Pangram helps you hide AI garbage on Hacker News

If you are tired of your feed being clogged by "10 Ways to Optimize Your Workflow" written by a bot, this Firefox extension and API uses Pangram scores to flag LLM-generated content on the Hacker News front page. It doesn't filter articles about AI—it filters articles written by AI. You can set it to automatically hide posts with high probability scores, effectively scrubbing the fluff from your view.

How to actually use it

Since this is a browser extension and a userscript, you aren't installing some heavy enterprise software. You just add it to Firefox, and it starts tagging the HN front page with scores.

1. Install the extension or the userscript in your browser.
2. Refresh the Hacker News homepage.
3. Look for the Pangram scores next to the threads.
4. Configure the "hide" threshold in the settings to automatically collapse any article that hits a specific AI-probability percentage.

If you're a developer and don't want to rely on a browser extension, there is a public API available. You can pull the scores for the current front page and pipe them into your own custom RSS reader or a CLI tool to filter out the noise before you even open a browser.

Is it actually useful or just another gimmick?

Most "AI detectors" are a joke because they rely on vague "perplexity" metrics that flag non-native English speakers as bots. Pangram is slightly more honest about its scoring, but it's still a probability game.

• The Win: It's great for high-volume skimming. When a specific topic trends on HN, 40% of the submissions are usually low-effort AI summaries. This hides those instantly.
• The Fail: It will occasionally hide a very dry, technical human who writes like a manual. Also, if you use Chrome, you're out of luck unless you find a compatible userscript manager, as the primary focus here is Firefox.

The technical trade-off

The project relies on data provided by user salahadawi to identify patterns. Because it's an API-driven extension, there is a slight delay between the page loading and the scores appearing. It's not an instant "blink and it's gone" experience; the scores pop in a second after the HTML renders.

If you are building your own tool using the API, be prepared for the fact that "AI-written" is a moving target. As models get better at mimicking human quirkiness, these scores will fluctuate. But for now, it's a decent way to avoid clicking on a link only to realize three sentences in that you're reading a GPT-4o hallucination.