AI Detective — User Guide

1. Getting Started

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.

Getting an API key

1. Go to console.anthropic.com and create a free account, or sign in.
2. Navigate to API Keys and create a new key.
3. Copy the key (it starts with sk-ant-).

Connecting

1. Paste your key into the field at the top of the page.
2. Click Connect.
3. Once connected, the button turns green and reads Connected. Hover over it and it turns red — click it any time to disconnect.

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.

2. Your API Key & Local Storage

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:

• "Settings saved on this device" — your API key (if you chose to remember it), theme, model, and other preferences persist across visits, stored in this browser's local storage — not a cookie, and never sent to or stored on our server.
• "Settings are not being saved" — nothing persists. Every setting resets to default the next time you load the page.

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.

3. Single Document Mode

Provide your text one of three ways, using the tabs at the top of the input card:

Options

Running an analysis

Click Analyze Document. The first time you do this in a new browser session, you'll see:

1. A Document Submission Notice — must be acknowledged once per session (see Privacy, Security & Legal).
2. A brief human-verification check (CAPTCHA) — happens roughly once per hour of activity, to keep the service available for everyone.

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.

4. Reading Your Results

Scores

The colored bar shows four percentages that always add up to 100%:

• AI — likely AI-generated
• Human — likely human-authored
• Edited — AI content edited by a human, or human content with AI assistance
• Unknown — could not be determined with confidence

Confidence

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.

Signals

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.

Citations

If citation checking was enabled, a list of every citation found, each marked valid, suspicious, or invalid. See Citation Validation below for full detail.

Summary

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.

Exporting

Two export buttons appear after an analysis completes:

• Save as Text — a plain-text report.
• Save as PDF — a PDF report that visually mirrors the on-screen results, including the colored score bar and color-coded signals and citations.

5. Batch Mode

Switch to Batch using the toggle at the top of the page. Batch mode analyzes many files in one run.

Selecting files

• Click the drop zone to choose a folder (the default) — everything inside it is added.
• By default, only files directly inside that folder are included. Check Include subfolders to also pull in files from nested folders.
• Or click "Or choose individual files instead" to hand-pick specific files.
• Default limit of 50 files per batch — configurable in Settings (including no limit at all). You'll still see an estimated cost and can abort before a large batch actually runs.

Batch-specific options

All the other options (Model, Effort, Thinking, Flagged Phrases, Citations, Chunking) work identically to Single mode and stay synced with it.

The batch log

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.

Exporting

Three export options once the batch completes: CSV, Text report, and PDF report (plus Signals Detail if that option was enabled).

6. Citation Validation

When enabled, Claude extracts every citation in the document and screens it for format validity and suspicious patterns. Optionally, uncertain citations are also checked against real external databases:

Each citation is marked:

• ✓ Valid — format looks correct, no red flags
• ⚠ Suspicious — possibly fabricated, or dated after Claude's training cutoff
• ✗ Invalid — clear formatting error
• A ? after a service tag (e.g. CR?) means that service could not find a match — not necessarily wrong, just unverified by that particular database.

7. Security & Prompt Injection Detection

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:

Where it's shown

Nothing appears if nothing is found — the common case stays uncluttered. If something is found:

• A banner appears above your results. Amber means hidden text was found but nothing read as manipulative. Red means something read as an actual attempt to manipulate the analysis.
• A Security tab (next to Signals, Citations, and Summary) lists every finding, shown verbatim, with the reason it was flagged.
• In Batch mode, a [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).

Known limitations

This is a best-effort screen, not a guarantee. Specifically:

• Structural (hidden-text) scanning covers DOCX, ODT, HTML/HTM uploads, and the HTML portion of EML files. It does not currently cover PDF (detecting truly invisible PDF text reliably requires deeper access to the file's internal structure than is practical here) or content fetched From URL. Both are still covered by the semantic check, since that runs on the extracted text regardless of source — just not the structural check specifically.
• A sufficiently determined attempt — splitting instructions across non-adjacent text, unusual character encoding — could evade both checks. Treat this as raising the bar significantly, not as absolute protection.

8. Settings

Open Settings via the gear icon in the top bar.

9. Cost & Billing

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.

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.

10. Keyboard Shortcuts

11. Troubleshooting & FAQ

"Could not connect" when I click Connect

Double check your key is copied correctly (starts with sk-ant-) and that your Anthropic account has active credits.

A human-verification check keeps appearing

This is normal — it runs roughly once per hour of activity as a bot-prevention measure, not a sign of a problem.

My document seems cut off in the results

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.

A model I selected won't accept Effort or Extended Thinking

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.

My theme or saved key disappeared

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.

12. Privacy, Security & Legal

This guide covers how to use the app. For the full legal and privacy detail, see:

• License — terms of use, trademarks, and exactly what data is and isn't logged.
• Document Submission Notice — important information about attorney-client privilege, confidentiality, HIPAA/PHI, and NDA considerations before submitting any document. This is shown before your first analysis each session, and can be reviewed any time from the Help menu.

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, for debugging and usage analysis — never document content.