D-slop

Highlights or hides AI-generated content on web pages.

D-slop detects AI-generated content and gets out of your way. Much of what comes up in search results these days was written by a language model. Some of it is fine. Most of it is a waste of time, and the frustrating part is that you usually can't tell until you're already three paragraphs in. D-slop flags suspicious content automatically, on every page you visit, so you know what you're reading before you commit to it. How it scores text: The extension runs five signals against each block of text on a page. The most obvious one is phrase matching. Language models reach for the same vocabulary over and over: "delve into," "it's worth noting," "the landscape," "robust," "moving forward," and a few dozen others. Any single phrase could appear in human writing. What's unusual is the rate. These phrases cluster in generated text at frequencies that don't show up in anything a person actually sat down and wrote. Sentence length is a quieter signal, but often more reliable. People write unevenly. A short observation, then something longer, because the thought required it. AI output tends toward regularity, with paragraphs full of sentences that all fall within a few words of each other in length, and that pattern is detectable even when the vocabulary looks clean. The remaining three signals work on similar logic. Punctuation density catches the flat, even patterns that generated text produces. List uniformity flags bullet points that are eerily parallel in length and structure. And conclusion markers catch phrases like "in summary" and "overall" that signal a model wrapping up rather than a person finishing a thought. Each signal contributes to a combined score. Blocks above your threshold get flagged. How it detects AI-generated images, video, and audio: For media, D-slop checks for C2PA provenance metadata, the open standard that AI image generators can embed in their output to declare a file's origin. When a valid manifest is present, the element gets flagged with the generator name if the file includes one. The practical limit is that most social platforms strip image metadata on upload, so C2PA detection works on directly-hosted images where the metadata survives, not on content pulled through a CDN that discards it. Coverage will improve as platform adoption grows. What happens when something gets flagged: Text and media each have independent controls. You can highlight suspicious text while collapsing AI-generated images, or run just one channel at a time. Highlight mode adds an orange outline and a badge to flagged content and leaves the rest of the page alone, which is useful when you want to stay aware without changing how you read. Collapse folds flagged content down so it's out of your visual path, but a single click expands anything you want to see. Hidden is the aggressive option: a full-page overlay appears with a short countdown, then navigates back. There's a "Show anyway" button if you change your mind. The threshold and mode for each channel are adjustable from the popup. The phrase list stays current automatically: Once a week, an automated pipeline searches for new research on AI writing patterns, runs frequency analysis against a corpus of known-AI and known-human text, and passes any candidates through a validation step before they're added to the list. Phrases that have spread into regular human writing get pruned the same way, so the signal doesn't degrade over time. Updates reach your browser within 24 hours, with nothing to reinstall. The phrase list is served from a private endpoint rather than a public repository, which makes it harder to write around. Who this is for: If you read a lot online and have noticed that something has quietly changed about the quality of what you find, this extension is for you. It's also useful for editors evaluating submissions, researchers who want a quick first-pass signal on a source, and anyone who has grown tired of "delve" appearing in every third paragraph. It handles the obvious cases automatically and doesn't claim to catch everything. Privacy and source: There are no accounts and no telemetry. Text analysis runs locally on whatever is already rendered on your screen. The only outbound requests the extension makes are a periodic fetch of the rules file, around 5KB once every 24 hours, and partial media fetches for C2PA scanning, the first 200KB of each media element, only while media detection is enabled. The extension is MIT-licensed and the source is on GitHub. The scoring logic, the phrase pipeline, and the extension code are all public. If you find a whole category the detector misses, open an issue and describe it.