Put on the glasses.

Named for the glasses in They Live.

In John Carpenter's 1988 film, a pair of sunglasses revealed the hidden messages embedded in ordinary reality — commands embedded in advertisements, billboards, dollar bills. Once you put them on, you could not unsee what was there.

The Hoffman Browser works the same way. It reads the page you are on the way an expert reads manipulation — not word by word, but all at once, understanding what the language is designed to do to you. When it finds a manipulation technique, it shows you the exact words being used, names the technique, and explains its purpose.

It does not tell you what to think. It does not block content. It does not decide what you can read. It shows you what is being done to you, and trusts you to think for yourself.

• ◈ Local AI reads the full rendered page text and identifies manipulation techniques — outrage engineering, war framing, false authority, tribal activation, and more
• ◈ Quotes the exact language being used against you and explains the mechanism — why the technique works and what psychological response it is designed to produce
• ◈ "Why is this here?" — shows who owns the site, their documented business model, revenue sources, and cases of documented harm linked to their platform
• ◈ Reads rendered screen text — not DOM structure, which platforms can manipulate or obfuscate
• ◈ Works on every website — news, social media, blogs, political content, advertising. Any page you can visit, the browser can analyze.
• ◈ Zero data retention. All processing local. The website you are reading never knows it is being analyzed.

Nine documented manipulation techniques.

The local AI model is trained to identify nine categories of behavioral manipulation technique commonly deployed in media content, news, political advertising, and social platform feeds. Each detection returns the technique name, a severity rating, the specific quoted text triggering the flag, and an explanation of why it qualifies.

outrage_engineering — Content calibrated for maximum emotional arousal rather than information. Language chosen to provoke visceral anger, moral disgust, or fear, because outrage is the emotion most reliably associated with share behavior and continued engagement.

war_framing — Framing of political, social, or diplomatic situations as armed conflict, invasion, or battle. Escalates perceived stakes and signals to readers which "side" they should be on.

false_authority — Invocation of credentials, institutional affiliation, or consensus ("experts agree," "studies show") without specific, verifiable sources. Borrows epistemic legitimacy the content has not earned.

tribal_activation — Language that signals group identity and invites readers to interpret content through the lens of in-group/out-group dynamics. Reduces complex issues to team membership.

enemy_framing — Identifies a specific person, group, or institution as an adversary whose actions are the cause of harm to the reader or their group. Personalizes structural problems as individual villain narratives.

Local AI. No internet connection required for analysis.

When you click "Analyze" on any page, the analysis pipeline runs entirely on your device. The AI model — a 2.2GB language model stored locally — never communicates with external servers. The page you are reading never receives any signal that analysis is happening.

The "Why is this here?" button on each flag queries the Behavioral Manipulation Intelligence Database (BMID) — a locally-running database of documented platform conduct, motives, and harm cases. If you run the BMID server alongside the browser, every flag card shows you the documented ownership, business model, and harm record of the site you are reading.

Nothing leaves your device. Nothing.

We built it this way deliberately. A surveillance tool that surveils its users would be no different from the systems it is designed to expose.

Common questions.

• What hardware do I need?
  Any modern Windows PC with at least 4GB of RAM. The AI model requires approximately 2.2GB of storage. Analysis runs on the CPU — no GPU required. On a modern CPU, analysis typically completes in 15–45 seconds depending on page length.
• What happens the first time I analyze a page?
  On first analysis, the browser downloads the AI model (~2.2GB) from HuggingFace. This is a one-time download. After that, all analysis runs fully offline. The model is stored in your user data directory and persists between sessions.
• Does the website I'm reading know I'm analyzing it?
  No. The browser reads the rendered text of the page the same way a screen reader or print-to-PDF function would. No network request is made to the page's server during analysis. The page receives no signal that analysis is occurring.
• How accurate is the detection?
  The model — Llama 3.2 3B Instruct, a 3 billion parameter language model — is capable but not infallible. It will miss techniques and occasionally flag language that is emphatic but not manipulative. The purpose is not to produce a verdict but to make the techniques visible. You are the judge of what you see.
• Does it work on paywalled or logged-in content?
  Yes. The browser is a full Chromium browser. You can log into any site and the analysis panel reads whatever text is rendered on screen — including content behind a paywall or login, because you are already logged in.
• What is the BMID?
  The Behavioral Manipulation Intelligence Database is a locally-running database of documented platform conduct — fishermen (platforms), their motives, their documented harms, and the evidence supporting each claim. When the BMID server is running alongside the browser, the "Why is this here?" button on any flag card shows you the documented record for the site you are reading. The BMID is open source and available at the same GitHub repository.
• Can I contribute to the project?
  Yes. The browser and BMID are fully open source under MIT and CC BY 4.0 licenses respectively. The GitHub repository welcomes contributions — particularly OCR support for reading text embedded in images, expanded platform data for the BMID, and ports for Mac and Linux.