AI Content Detector

How to use this AI content detector

1. Paste the text you want to check.
2. Click "Analyze" for a rough likelihood assessment and reasoning.
3. Treat the result as a signal to investigate further, not a final verdict.

Why do I need my own API key?

Running AI models costs money per request. Instead of OmniDeck paying for (and rate-limiting) everyone's usage, you use your own account's credits, and we simply provide the interface.

Is my key safe?

Your key is used only to call the provider directly from your browser. It is never transmitted to or stored on OmniDeck's servers. If you check "remember," it is saved in your browser's local storage, which only your device can read.

Getting a "blocked by CORS" error?

Some providers restrict direct browser requests for security. If a provider fails, try a different one from the dropdown — Anthropic's API is the most reliable option for this kind of direct-from-browser use.

Why AI text detection is fundamentally unreliable, not just imperfect

AI content detectors, including this one, work by looking for statistical patterns that tend to correlate with machine-generated text — things like unusually consistent sentence length, lower "perplexity" (how predictable each word is given the words before it), and less of the natural variation ("burstiness") that human writing tends to have. The problem is that none of these signals are exclusive to AI text: a careful human writer, a non-native speaker following textbook grammar rules closely, or simply a topic with limited natural vocabulary can produce text with the same statistical fingerprint a detector associates with AI. This isn't a bug that better detectors will eventually fix — it's a structural limitation of trying to infer authorship from text statistics alone, which is why no publicly available AI detector, from any vendor, has been shown to be reliably accurate across real-world writing.

False positives are a real, documented harm, not a theoretical edge case

This isn't an abstract concern — AI detectors have a well-documented history of disproportionately flagging text written by non-native English speakers, since these writers often follow grammar rules more rigidly and use less idiomatic variation than native speakers, which happens to resemble the statistical patterns detectors associate with AI generation. This has caused real harm in academic settings, where students have faced accusations of AI-assisted cheating based substantially on detector output, despite having written the text themselves. Any use of a tool like this to make a consequential decision about someone — academic, professional, or otherwise — carries real risk of a false accusation, which is exactly why the disclaimer above this tool exists and should be taken seriously.

Using an AI model to detect AI text is a somewhat circular approach

This specific tool works by asking an AI model to assess whether a piece of text looks AI-generated, which is worth understanding as a distinct approach from a dedicated statistical detector: the model is essentially pattern-matching based on its own training, offering a judgment rather than a mathematically calibrated probability score. It has no access to any ground truth about how the text was actually produced, no ability to verify authorship through any external means, and can be wrong with the same unwarranted confidence a human reviewing the same text might have. Treat its output as one model's educated guess, phrased as an assessment — useful as a starting point for further investigation, not as evidence in itself.

Why even small edits defeat detection reliably

The statistical patterns detectors rely on are fragile enough that lightly rewriting a handful of sentences, replacing a few words with synonyms, or manually varying sentence structure in an AI-drafted piece of text is often enough to shift the assessment noticeably, even though the text is still substantially the same content. This has spawned an entire category of "AI humanizer" and paraphrasing tools marketed specifically to help text evade detection, which puts detector and evader in a permanent cat-and-mouse dynamic where the detector is usually a step behind — a genuinely determined effort to disguise AI-assisted text will very often succeed against any detector, including this one.

Limitations of this tool

This tool asks an AI model to assess a piece of text and explain its reasoning — it doesn't run a dedicated, purpose-built detection algorithm, doesn't have access to any database of known AI outputs to compare against, and can't distinguish AI-written text that's been lightly edited by a human from text a human wrote from scratch. Its accuracy also depends on which provider and model you select, and none of them should be treated as authoritative. Use this only as a rough, informal signal worth investigating further through other means — never as standalone proof in any situation where the answer genuinely matters, particularly one involving an accusation against a specific person.