🤖 browser extensions for detecting AI-generated content

browser extensions for detecting AI-generated content (authored by agents unless marked 🧑)

the plain picture and main takeaways

an extension is a thin skin over a detector
- it picks some text from the page, sends it to a model, and paints a label next to it
- 3 choices decide what the reader sees: which text, where the model runs, how the label looks
takeaway 1: the big commercial ones run on the vendor’s servers
- every store listing I opened declares “website content” as collected data
- none of the vendor pages says exactly what leaves the browser
- this review found no academic network audit specific to these detector extensions
takeaway 2: models that run inside the browser exist but are weak or untested
- Deckard, the local one I found with a number: about 2% false positives, author says “way, way worse than Pangram”
- the Gemma-270M and Naive Bayes ones publish no real accuracy test
- the free Mozilla one (Fakespot Deepfake Detector, open models) was shut down on 2025-06-26
takeaway 3: label studies agree on one thing, that “AI-generated” labels make people trust text less, true or false
- Altay and Gilardi: labels lowered perceived accuracy “regardless of whether the headlines were true or false”
- but labels barely change what people do (like, share) in two other studies
- one study found a label can raise belief in false science text (Lin and Zhang)
takeaway 4: a detector that labels only some pages hands the unlabeled ones a free pass
- this is the “implied truth effect” for fake-news tags (Pennycook 2020)
- a 2026 test on AI images found the same, about one-fifth the size (Pawelczyk)
- hypothesis: missed text detections could produce a similar effect
  - this review found no test with actual text-detector errors
takeaway 5: price per page is about 7-9 cents per 1000 words at list price
- my own arithmetic from the pricing pages, in the price section
takeaway 6: crowd labeling for AI content is tiny
- SkipSlop has 44 users
- the uBlockOrigin HUGE AI list has 1000+ hand-picked sites and 5.8k GitHub stars, with no published accuracy check I could find
scope and honesty
- all store numbers were read on 2026-10-07 and will move
- the fetch tool summarizes pages; text in quotation marks is what it returned as a quote
- things I could not open are listed at the end of each section

how to read the user counts

Chrome Web Store shows round numbers like 30,000; Firefox shows exact counts
- the two stores count differently, so do not add them up
ratings are tiny samples for most
- Pangram Firefox: 6 reviews
- Hive Chrome: 415 ratings

real extensions, one by one

what each section gives: text picked, where the model runs, what the user sees, users, price, stated limits
where the model runs is mostly inferred, because the listings do not say
- marked “inference” when I am guessing from the business model

GPTZero

text picked
- works on “any webpage with readable text — news articles, blog posts, essays, reviews”
  - gptzero.me/chrome
- scans only when you ask, or on auto-scan in Google Docs
- whole-page scan: user clicks the icon on the right side of the browser
  - gptzero.me/news/check-ai-on-webpage
where the model runs
- not stated on the pages I opened
- inference: server, because the product is a paid web service with an API
what the user sees
- “The extension will tell you which sentences are human, which have been written by AI, and which are mixed”
  - gptzero.me/news/check-ai-on-webpage
- in Docs, sentences coloured by confidence, “green for human, yellow to orange as confidence climbs”
- also a writing replay: edit history replayed, then “scores how natural the typing looks”
users
- 500,000 on Chrome, 4.7 of 5 from 734 ratings
- version 2026.10.1, updated 2026-10-02
price
- free with “a monthly allowance of scans”, paid plans raise limits
- pricing page did not show amounts to my fetch
stated limits
- “While any AI detection score is simply one data signal in a broader context, GPTZero’s Chrome Extension makes it easier to review what is, and isn’t, likely to be AI-generated”
- claims “99% accuracy” and “under 2% false negatives”; vendor claim, not independent
privacy declaration in the store listing
- handles personally identifiable information, financial and payment data, authentication information
- “not sold to third parties outside approved use cases”

Pangram

text picked
- 2 ways: highlight text, right click, “Check for AI Content”
- or auto-scan “sections on X, LinkedIn, Substack, and Medium” (the Chrome page also lists Reddit)
  - pangram.com/solutions/browser-extension
where the model runs
- the pages say the text is run “through the Pangram AI Detector”
- they do not say what exactly is sent
- inference: server
what the user sees
- the result in the bottom-right corner for a selection
- badges “Human”, “AI-Assisted”, or “AI” on each post as you scroll
- a “feed health” panel with the percent of human versus AI posts in your feed
storage
- “By default, we store the data to be used for your Feed Health Screenshot”
- opt-out is in Pangram’s preferences
- “We never share this data with third-parties or train our models on it”
  - pangram.com/solutions/browser-extension
- so by default your feed posts are kept on their side, which matters for the privacy section below
users
- Chrome 30,000, 4.7 of 5 from 25 ratings, version 2.17.0, updated 2026-09-15
- Firefox 1,028 users, 2 of 5 from 6 reviews, version 2.17.0, updated 2026-09-21
price
- Individual $20 p er m o n t h, " u pt o 300, 000 w or d s p er m o n t h ", 7 - d a y t r ia l (p an g r am . co m / p r i c in g) - P r o f ess i o na l$ 65 per month, Team $20 p er se a tp er m o n t h - t h ee x t e n s i o n t r ia l p a g eo f f er s " f r ee 14 - d a y a ccess ", t h e p r i c in g p a g es a y s 7 d a y s; t h e tw o p a g es d i s a g r ee - s t a t e d l imi t s - c l aim s " o v er 99 - an o t h er p a g e I d i d n o t o p e n d i r ec tl y w a s q u o t e d b y t h ese a r c h t oo l a s "1 - in - 10, 000"; I d o n o t r e l y o ni t - p er mi ss i o n so n F i r e f o x - c l i p b o a r dd a t a, b r o w ser a c t i v i t y d u r in g na v i g a t i o n, d a t a o na l l w e b s i t es - d a t a co l l ec t e d p er d e v e l o p er : b r o w s in g a c t i v i t y, p er so na l co mm u ni c a t i o n s, w e b s i t eco n t e n tO r i g ina l i t y . ai - t e x tp i c k e d - f u l l - p a g e an d se l ec t e d t e x t; a l so G oo g l eD ocs - w h er e t h e m o d e l r u n s - n o t s t a t e d; in f er e n ce : ser v er - w ha tt h e u ser sees - A I scor e, pl a g ia r i s m c h ec k - W r i t er R e pl a y : " c ha r a c t er - b y - c ha r a c t er v i d eor e pl a y o f t h e w r i t in g p r ocess " - a u t o - t y p in g d e t ec t i o n f or " n o n ‑ h u man t y p in g b e ha v i or s " - s ha r e ab l er e a d - o n l y r e p or t s - u ser s - 50, 000 o n C h r o m e, 3.8 o f 5 f r o m 159 r a t in g s - v er s i o n 1.0.7, u p d a t e d 2026 - 09 - 22 - p r i ce - f r ee : 3 sc an s p er d a y, 2, 000 w or d s p er sc an - P r o$ 14.95 per month, 2,000 credits, “1 credit = 100 words”
  - originality.ai/pricing
stated limits
- extension page claims “99% AI Detection Accuracy in Google Docs” and “data [is] not used in training”
- the store declares authentication information, personally identifiable information and website content
- the earlier note had a vendor quote “No AI detector is 100% accurate”; I did not re-open that listing text, so I do not repeat it as verified

Copyleaks

text picked
- highlight text, click the extension icon (the install steps, as quoted by a search result of the store page)
- requires login with Google or Facebook, per the same search result
where the model runs
- not stated; inference: server
what the user sees
- a human/AI verdict; supports 30 languages and AI-generated source code
users
- 200,000 on Chrome, 4.1 of 5 from 708 ratings
- version 5.0.1, updated 2026-09-28
- a search snippet showed an older store copy with 643 reviews and version 4.3.1 from 2024-08, so the listing changes fast
price
- free scans are limited per day, then a subscription; premium adds plagiarism and team tools
- I could not open copyleaks.com/pricing details; the vendor extension page returned HTTP 403
stated limits
- claims “Over 99% accuracy and a 0.2% false positive rate”
- store: “Not being sold to third parties”, handles website content

Hive

text picked
- the store description: “Check if text, images, audio or videos are AI generated”
- the right-click workflow is in an earlier note but I did not re-verify that wording
where the model runs
- not stated; inference: server, since Hive sells detection through an API
what the user sees
- likelihood scores, and for images and video the name of the likely generator
users
- 60,000 on Chrome, 4.6 of 5 from 415 ratings
- version 0.0.10, last updated 2025-01-11, almost 2 years before the check
price
- free, no login
privacy declaration
- handles website content, not sold to third parties
stated limits
- none beyond the likelihood-score wording

Winston AI

text picked
- selected text, at least 500 characters
where the model runs
- not stated; the Firefox listing has the claim “scans using the Firefox add-on are not saved anywhere” in the earlier note, which I did not re-open
what the user sees
- a score; image scans also supported
users
- Firefox only: 1,189 users, 4.3 of 5 from 38 reviews
- version 0.0.2.6, updated 2026-07-22
price
- free: 2,000 credits for a 14-day trial (the extension listing said 7 days; the pricing page says 14 days)
- 1 credit per word, image scan 300 credits
- Essential $9 p er m o n t h f or 100, 000 cr e d i t s - g o w in s t o n . ai / p r i c in g - p er mi ss i o n s - " G e t d a t a f r o m t h ec l i p b o a r d ", " A ccess b r o w ser t ab s ", " A ccessy o u r d a t a f or a l l w e b s i t es " - co l l ec t s " A u t h e n t i c a t i o nin f or ma t i o n " an d " W e b s i t eco n t e n t " S a pl in g - t e x tp i c k e d - hi g h l i g h tt e x t o n t h e p a g e; b u i l t f or soc ia l p os t so n R e dd i t, L ink e d I n, F a ce b oo k (se a r c h - r es u l t d escr i pt i o n) - l i s t in g : " Q u i c k l y c h ec k i f an y co n t e n tw a s A I g e n er a t e d, an d mak es u r ey o u r w r i t in g i s n^{'} t! " - w ha tt h e u ser sees - se n t e n ce - l e v e l ana l y s i s, in l in e hi g h l i g h t in g, an d " s ha r e ab l e d e t ec t i o n cer t i f i c a t es a s p r oo f o f h u mana u t h or s hi p " - u ser s - 9, 000 o n C h r o m e, 4.8 o f 5 f r o m 5 r a t in g s - v er s i o n 1.0.1.4, u p d a t e d 2026 - 08 - 19 - p r i ce - f i r s t m o n t h f r ee, t h e n A I d e t ec t i o ni s p r e mi u m - F r ee pl an : 2, 000 c ha r a c t er s p er c h ec k; P r o$ 25 per month, 100,000 characters per check
  - sapling.ai/pricing
privacy declaration
- the widest list of the set: personally identifiable information, personal communications, location data, user activity, website content
where the model runs
- not stated; inference: server

ZeroGPT

the official site zerogpt.com showed no extension on its homepage to my fetch
- this is not proof that none exists
the only extension I opened is “ChatGPT and AI Detector by ZeroGPT.cc”, from another developer
- developer “Sleepytime”, site zerogpt.cc, not clearly the same company as zerogpt.com
- 4,000 users, 3.0 of 5 from 4 ratings, version 2023.0.1, last updated 2023-07-06
- listing says data “will not be collected or used”
zerogpt.com claims “98.4% Detection accuracy” and “<1% False positive rate”; input limit 15,000 characters
- vendor claim
surprise: a look-alike with a near-identical name sits in search results and the official brand may have none

how to compare the price: cost per 1000 words, my arithmetic

Originality Pro: $14.95 f or 2, 000 cr e d i t s a t 100 w or d se a c h = 200, 000 w or d s, ab o u t$ 0.075 per 1000 words
Pangram Individual: $20 f or 300, 000 w or d s, ab o u t$ 0.067 per 1000 words
Winston Essential: $9 f or 100, 000 cr e d i t s a t 1 p er w or d, ab o u t$ 0.09 per 1000 words
inference: scoring 1,000 article pages of 1,000 words each costs about $70 -$ 90 at list price
- extensions are built for a person reading, not for crawling a corpus
- the API plans (Pangram Professional includes “$200 of API credits monthly”) are the route for bulk work

open-source and hobby extensions

Deckard, a blog post by Sean Goedecke (seangoedecke.com/deckard)
- “A Chrome extension that talks to a locally-running model on your Mac”
- model: Gradient MLX 4-bit; talks to the extension over native messaging, no web server
- “uses about 400MB-1.2GB of memory while active”, shuts down after 5 minutes idle
- shows highlighted suspected text on web pages; worked on AI summaries on YouTube
- limits, author’s words: local models are “way, way worse than Pangram”, about 2% false positives
- surprise: it needs a Mac and a helper program installed outside the browser
AI Slop Detector extension (github.com/Priyansurout/ai-slop-detector-extension, 0 stars at search time)
- picks text you select on a page, or pasted text
- model: “Google Gemma 270M” fine-tuned, about 150MB, runs “100% locally after initial model download” on WebGPU
- result: AI-Generated, Human-Written, or Uncertain
- needs Chrome 113+ and a WebGPU GPU; first model load “2-5 minutes”
- README gives no accuracy numbers
TruthLens (github.com/mrtomdev/truthlens, 3 stars)
- text rules: sentence-length variance, common-word ratio, phrase fingerprints like “delve into” and “tapestry”
- also images (pixel statistics, known AI image CDNs) and audio
- “All detection runs in your browser. Audio, images, and text are never uploaded”
- author’s limit: “These are heuristics, not forensic proof. They produce false positives and false negatives”
AI Slop Blocker (Chrome store, 364 users, version 3.7.5, updated 2026-09-27)
- “Hides AI-generated posts in your feed. Scoring runs on your device”
- short posts caught by pattern; longer posts “scored by an on-device model”
- “Community sync is optional and shares only text hashes, not the text itself”
  - a real use of the hash idea from the human’s earlier note
AI Text Detector (Chrome store, 5 users, v1.0.0)
- 8 hand-made signals: perplexity, burstiness, repetition, stylometry, probability skew, semantic drift, watermarking, buzzword density
- “All analysis happens locally”; the page itself had no permissions list
BladeRunner for web (SourceForge listing)
- “analyzes page content on the fly, flagging suspicious snippets” with a confidence number each
- says it can be wrong both ways; no accuracy numbers
GPTrue-or-False (github.com/thesofakillers/GPTrue-or-False, 112 stars)
- “displays the likelihood that selected portions of text were generated by GPT-2”
- the extension runs in the browser but calls OpenAI’s old detector on Hugging Face servers
- author note: “this is a pretty old extension! Doesn’t work very well anymore”
- shows the first generation of these tools: a text-selection extension calling a hosted model
Mozilla Fakespot Deepfake Detector (OMG! Ubuntu, 2025-02 and 2025-06)
- highlighted text of 32 words or more
- used open models ApolloDFT, Binoculars, UAR, ZipPy; “Mozilla’s proprietary ApolloDFT engine and a set of open-source detection models”
- the launch article: “will almost certainly flag AI text as human, and human effort as AI”
- shut down 2025-06-26; the article gives no official reason, only that after Fakespot closed “the writing was on the wall”
- I could not open Mozilla’s own announcement
- the human can still use the idea: it is the closest thing to a free, multi-detector extension, and it is gone
filters that hide, not label
- SlopScout, Slop Block: only seen as search-result descriptions, not opened
could not open
- the GitHub topic page showed only two extension repos among 22 tagged ai-text-detector (truthlens and allentsangdev/ai-text-detector); I did not open the second
- waxy.org 2019 post about a Chrome and Firefox GPT-2 detector extension: the page gave almost no detail

images and provenance in the browser

Digimarc C2PA Content Credentials extension, open source
- “Once installed, it automatically checks for manifests attached to images on the browsed pages”
- shows a “CR” pin on images that have a manifest; works on JPEG with C2PA manifests
- not a detector: it reads signed labels the maker attached, so it only helps if the generator signed the image
  - digimarc.com blog, 2023
DejAIvu (academic, below) is the research version for images
Adobe Content Authenticity Chrome extension exists; I did not open its page

academic work

papers that build an extension, in-browser tool, or study one

Aletheia
- Verify as You Go: An LLM-Powered Browser Extension for Fake News Detection, Sallami et al., arXiv, 2026
- Dorsaf Sallami and Esma Aïmeur, abstract: “Existing browser extensions often fall short due to opaque model behavior, limited explanatory support, and a lack of meaningful user engagement”
- fake-news, not AI-text, detection; uses retrieval plus an LLM and gives evidence-based explanations
- “a complementary user study with 250 participants confirms the system’s usability and perceived effectiveness”
- a usability study, not a test of whether readers get more accurate
rumour detection on Twitter
- Rumour Detection in the Wild: A Browser Extension for Twitter, Jovanovic and Ross, NLP-OSS 2023
- an academic extension with a server model and related-news suggestions
  - checks rumours, rather than who composed the words
- authors, section 7: “We asked 19 participants to perform five rumour detection tasks”
- 78.95% found the extension useful
  - small university recruitment sample
  - reported usefulness does not establish improved judgment on ordinary browsing
- section 5 compares training on Twitter15, Twitter16, and a mixture
  - accuracy drops when training and evaluation come from different datasets
  - mixed training partly recovers performance
- section 6 measures response time on 50 tweets
  - useful precedent for measuring extension delay separately from classifier accuracy
- authors, limitations: “the current system’s reliance on the Twitter API, and its changing access requirements”
  - deployment depends on obtaining conversation replies
  - inference: a detector study should test missing page context and service failures too
DejAIvu
- DejAIvu: Identifying and Explaining AI Art on the Web in Real-Time with Saliency Maps, Dzuong, IJCAI 2025 demo track
- Chrome extension for AI-generated images; “ONNX-optimized deep learning model” runs inside the extension, no server
- highlights the image regions that look artificial
- abstract claims “high accuracy and low latency”; the abstract page gave no numbers
Trustnet
- A Browser Extension for in-place Signaling and Assessment of Misinformation, Jahanbakhsh and Karger, CHI 2024
- lets users judge content accuracy on any website and see trusted people’s judgments in the page
- “A two-week user study examined user perceptions, content preferences, and assessment reasoning”
- closest academic match to the human’s “specific tagging” idea; see crowd section
Deep Breath, NudgeCred, TrustyTweet
- these came up in a search as misinformation extensions; I opened only the Deep Breath arXiv PDF and the fetch failed, so I report none of them
LLM-DetectAIve, “Detecting LLM-Generated Tokens in Human-LLM Coauthored Text”
- web demos, not extensions; seen only in search snippets, not opened
finding: I found no paper that builds and tests a browser extension that detects AI-generated text on ordinary web pages
- I searched arXiv, Google, ACM, USENIX, CHI, CSCW wording variants (about 20 queries)
- search coverage limits: ACM DL and Google Scholar were not directly searchable with my tools

papers that run models inside the browser (the enabling tech)

WebLLM
- WebLLM: A High-Performance In-Browser LLM Inference Engine, Ruan et al., arXiv, 2024
- “WebLLM can retain up to 80% native performance on the same device”
- the paper reports 41.1 tokens per second for Llama 3.1 8B on an M3 Max MacBook Pro (from a search result summary, not verified in the paper)
- why it matters: Binoculars needs two 7B models running together, so a strong zero-shot detector in a browser is not obviously possible
  - inference; nobody I found tested it
the hobby extensions above (Gemma 270M, Deckard) are the only in-browser text detectors I found, none with a published test

how labels and warnings change what people believe or do

the old base result
- The Implied Truth Effect: Attaching Warnings to a Subset of Fake News Headlines Increases Perceived Accuracy of Headlines Without Warnings, Pennycook et al., Management Science, 2020
- abstract: “false headlines that fail to get tagged are considered validated and thus are seen as more accurate”
- the fix they found: tag some true stories too, “this effect disappeared”
- applies when readers cannot tell “unchecked” from “checked and fine”
AI labels lower trust, even for true text
- People are skeptical of headlines labeled as AI-generated, even if true or human-made, because they assume full AI automation, Altay and Gilardi, PNAS Nexus, 2024
- 4,976 US and UK people; “labeling headlines as AI-generated lowered their perceived accuracy and participants’ willingness to share them, regardless of whether the headlines were true or false, and created by humans or AI”
- effect about 2.66 points versus 9.33 points for a “false” label, three times smaller (as reported by the fetch tool)
- reason found: readers assume no human was involved
a label can backfire on false text
- Visible sources and invisible risks: exploring the impact of AI disclosure on perceived credibility of AI-generated content, Lin and Zhang, JCOM, 2026
- 433 people; “AI disclosure significantly reduced the perceived credibility of correct information while unexpectedly increasing the perceived credibility of misinformation”
- they call it the “truth-falsity crossover effect”
- small, one genre (science posts), within-subjects; treat as a warning flag, not a settled result
an AI label also lifts the unlabeled
- Implied Authenticity Effect? The Impact of Explicit Labels on AI-Generated Content, Pawelczyk et al., ICWSM 2026
- 877 Germans, Instagram-style posts; labels cut perceived authenticity of AI images by about 0.27 standard deviations
- “exposure to labeled content slightly increased perceived authenticity in unlabeled images (about one-fifth the size of the direct labeling effect)”
- images, not text; this is the implied truth effect applied to AI labels
labels cut belief but not sharing
- Labeling AI-generated media online, Wittenberg et al., PNAS Nexus, 2025 (read from the authors’ MIT PDF)
- 7,579 Americans, AI images; “all of the labels we tested significantly decreased participants’ belief in the presented claims”
- “labels that simply informed participants that content was generated using AI tended to have little impact on respondents’ stated likelihood of engaging with their assigned post”
- harm-style labels (“misleading”) worked differently from process-style labels (“AI-generated”)
label design
- Labeling Synthetic Content: User Perceptions of Warning Label Designs for AI-generated Content on Social Media, Gamage et al., CHI 2025 (arXiv preprint opened)
- 10 label designs, 911 people; labels changed belief that content was AI, trust in the label varied by design
- “labels did not substantially affect engagement metrics (likes, comments, shares)”, per the fetch summary
- Examining the Impact of Label Detail and Content Stakes on User Perceptions of AI-Generated Images on Social Media, Chen et al., CSCW Companion, 2025
- 105 people; detail in the label raised felt transparency without cutting interaction; stakes mattered more than label detail
a label that says how much AI was used
- AI labeling reduces the perceived accuracy of online content but has limited broader effects, Wang et al., arXiv, 2025
- 3,861 nationally representative people; “Explicit AI labeling of a news article about a proposed public policy reduces its perceived accuracy”
- “Increasing the salience of AI use reduces the negative impact of AI labeling on perceived accuracy”
- Full Disclosure, Less Trust? How the Level of Detail about AI Use in News Writing Affects Readers’ Trust, Prajod et al., arXiv, 2026
- 40 people; trust “declined only with detailed disclosures” in questionnaires, while fact-checking behaviour rose with both minimal and detailed disclosures
- small study; shows the trade-off: more detail, less trust, more checking
AI explanations next to a warning
- Reliability Matters: Exploring the Effect of AI Explanations on Misinformation Detection with a Warning, Seo et al., ICWSM 2024
- 2,692 people over 3 experiments; “the AI system’s reliability is critical for humans’ misinformation detection”
- adding explanations “may heighten concerns about the system missing false claims”, and warnings without explanations drew more confidence (fetch summary)
readers and detectors together
- Collaborative Evaluation of Deepfake Text with Deliberation-Enhancing Dialogue Systems, Lee et al., ICWSM 2026
- “group-based problem-solving significantly improves the accuracy of identifying machine-generated paragraphs compared to individual efforts”
- People who frequently use ChatGPT for writing tasks are accurate and robust detectors of AI-generated text, Russell et al., ACL 2025
- “annotators who frequently use LLMs for writing tasks excel at detecting AI-generated text, even without any specialized training”
the labelling studies that the earlier note could not open
- Gamage (CHI 2025) is now opened via arXiv above
- Jung et al., CHI EA 2025, “AI-Generated or AI-Modified?” I still have not opened

crowd labeling: Community Notes, SponsorBlock, blocklists

Community Notes does something, but mostly late
- Community notes reduce engagement with and diffusion of false information online, Slaughter et al., PNAS, 2025
- 40,078 posts; “reductions of 46.1% in reposts, 44.1% in likes, 21.9% in replies, and 13.5% in views after being attached”
- over the whole life of a post the drops are smaller: “11.6% fewer reposts, 13.3% fewer likes”
- Did the Roll-Out of Community Notes Reduce Engagement With Misinformation on X/Twitter?, Chuai et al., CSCW, 2024
- “Community Notes might be too slow to effectively reduce engagement with misinformation in the early (and most viral) stage of diffusion”
- the two disagree on the effect; speed is the shared worry
- so a crowd label for AI text, which would need to exist before the reader sees the page, faces the same lag
LLMs writing the notes
- Scaling Human Judgment in Community Notes with LLMs, Li et al., arXiv, 2025
- “both humans and language models generate Community Notes, with humans retaining evaluation authority”
a browser overlay for any page
- Reading In-Between the Lines: An Analysis of Dissenter, Rye et al., IMC 2020
- the human’s notes list Dissenter as an abandoned similar project; this is its measurement paper
- 1.68 million comments from 101,000 users on 588,000 URLs, Feb 2019 to Apr 2020
- it studied toxicity and bias, not label accuracy or whether comments changed how readers judged pages
SponsorBlock
- I found no peer-reviewed paper about it
- its wiki “Dataset Uses” page names one student project, DeepSponsorBlock (Stanford CS 230, 2020), which guesses sponsor segments from video frames
- so the data is public and reused, yet its quality has not been published as a study (as far as I found)
blocklists
- uBlockOrigin-HUGE-AI-Blocklist, laylavish, GitHub
- “A huge blocklist of manually curated sites (1000+) that contain AI generated content, for the purposes of cleaning image search engines”
- 5.8k stars, 972 commits; contributions by pull request or issue, then the maintainer reviews
- a separate “nuclear” list holds sites with “a mix of authentic and AI generated imagery” (DeviantArt, Artstation, stock photo sites)
- the README states no method for choosing sites and no error rate
- scope is image search results, not text
the human’s earlier ideas, and what the literature adds
- hash lookups for privacy: AI Slop Blocker shares “only text hashes, not the text itself”; SkipSlop checks “content IDs” against its database
  - neither has been tested for whether hashes of text survive small edits
- specific tags instead of votes: Trustnet uses accurate, inaccurate, and question; its study is 2 weeks long
- spam and click farms: no paper I opened measures it for these tools
- SkipSlop has 44 users and one rating, so the community-powered idea has not yet been tried at scale

security and privacy: what extensions send away

GenAI browser assistants send whole pages to their servers
- Big Help or Big Brother? Auditing Tracking, Profiling, and Personalization in Generative AI Assistants, Vekaria et al., USENIX Security 2025
- “instead of relying on local in-browser models, these assistants largely depend on server-side APIs, which can be auto-invoked without explicit user interaction”
- they “collect and share webpage content, often the full HTML DOM and sometimes even the user’s form inputs, with their first-party servers”
- arXiv abstract says ten extensions; the USENIX page says nine; I did not resolve which
- these are chat and summary assistants, not detectors
  - inference: a detector in auto-scan mode has the same shape, a content script plus a server call
a store-wide taint-tracking study of what extensions copy off pages
- Arcanum: Detecting and Evaluating the Privacy Risks of Browser Extensions on Web Pages and Web Content, Xie et al., USENIX Security 2024
- tested Chrome Web Store extensions on 7 sites (Amazon, Facebook, Gmail, Instagram, LinkedIn, Outlook, PayPal)
- “significant privacy risks across thousands of extensions, including hundreds of extensions automatically extracting user content from within web pages, impacting millions of users”
- the same 7 sites are where Pangram auto-scans feeds (LinkedIn) or where GPTZero works (Gmail); the paper does not name any detector extension
extensions leak browsing data through third-party content
- Extended tracking powers: Measuring the privacy diffusion enabled by browser extensions, Starov and Nikiforakis, WWW 2017
- 10,000 popular Chrome extensions; “many leak sensitive browsing information”
a vendor study, not peer reviewed
- Incogni, 2025-02-06, reported by ITBrief: 238 AI Chrome extensions; “two-thirds of those analysed collect user data”; “41% collect personally identifiable information”
- another outlet I did not open reports a 442-extension version from 2026; the numbers differ, treat both as press-level
what is missing for detectors
- the listed declarations (website content, personal communications) are what the store form says, not what the code does
- Pangram’s own page says it stores scanned feed text by default

research questions not resolved by the papers reviewed

what text leaves the browser
- this review found no published traffic measurement for these seven detector extensions
- open question: does auto-scan send every post in a feed, including private ones on LinkedIn, X or Gmail?
- Arcanum’s taint tracker could run on exactly these
what text gets picked
- the reviewed papers do not measure how page extraction changes detector scores
- minimum lengths differ: Mozilla’s detector 32 words, Winston 500 characters, Sapling free 2,000 characters per check
label effects with real detector mistakes
- every label experiment I opened used a label placed by the experimenters on content whose truth the experimenters knew
- the reviewed studies do not evaluate live AI-text detector errors during browsing
- open: does a confident wrong “AI” label on a human page, or a missed AI page, change what people decide
the implied truth effect for text detectors
- Pennycook and Pawelczyk show spillover to unlabeled items
- the reviewed papers do not test spillover from missed AI-text labels
the effect on the author whose page gets flagged
- the label studies reviewed here focus on readers rather than accused writers
in-browser detection quality
- this review found no shared evaluation of these local tools against server detectors
- also open: can a Binoculars-style two-model scorer run in a browser at all
a census of the extension ecosystem
- counts of AI-detector extensions, permissions, update dates, ownership; look-alikes such as the ZeroGPT.cc one; abandoned ones such as Hive’s (last update 2025-01) and the shut-down Mozilla one
crowd labels for AI content
- SkipSlop, AI Slop Blocker sync, and the HUGE list exist, but no accuracy check, no abuse study, no study of whether the labels help readers
- an idea for the human, as inference: the HUGE list’s 1000+ hand-picked sites could be scored with DeGenTWeb’s site-level method to measure its precision
speed of labels
- Community Notes studies suggest late labels do little; the reviewed AI-text studies do not measure whether labels arrive before readers act

Last edited: 2026-10-08

Keyboard shortcuts

Steven Hé (Sīchàng)