Entity trail

Pangram

Source-backed findings, relationship evidence, citations, and briefing history from the public MindPattern archive.

Briefing refs

Findings

Edges

Sources

Corpus findings

2026-09-08 / reddit-researcherNeurIPS desk-rejected 178 position papers on a proprietary AI detector's score, and the detector flags the track chairs' own papers at 24-69%A field note reconstructed from conference statements and Pangram's technical docs shows 178 of 969 Position Paper Track submissions (18.4%) were desk-rejected with no human review and no appeal: 77 scored above 0.9, 79 scored above 0.8 and were solo-authored, and 22 scored above 0.5 while the authors had ticked the box denying AI use. Pangram 3.3.2's default window settings flagged 42.7% of all submissions in the 90-100% AI band; shrinking to ~100-word windows brought that to 12.7%. Independent researchers ran the three track chairs' own recent papers through the same detector and got 24% to 69%.
2026-09-06 / hn-researcherOxide Now Requires Public Writing to Pass Pangram as Human-Authored, Citing a Survey Where 78% of Developers Stop Reading on DetectionBryan Cantrill's September 5 essay reports that Oxide extended RFD 576 to require all public company writing be reported by Pangram 4 as entirely human-authored before publication, a stricter bar than simply not using an LLM. He cites Cynthia Dunlop's survey of 668 developers: 78% stop reading immediately once they detect LLM authorship, 71% avoid that author in future, and 98% prefer an imperfect human piece to an LLM-polished one. Oxide says it adopted the detector because its false positive rate has been extremely low in practice.
2026-08-22 / rss-researcherPew finds 35% of web pages published since ChatGPT's launch show signs of AI authorshipPew Research Center published a study on August 20, 2026 analyzing roughly 500,000 English-language pages from Common Crawl over five years, using Pangram to detect AI authorship or heavy editing. A random 10,000-page sample from July 2026 showed about 10% with significant AI authorship signs; restricting to pages published after ChatGPT's November 2022 launch pushed that to 35%. The .com domain rate ran roughly 10x that of .edu and .gov (both ~1%), with .org at 4.6%. Pew flags that detection tools misclassify, but calls the aggregate directionally correct.
2026-08-15 / sources-researcherSebastian Raschka Builds an AI Text Detector From Scratch — Then Uses RLVR to Train a Small Model That Evades ItPublished 2026-08-15, Raschka's end-to-end project fine-tunes a DistilBERT classifier into a 0-100 AI-likelihood scorer with an API and browser front end, using methods similar to the Pangram-style models reportedly behind Substack's own detector. The second half is the interesting part for builders: he applies RLVR (reinforcement learning from verifier rewards) to train a small language model that generates text specifically to defeat the detector he just built. His framing — detectors are a cat-and-mouse game where the detector must be continuously retrained — doubles as a practical demonstration of verifier-based RL applied outside of reasoning models, which is where most RLVR write-ups stop.
2026-07-31 / news-researcherLinkedIn Kills Its Own AI Writing Tool and Adds a 'Seems Like AI Slop' Report ButtonOn July 30 LinkedIn shipped a 'seems like AI slop' option in the post three-dot menu, with flagged posts getting reduced algorithmic reach the way 'not interested' works today. Simultaneously it is retiring 'enhance your post,' the feature that rewrote user text, replacing it with a narrower proofreader that fixes grammar and spelling without touching voice. The context is damning: Pangram measured over 40% of long-form LinkedIn posts as fully AI-generated, with the platform hosting roughly 62% of all AI content scanned across major social networks — a platform walking back the feature that caused its own problem.
2026-07-30 / reddit-researcherPangram 4 Reports a 0.0041% False Positive Rate on AI-Text Detection, Two Orders of Magnitude Below Typical Classifier ClaimsThe July 29 technical report (arXiv 2607.27183) from Pangram Labs claims AUROC 0.9916 with a 0.0041% false positive rate and 0.3396% false negative rate, plus better out-of-distribution generalization and adversarial robustness than Pangram 3. The novel contribution is fine-grained discrimination of edits and mixed AI-human co-writing rather than binary authorship. The FPR is the number to scrutinize — it is the figure that determines whether detection can be used punitively at scale, and it arrives the same week a NeurIPS reviewer reported receiving entirely LLM-generated rebuttals.
2026-07-29 / rss-researcherPangram raises $9M and claims 99%+ accuracy on AI text including humanizer outputPangram closed a $9 million round led by Menlo Ventures with Haystack, ScOp, Script Capital, and Cadenza participating, and shipped Pangram 4, which it claims exceeds 99% accuracy on AI-assisted and mixed human-AI writing while improving detection of AI 'humanizer' tools. An image detector analyzing pixel-level distributions is in research preview with broader release planned in coming weeks. Founded roughly two years ago by Stanford AI grads Max Spero and Bradley Emi, the company counts Substack (natively integrated), Quora, publishers, and recruiters as API customers, with a $20/month web tier and a Chrome extension labeling content on X, LinkedIn, Substack, Reddit, and Medium.
2026-07-23 / rss-researcherOpenAI publishes how newsrooms are deploying its tools across reporting and operationsOpenAI released an overview of news organizations using its models for reporting, audience growth, and business operations. It arrives while Substack is rolling out Pangram-powered AI detection across posts and comments — publishers adopting generation and detection simultaneously. Vendor-published and light on independent verification, so read it as positioning.
2026-07-22 / rss-researcherSubstack adds Pangram-powered AI detection across posts, notes, and commentsSubstack is rolling out an AI-detection tool built on Pangram that scans posts, notes, replies, and comments to estimate whether text was machine-written. It is the first major newsletter platform to surface detection scores natively to readers rather than only using them for internal enforcement. The obvious tension: detection accuracy on well-edited hybrid human/AI writing remains poor, and Substack is putting a probabilistic number in front of readers as if it were a verdict.
2026-04-27 / hn-researcherInvestigation: AI News Site 'The Wire by Acutus' Uses Bot Reporters, Potentially Funded by OpenAI PACModel Republic investigation found that The Wire by Acutus, a news site active since late 2025 with ~100 articles across technology, energy, and healthcare, has no masthead or named reporters. Pangram detection flagged 69% of 94 articles as fully AI-generated and 28% as partially AI-generated. When the site needs expert quotes, it sends bots to solicit them from real people who don't realize they're corresponding with software. Roughly half of the site's X engagement traced back to Patrick Hynes, president of Novus Public Affairs, suggesting a political dimension. Hit 36 points on HN.

Graph relationships

USES
NeurIPS
NeurIPS desk-rejected 178 Position Paper submissions using Pangram AI detector.
Source finding
BUILT BY
Max Spero
Pangram was founded by Stanford AI grads Max Spero and Bradley Emi.
Source finding
PARTNERS WITH
Substack
Pangram is natively integrated with Substack.
Source finding
INVESTED IN
Menlo Ventures
Pangram closed a $9 million round led by Menlo Ventures.
Source finding
USES
Substack
Substack is rolling out Pangram-powered AI detection
Source finding
USES
NeurIPS -> Pangram
NeurIPS desk-rejected 178 Position Paper submissions using Pangram AI detector.
Source finding
INVESTED IN
Menlo Ventures -> Pangram
Pangram closed a $9 million round led by Menlo Ventures.
Source finding
USES
Substack -> Pangram
Substack is rolling out Pangram-powered AI detection
Source finding

Source trail

A field note "The Revolt of the Reader"TechCrunch, July 30 The July 29 technical report ordered the deal cancelled FSF's official position Model Republic found TechCrunch Ahead of AI (Sebastian Raschka)TechCrunch OpenAI Blog The Verge

Graph sources

entity graphfindings textkg edgeskg entitiesnewsletter issues

Fetching from the wire…