Sean Goedecke, a software engineer who writes about AI tooling on his own site, built a Chrome extension called Deckard that scans web pages for AI-generated text using a model running entirely on his laptop. He shared the project, along with benchmark numbers, in a post on his own blog.
Goedecke’s stated motivation is a workflow gap. Pangram, the detection service he already trusts, only helps once a reader is suspicious enough to paste text in and check. He wanted something scanning continuously in the background, without sending every page he visits to a third-party API and without a per-check bill.
That framing points to a real difference in what gets asked. A tool you open when something smells synthetic answers “is this one paragraph AI.” A detector running the whole time you browse answers a different question: what share of everything crossing your screen today came from a model. The second question is the one likely to change how someone reads, because it produces a running signal instead of a single verdict.
Building that meant giving up accuracy. Goedecke benchmarked eight small, open detection models against combined AI-detection datasets and picked the strongest tradeoff for the extension: a 4-bit variant of a model called Gradient, run through MLX. By his own numbers, it flagged 52.35 percent of AI-involved text while misflagging 2.712 percent of human writing as machine-generated. He compares that against Pangram’s own claimed figures of a 99.66 percent detection rate and a 0.004 percent false-positive rate, drawn from the company’s published technical writeup rather than independent testing. Every model he tested with a lower false-positive rate also caught noticeably less AI text, and he did not test Pangram’s own larger EditLens model because it is too heavy to keep running continuously on a laptop.
Goedecke is explicit that this is closer to a shipped prototype than a validated tool. His benchmark draws on datasets he assembled himself, and he offers no confidence interval or third-party check on the false-positive rate. A roughly 2.7 percent chance of tagging genuine human writing as AI is high enough that a single flag from Deckard should read as a hint, not a verdict, especially against Pangram’s much lower claimed error rate that itself comes only from the vendor.
The engineering tradeoff is the more durable story here. Running the model locally means no text leaves the browser, no per-page charge accrues, and the extension can stay on indefinitely: Goedecke reports it idles down after five minutes of inactivity and otherwise uses 400MB to 1.2GB of memory, comparable to a handful of extra open tabs. What it gives up against a server-hosted service like Pangram is model size and, with it, raw accuracy: a laptop cannot run the multi-billion-parameter models that make Pangram’s detection rate possible.
For any team weighing whether to add AI-detection to a product surface, Deckard is a data point on that tradeoff rather than a reason to skip server-side detection. A background always-on signal is useful for triage across large volumes of text, but a compliance or moderation decision that hinges on a single flagged item still needs the more accurate, heavier-weight check.
Sean Goedecke described Deckard’s design and benchmark results in a post on his own site, seangoedecke.com.
