AI Detective analyzes text and estimates how likely it is to have been written by an AI, a human, or some combination of both. It runs entirely through Anthropic's Claude API, using your own API key — there is no account to create and nothing to sign up for on this site.
sk-ant-).Connecting does two things: it confirms your key is valid, and it fetches the actual list of models available to your account, replacing the three default options in the Model dropdown with your real list.
By default, your API key lives only in this browser tab's memory. Close the tab and it's gone — nothing is saved anywhere.
A status bar at the bottom of the screen shows whether your settings are currently being saved to this device, with a one-click toggle:
The first time you successfully connect a key, you'll be asked whether you'd like to remember it. Saying yes turns saving on and stores the key together, in one step.
Provide your text one of three ways, using the tabs at the top of the input card:
| Tab | What it does |
|---|---|
| Paste Text | Type or paste directly into the box. |
| Upload File | Upload a document. Supported formats: TXT, MD, HTML, XML, JSON, DOCX, PPTX, ODT, PDF, RTF, EML. All parsing happens in your browser — files are never uploaded anywhere except the extracted text, which goes to Anthropic when you click Analyze. |
| From URL | Enter a web page URL and click Fetch page text. The server fetches the page and extracts the main article text (skipping navigation, ads, etc.), shown in a box below for review before analyzing. Only public http and https web pages on standard ports can be fetched (pages are downloaded compressed, up to 15 MB of page code); addresses inside private networks and pages on this site itself are refused for security. Some sites can't be fetched: pages built with JavaScript, behind a login or cookie wall, or that block automated access. AI Detective tells you which happened; for those, copy the article text and use Paste Text. A link to a PDF should be downloaded and analyzed with Upload File. You can fetch up to 30 pages every 5 minutes and 500 a day (see Limits). |
| Option | What it does |
|---|---|
| Model | Which Claude model runs the analysis. More capable models cost more per analysis. |
| Effort | Reasoning depth: Standard, Low, Medium, High, X-High, or Max. Higher levels can improve accuracy but cost more. Not every model supports every level — if rejected, you'll see a plain message and the app automatically resets it to Standard. |
| Extended Thinking | Lets Claude reason step-by-step before answering. Works independently of Effort. Increases cost. |
| Flagged AI phrases | Lists up to 15 phrases from your text that most resemble AI writing, shown under the Signals tab and included in copied, TXT, and PDF results (and on each file in Batch). A phrase on this list is a clue, not proof. Adds a small cost. |
| Check citations | Validates any citations found in the document. See Citation Validation below. |
| Split large files into chunks | For documents beyond the recommended size limit, splits the text into pieces, analyzes each, then combines them into one length-weighted result. Without this on, oversized documents are truncated instead (you'll be warned either way). |
| Pre-AI-era warning | Upload File only. When enabled in Settings (see below), warns if the file's date predates your AI era cutoff before analyzing it — such files are unlikely to contain AI-generated text. Not a separate toggle here; controlled from Settings. |
Click Analyze Document. The first time you do this in a new browser session, you'll see:
While an analysis runs, the button becomes Cancel Analysis — click it to stop (you'll be asked to confirm). Once complete, results appear on the right, and a Clear button appears to reset for a new document.
The colored bar shows four percentages that always add up to 100%:
A separate measure of how certain Claude is in the overall assessment — not one of the four categories above. Low confidence often means the text is short, technical, or has mixed characteristics.
The analysis method was rebuilt in November 2026 to reflect the latest research on how AI writing differs from human writing, and on where AI detectors go wrong. In brief:
Even so, no detector is proof. Treat results as one input alongside what you know about the writer and how the text was produced.
Specific observations that informed the assessment, each tagged by type: Strong AI or Moderate AI (confidence in an AI indicator, strong being the higher-confidence of the two), Weak AI (ambiguous, could be either), Human (a specific human writing trait), or Edited (evidence of human revision to AI-generated text, or AI polish applied to human-written text). Color-coded to match the score bar above: AI-family tags in red, Human in green, Edited in blue, Weak AI in gray. Light AI-assisted proofreading of otherwise human-written text is judged as human-authored, not edited — see the FAQ for why this distinction matters.
If citation checking was enabled, a list of every citation found, each marked valid, suspicious, or invalid. See Citation Validation below for full detail.
A plain-English paragraph explaining the verdict. If the document was chunked, this notes how many chunks were analyzed and the range of AI% scores observed across them.
Two export buttons appear after an analysis completes:
Switch to Batch using the toggle at the top of the page. Batch mode analyzes many files in one run.
| Option | What it does |
|---|---|
| Spending limit | The batch stops automatically once this total is reached (checked after each file completes, not mid-file). |
| Pre-AI-era skipping | When enabled in Settings (see below), automatically skips files dated before your AI era cutoff without analyzing them — saves time and cost. Not a separate toggle here; controlled from Settings. |
| Produce detailed Signals file | Collects every file's full signal list into one downloadable text file, in addition to the regular exports. |
All the other options (Model, Effort, Thinking, Flagged Phrases, Citations, Chunking) work identically to Single mode and stay synced with it.
Each file produces a [SIG] line showing its scores, color-coded by the dominant result (red = AI, green = Human, blue = Edited), followed by a [CIT] section if citation checking is on, with per-citation results and a summary count.
If chunking is enabled and a file is too large, it's split and each chunk appears as its own labeled row,
e.g. report.docx [1/3], report.docx [2/3] — unlike Single mode, batch does not
combine chunks into one result; each is shown and counted separately.
Three export options once the batch completes: CSV, Text report, and PDF report (plus Signals Detail if that option was enabled).
When enabled, Claude extracts every citation in the document and screens it for format validity and suspicious patterns. Optionally, the citations Claude is unsure about (or every citation, if Claude screening is off) are also checked against real external databases. Only those citations' text is sent to the selected services, directly from your browser; citations Claude found no problem with are not checked externally:
| Tag | Service | Checks | Setup |
|---|---|---|---|
| Claude | Anthropic Claude | Format and plausibility screening | Uses your existing API key — nothing extra needed |
| CR | CrossRef | Academic journals | Free, no key needed |
| OA | OpenAlex | Academic (broad) | Free, no key needed |
| CL | CourtListener | US case law | Free token — register at courtlistener.com, add it in Settings |
| PM | PubMed | Medical literature | Free; an optional NCBI key raises the rate limit from 3 to 10 requests/sec |
| OL | OpenLaws | Statutes & regulations | Coming soon — paid service |
Each citation is marked:
Every analysis is automatically checked for content that appears designed to manipulate the tool itself — this runs on every analysis, in both modes, with no setting to turn it off, since the people most exposed to this risk are usually the ones who don't know to look for it. There are two independent checks:
| Check | What it looks for |
|---|---|
| Structural | Text hidden in the document's actual file format — Word's "vanish" formatting, extremely small font sizes, CSS display:none/visibility:hidden tricks in HTML or HTML email content. This inspects the file's real markup, not just what's visibly rendered. |
| Semantic | Content — hidden or in plain view — that reads as an attempt to instruct or override the analysis itself (e.g. text addressed to an AI telling it to ignore prior instructions or report a predetermined result). Runs as part of the same analysis call, on every input method including pasted text. |
Nothing appears if nothing is found — the common case stays uncluttered. If something is found:
[SEC] line appears in the batch log for any affected file, and findings are included in every export (Text, PDF, and a compact count in CSV).This is a best-effort screen, not a guarantee. Specifically:
Open Settings via the gear icon in the top bar.
| Section | Contents |
|---|---|
| Appearance | Color theme: System (follows your OS/browser setting), Light, Lark, or Dark. |
| AI Detection | One shared switch enabling the pre-AI-era checks in both modes, plus the cutoff date itself (default: November 30, 2022 — roughly when AI writing tools became widely available). |
| Batch Processing | Max files per batch (default 50) — set to 0 or leave blank for no limit. The cost estimate and pre-run confirmation still apply regardless, so a very large batch can still be reviewed and aborted before it actually runs. |
| Remembered API Key | Shows whether a key is currently saved on this device, with a Forget button. |
| Citation Services | CourtListener token and PubMed/NCBI key fields. |
| Spending Summary | Running total of what you've spent and how many analyses you've run (persists across sessions if you're saving settings; session-only otherwise), with a reset button. |
| Session | Clear your session history (the list of recent analyses, viewable via the clock icon). |
Every analysis uses your own Anthropic API credits — this site never sees or charges you anything itself. An estimated cost is shown before you click Analyze, and the actual cost (calculated from the real tokens used) appears afterward.
The price shown next to each model is its input/output cost per million tokens. AID checks Anthropic's published pricing page once a day and uses those prices for both the estimate and the actual cost; if the check isn't possible, it uses prices built into the current version, and the note under the Model menu says which. A ≈ before a price means the model is newer than Anthropic's price list, so its price is estimated from the same model family. The pre-analysis estimate is based on your document's length (in Batch, on file sizes, which is rougher for compressed formats such as Word files). Anthropic's invoice is the final word on what you're charged.
Extended Thinking and higher Effort levels can significantly increase token usage and therefore cost — the pre-analysis estimate accounts for this roughly, but actual usage varies with the specific document and model. Citation checking adds a second API call, and each chunk in a split document is its own full call.
The same text limits apply however you bring text in: pasting, uploading a file, or fetching a web page.
| Limit | What it means | How to work around it |
|---|---|---|
| 100,000 characters per analysis (about 15,000 words) | With Split large files into chunks off, only the first 100,000 characters are analyzed. AI Detective warns you before this happens in Single mode, and notes it on the file's line in Batch. | Turn on Split large files into chunks. The document is split at natural breaks and each part is analyzed, then combined into one result (each part is its own API call, so it costs more). |
| About 3.2 million characters with splitting on (roughly 500,000 words, 40 parts) | Text beyond 40 parts isn't analyzed. | Split the document into separate files and run them in Batch mode. |
| File size: no limit; a warning above 50 MB | Files are read in your browser, never uploaded to this site. Very large files can make the page slow while they're read. Big PDFs are often scanned pages or images with little real text. | If you only need part of a document, copy that part and use Paste Text. For scanned PDFs, run them through a text-recognition (OCR) tool first. |
| Web pages: up to 15 MB of page code | This site's server downloads the page for you, so it caps the size. In practice that's more text than an analysis can use anyway. | For a bigger page, copy the article text and use Paste Text. |
| Pages that can't be fetched | Pages built with JavaScript, behind a login, cookie wall or paywall, or that block automated access. AI Detective tells you which; a link to a PDF or other file is refused too. | Open the page in your browser, copy the text and use Paste Text. Download PDFs and use Upload File. |
| Page fetches: 30 every 5 minutes, 500 a day | Fetching uses this site's server without an API key, so it has its own limits to prevent abuse. | Wait a few minutes (the message tells you how long), or use Paste Text or Upload File, which don't count as fetches. |
| Requests: 120 every 5 minutes, 2,000 a day | Every analysis, citation check and page fetch counts. Limits are per device and network. | You rarely need to do anything: if a long batch or a document split into many parts reaches the 5-minute limit, AI Detective pauses and continues by itself, showing a countdown (Cancel/Abort still works). The daily limit resets the next day. |
Anthropic also applies its own limits to your API key (based on your account's usage tier). If Anthropic's limit is reached, you'll see Anthropic's message; waiting a minute or raising your tier in the Anthropic Console resolves it.
| Shortcut | Action |
|---|---|
| F5 or Ctrl+Enter | Analyze (Single mode) or Run Batch (Batch mode) |
| Ctrl+Tab | Switch between Single and Batch mode |
| Ctrl+S | Save report as text (Single mode, once a result exists) |
| Ctrl+Shift+C | Copy results (Single mode, once a result exists) |
| Esc | Cancel the current analysis, abort a running batch, or close an open panel |
No, and this is a deliberate limitation, not an oversight. Some AI vendors now embed real cryptographic watermarks in generated text — Google's Gemini since 2024, and Claude models released after August 2026, as part of EU AI Act compliance. Verifying one of these requires the originating vendor's own private detection system, which is not available to third-party tools like AID — Anthropic's detection API, for example, is currently restricted to regulators, fact-checkers, researchers, and compliance-obligated enterprises via application. Any tool claiming to check for a specific vendor's watermark without that vendor's private key is not actually doing so, regardless of what it's named. AID's method has always been stylistic and linguistic pattern analysis, not cryptographic verification, and that's still true today. Worth knowing too: even a genuine watermark hit only confirms a model processed the text at some point — a person's own writing that was just lightly proofread by AI carries the same mark as text the AI substantially drafted, so a watermark alone can't tell you how much of the writing was actually AI-authored. OpenAI's ChatGPT does not currently watermark text output at all.
Double check your key is copied correctly (starts with sk-ant-) and that your Anthropic account
has active credits.
This is normal — it runs roughly once per hour of activity as a bot-prevention measure, not a sign of a problem.
Very long documents are either truncated (with a warning first) or, if you've enabled chunking, split and analyzed in pieces. Turn on Split large files into chunks for full coverage of long documents.
Not every model supports every setting — this varies and changes over time. If a request is rejected, you'll see a plain message saying so, and the app automatically resets that setting (Effort back to Standard, or Extended Thinking off) so you don't have to fix it manually before trying again.
Check the status bar at the bottom of the screen — if it says settings aren't being saved, nothing persists between visits. Turn saving on to keep your preferences.
This guide covers how to use the app. For the full legal and privacy detail, see:
See also Section 7 for how the app checks documents for hidden content and prompt-injection attempts.
In short: your document text goes only to Anthropic's API, using your own key. This site's server never logs, stores, or retains document text, citation content, or your full API key. An anonymous activity log tracks which actions were taken and whether they succeeded, plus each analysis's resulting numbers (AI percentage, confidence, tokens, cost, file type), for debugging and usage analysis — never document content, file names or URLs. Citation checking, when on, sends the text of uncertain citations to the services you select, and a quick Cloudflare human check appears only when needed. Everything else the page uses (fonts, document-reading libraries) is served from this site itself.