The question is whatever you want to ask of a paragraph, and you can stack several — they cost one request between them. Open the popup, hit + Add question, pick a type, and write the instructions.

Pick the type by the shape of the answer you want:

• noul for a yes/no — "Does this cite a source?", "Is this a sales pitch?". You get one number: the probability of yes. Criteria are optional; add them when "yes" needs pinning down.
• score for a spectrum — "How technical is this?" with levels Layperson → Practitioner → Specialist. Write the levels in order, low to high. The answer is a weighted position across them, so it lands between levels rather than snapping to one.
• choice for named buckets with no ordering — "What is this paragraph doing?" with options argument, evidence, anecdote, filler. Give each option a description saying when to pick it; that description does most of the work.

A few things that make answers better:

• Describe the levels, don't just name them. Broken, but a workaround exists beats Medium. The model reads the descriptions.
• Order matters for score, not for choice. Score levels must run low to high; choice options are unordered, which is why a choice answer leaves the outline neutral.
• Ask one thing per question. A level that mixes two dimensions produces a flat distribution — the answer washes out, which shows up as a desaturated outline.
• The question id is the key in the API request, so it gets slugified on save: Is it technical? becomes is_it_technical.

Changing a question's type converts its criteria to the new shape, because sending an array where the API expects an object is a 422.