Telltail by CogShift 作成者: Cogshift

Telltail by CogShift spots likely AI-written and human-written prose as you browse, and tells you when it isn’t sure.

Telltail is part of CogShift’s work exploring how far small, highly optimised local models can go on tightly bounded tasks — instead of sending every problem to a large cloud model.

For this tightly bounded task, a 36 MB model on your CPU can outperform the cloud LLMs we tested — quickly and quietly enough that you barely notice it running.

Local doesn’t mean weaker. On our held-out benchmark, Telltail achieves 95% of Pangram’s score and matches or beats the cloud LLM judges we tested. It does that locally on your CPU:

• $0 per check
• No API call
• No cloud round-trip
• No text leaving your browser

Public web. Private work. Same local model.

It works locally across the web — on public sites, private email, workplace tools and other pages you choose to enable. Page content never leaves your browser, and Telltail records no browsing telemetry.

Telltail draws a quiet outline around prose, giving you immediate context without getting in the way. Nothing on the page is hidden or removed. Outlines can be customised, filtered or disabled from the extension controls.

Private by design. Telltail does not send the text you read to an AI service. Its optimised ~36 MB classification engine runs directly inside your browser on your own computer.

No cloud inference. No uploads. No account. No API key. No token charges. No ads. No usage limits.

On first run, Telltail downloads its signed model data — model weights, vocabulary, n-gram tables and calibration thresholds — and periodically checks for updates. Those requests contain no browsing content, account information or persistent Telltail user identifier.

Built for high accuracy without false certainty.

Telltail uses deliberately asymmetric thresholds designed to avoid calling human writing AI. A solid outline means the model has crossed a high-confidence bar.

• Solid green — high-confidence human. In held-out testing on about 2,800 social posts the model had never seen, 1,110 of 1,111 solid-green calls were human-written.
• Solid orange — high-confidence AI. In the AI subset of the same benchmark, 116 of 122 (95%) solid-orange calls were AI-written.

Across high-confidence calls, Telltail exceeds 99% accuracy in this benchmark.

Honest about the middle.

Human and AI writing overlap. When the evidence is not strong enough for a high-confidence call, Telltail does not force one.

• Solid outline — high-confidence call: orange for AI, green for human
• Dotted outline — uncertain, coloured in the direction it leans
• Grey dotted outline — toss-up; could be either
• Dashed blue — skipped while busy; never a verdict
• Text too short to judge is left without an outline

Hover any outline to see the confidence score behind it.

No detector is proof. Telltail is an estimate — and when it doesn’t know, it says so.

Setup is simple. Grant Telltail access to the sites where you want to use it. The local model downloads on first launch and outlines begin appearing as you browse. Other sites can be enabled with one click from the popup.