There is a quiet arms race happening on Hacker News. As AI-generated comments flood discussions, the community wants to know which posts are real and which are slop. A new open-source tool called hnslop makes that question answerable at scale.
hnslop is a cached JSON API built on top of Salah Adawi's Hacker News AI Detector. It takes every frontpage Hacker News post and runs it through Pangram, an AI text detection model, then serves the results as a simple HTTP endpoint. The hosted instance at hnslop.nilsherzig.com lets anyone query AI scores for any HN post without running their own infrastructure.
What the API actually does
The tool is straightforward. Send a request with one or more Hacker News post IDs, get back a score between 0 and 100. A score of 33 suggests the text is likely human-written. A score of 99 flags it as almost certainly AI-generated. Each response includes a link to Adawi's detector page for that specific post, where you can see the full analysis.
The API supports two formats: batch queries for multiple posts in one call, and single-post lookups. Responses include a cache status field, hit or miss, so callers know whether they are getting a fresh result or a stored one. This caching layer matters because HN traffic patterns are predictable. Frontpage posts get most of their engagement in the first few hours, and those are exactly the posts people want to check.
The Pangram model behind it
Adawi's detector uses Pangram v3.3, which is not the latest version of the model. Pangram is a commercial AI text detection service, and the cost of running detection at HN scale is nontrivial. The hnslop README gives a nod to Adawi's likely budget for Pangram API calls, and the caching strategy is clearly designed to minimize those costs while still keeping results reasonably current.
Pangram works by analyzing statistical patterns in text that are characteristic of language model output. It does not look for specific phrases or watermarks. Instead it measures how closely a piece of text matches the probability distributions that LLMs produce when generating tokens. The 0 to 100 score represents a confidence estimate that the text was machine-generated.
Why this matters for the HN community
Hacker News has an authenticity problem. The site's voting and comment ranking systems assume human participants. When AI-generated posts reach the frontpage, they distort the signal the community relies on. Adawi's detector showed this was possible to measure, but checking individual posts by hand does not scale. hnslop solves that by making the data programmatic.
The tool also creates a dataset. By caching results for frontpage posts, hnslop builds a running record of AI-generated content on HN over time. Researchers, journalists, and community moderators could use that data to study trends in AI slop, measure the effectiveness of detection models, or build filters that surface or suppress likely-AI content.
For developers building HN readers, aggregators, or moderation tools, the API is a drop-in signal. You can add an AI score to any post display, flag suspicious comments, or sort results by likelihood of being human-written. The HTTP interface is simple enough that integrating it takes minutes, not hours.
The limits of detection
No AI text detector is perfect, and Pangram v3.3 is an older model. False positives will happen. Short posts are harder to classify than long ones. Text that has been edited by a human after AI generation can confuse the model. And as language models improve, detection models need to keep pace, which is why the gap between v3.3 and Pangram's latest version matters.
hnslop does not try to solve the detection problem. It makes an existing detector accessible. The value is in the plumbing, not the model. By wrapping Pangram in a cached API and indexing it against HN post IDs, the tool turns a manual research project into a queryable service. That is the kind of infrastructure that lets a community respond to a problem rather than just complain about it.