How it works

Short text

Under 12 words, the detector is skipped. The model signal counts as 0.5 and confidence is 0.3. These items are marked heuristics only and count toward the ELO right away, because there is no model score anyone could fake.

The ELO

Every verified item is a match against a baseline human rated 1000. A sloppy item is a win, a human one is a loss.

E  = 1 / (1 + 10^((1000 − R) / 400))
S  = 1    if slop ≥ 70
     0.5  if slop is 40 to 69
     0    if slop < 40
K  = 32 for the first 30 items, then 16
T  = min(1.5, 1 + slop_weeks / 52)
R' = R + K × (S − E) × T   if S ≥ E
R' = R + K × (S − E)       otherwise

• R is the current rating and E is how likely this author was to post slop, given that rating.
• S is the result of this item. K makes new authors move faster.
• T is tenure: every week in a row averaging 70+ slop adds a little, up to 1.5x. It only multiplies gains, never losses, which is why the rating climbs the longer you post it.
• Example: a new author at 1000 posts something that scores 90. They gain 32 × (1 − 0.5) = +16.

Tiers

Every author gets a tier as soon as one of their posts is scanned. The tier comes from the estimated slop score: the average slop of their newest 20 scanned posts, including posts still waiting for a second install. Comments do not count toward it.

• Pure Slop guard: Pure Slop needs 3 or more scanned posts. With fewer, an estimate of 85+ shows as Generated.
• Leaderboard: the all-time board ranks every author with 3 or more scanned posts by estimated slop, ties broken by the number of scanned posts. The 7 and 30 day boards rank by the average slop of posts first seen in that window, from 2 posts.
• The ELO still runs as described above and shows next to the tier as a secondary number. It only moves with verified items, so it is slower and harder to game. On the ELO, a rating of 1700+ needs mean detector confidence above 85% over the last 30 days to count as Pure Slop.

Verification

The detector runs on users' computers, so one client could lie. A score only moves an ELO after independent installs agree.

1. First upload: the item is stored as provisional with one report. Badges show its score, marked provisional. The ELO does not move.
2. Later viewers: their extension runs the detector on the text already on their screen and sends only the hash and a number, never the text again.
3. Verified when two reports from different installs are within 0.20 of each other, or three or more reports exist (the median counts), or one report comes from a trusted admin install.
4. On verification the slop score is recomputed with the median, the ELO moves exactly once, and phrase counts go up.
5. Each item accepts at most 5 reports.
6. Reputation: a report more than 0.35 from the final median is an outlier. An install with at least 20 reports and 30% or more outliers is ignored from then on.

Decay

Once a night, authors with no new items in the last 7 days move 2% of the way back toward 1000. A dormant 1500 becomes 1490 the next day. Slop streaks are recomputed at the same time: a week with no items breaks the streak.

Guardrails

1. No crawler. The extension only reads what a real person's browser already rendered, and never loads LinkedIn pages in the background.
2. Every badge opens its evidence. There is never a bare number with no explanation.
3. Authors join the leaderboard after 3 scanned posts.
4. Pure Slop requires 3 or more scanned posts.
5. Opt-out and contest forms exist and are linked from every profile.
6. Post text is stored for 30 days at most. The hash, scores and evidence are kept after that.
7. Detector bias is real, so the drift signal and two-install verification stay on. See limitations below.
8. All collection stays inside the user's own session and visible pages.

Limitations

AI text detectors flag writers whose first language is not English more often than native speakers, and the Fakespot model also flags some writing that a human wrote. That is why the detector is never the whole score. We keep the drift signal, which compares people with their own past writing rather than with an idea of “normal” English, and we require two installs to agree before anything counts. Pure Slop needs high confidence on top of that, and every score can be contested.

A slop score is a guess about how a post was written, not a judgment of the person who posted it.

What we collect

Sent to the server

• A post or comment's text, the first time anyone with the extension sees it. Everyone after that sends only its hash.
• The author's public handle, name, headline and avatar URL, and the item's URL.
• A number from the detector on your computer, and a random install id.

Never sent

• Anything about you: no name, no LinkedIn account, no browsing history, no list of what you looked at.

Kept

Post text is deleted after 30 days by a nightly job. The hash, the scores and the evidence chips stay, so a rating can still be explained.

An opt-out is instant and deletes everything stored for that handle. The full privacy notice is on the terms and privacy page.

LinkedIn's terms

LinkedIn's User Agreement prohibits scraping, including with browser add-ons. slop elo is built to stay inside your own session: it only reads posts LinkedIn already showed you, never opens profiles, never scrolls for you and never fetches LinkedIn pages in the background. Using it is still your decision. slop elo is an independent project and is not affiliated with or endorsed by LinkedIn.

Install

The Chrome Web Store listing is on its way. Until it is live, you can load the extension unpacked: build it from the repository, open chrome://extensions, turn on Developer mode, choose Load unpacked and pick the extension's build folder. The README has the exact steps.

On first run the detector downloads once (about 125 MB) and is cached. Until it is ready the extension only looks up existing scores.