Urgent.News

What's breaking now, across thousands of outlets.

AI

I built a local heuristic to catch AI sounding product copy. Here is what it actually catches and what it misses.

This week I packaged a small analyzer I had been testing as a Chrome extension: Copy Tell. It checks product descriptions and posts on Gumroad and Substack for the kind of stock phrasing an unedited AI writing tool tends to produce, and shows a small notice if enough of it shows up. No ML model. Pure lexical and stylometric scoring: stock phrase density, symmetric list framing used as filler,…

A Chrome extension called Copy Tell has been developed to identify AI-generated product descriptions on marketplaces like Gumroad and Substack. The tool works by analyzing the text for patterns commonly found in machine-generated content, such as stock phrases, symmetric list framing, and sentence length uniformity. It runs entirely in the browser, without sending any data outside the webpage.

To evaluate its performance, Copy Tell was tested against a dataset of 100 product descriptions - 50 AI-generated and 50 human-written. The results showed 39 true positives, 1 false positive, 49 true negatives, and 11 false negatives. The precision of the tool was strong at 97.5%, meaning that when it flagged a description as potentially AI-generated, it was correct 97.5% of the time. However, the recall was lower at 78%, indicating that the tool missed 22% of the AI-generated samples.

The 11 false negatives shared a common characteristic - they were written in a more straightforward, literal style, lacking the typical stock phrases that the analyzer is designed to catch. This limitation stems from the fact that a rule-based approach like Copy Tell cannot easily incorporate natural language understanding, which would be necessary to improve recall.

One false positive was identified: the README for the axe-core library was flagged primarily due to the presence of a single phrase that a human author also used. While this is a fair miss, it's not a case of the tool being cherry-picked for an easy target.

It's important to note that the test set primarily consisted of AI-generated copy versus real technical documentation, rather than actual marketplace listings. In real-world scenarios, where sellers often use more superlative language, the tool's precision might be lower than the 97.5% reported in the test.

The extension is currently live and free to use at https://chromewebstore.google.com/detail/copy-tell-ai-description/ehpkbnkcpbfkagcfmbofldfnpnmopcbi. The developer encourages users to try it out on their own listings and report any errors, whether false positives or false negatives, to help improve the tool's accuracy.

Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.

Read the original at dev.to →

More in AI

Why this blog exists

Originally published on Loop & Retry — field notes on building LLM agents that survive production. Most writing about LLM agents is either a demo that works once on stage or a thread promising the…

  • Loop & Retry blog fills gap between demos and tech singularity claims
  • Author focuses on realistic agent deployment with unexpected issues
  • Posts demonstrate failing version and subsequent fixes with measured data

Mix and Match: One Agent, Three Clouds, One Protocol

This article provides a step by step comparison of the same research agent built three times, on Google ADK, AWS Strands and Microsoft Agent Framework, all three speaking A2A to one coordinator.

  • Same research agent deployed on Google ADK, AWS Strands, and Microsoft Agent Framework
  • A2A protocol serves as common language for agent interoperability
  • Comparison project identifies differences between three hyperscaler agent frameworks

Model Cascade: making LLM classification cheaper

Many LLM workloads are classification tasks. This can get expensive, and I believe it is going to become more and more important, especially with the proliferation of software factories.

  • Model Cascade uses a cheap proxy model and an expensive oracle model
  • BARGAIN paper shows up to 86% cost reduction on various datasets

More from Sunday 23 August →