Google Gemini Watermark Checker

Google Gemini Watermark Checker takes pasted text or uploaded files and returns structured JSON on whether the writing is AI generated. It checks for Google DeepMind SynthID-Text — the watermark Google applies to Gemini-generated text — and answers the same question people ask in the Gemini app: Is this AI generated? You get isAiGenerated, confidence, and a short rationale you can store, alert on, or pipe into a moderation workflow.

---

What is Google Gemini Watermark Checker?

Google watermarks text produced by Gemini with SynthID-Text, an invisible token-probability watermark from DeepMind. You can paste a passage, upload a PDF / Word / text file, and get one dataset row per sample: isAiGenerated, watermarkDetected, confidence, likely source, and token usage.

No login to Google. No browser. Files are converted to text on the Actor so the detector only sees the writing — not a 20 MB binary.

---

What can this Actor do?

• Check pasted text or uploaded files (PDF, DOCX, TXT, Markdown, HTML, JSON, CSV)
• Return isAiGenerated plus SynthID-Text / watermark fields in structured JSON
• Sample long files from the start, middle, and end instead of sending a whole book
• Cap characters per item so a huge upload cannot silently burn tokens
• Charge per check and per 1,000 input tokens so large files cost more than a tweet
• Export JSON, CSV, or Excel, or pull results via the Apify API

---

Why check Gemini text for watermarks?

AI writing is now in essays, support tickets, SEO drafts, and newsroom copy. A watermark check will not replace editorial judgment, but it gives you a machine-readable signal you can log next to the document.

Typical jobs:

1. Newsroom provenance — flag Gemini-shaped copy before it ships
2. University integrity — screen submitted essays and lab reports
3. Hiring screens — review cover letters and take-home writeups
4. SEO / content QA — catch fully generated blog drafts
5. Support quality — see whether macros or Gemini wrote the reply
6. Legal discovery — attach an isAiGenerated label to produced documents
7. Brand safety — scan UGC captions and comments at moderate volume
8. Dataset labeling — tag training or evaluation text as AI vs human
9. Publisher intake — check freelance submissions
10. Policy compliance — EU AI Act-style logging of synthetic text
11. Education tools — show students a confidence score, not a verdict
12. Internal audits — sample Slack/docs exports for Gemini-heavy writing

---

How to use Google Gemini Watermark Checker

1. Open the Actor in Apify Console
2. Paste Text / Texts, or upload Files
3. Keep Max characters per item at the default unless you know you need a longer sample (you pay for tokens)
4. Click Start
5. Download the dataset or read it via the Apify API

Input example

{
  "text": "Four score and seven years ago our fathers brought forth on this continent, a new nation...",
  "texts": [
    "In today's rapidly evolving digital landscape, it is important to note that leveraging synergistic paradigms..."
  ],
  "files": ["https://example.com/essay.pdf"],
  "maxItems": 5,
  "maxCharsPerItem": 24000
}

Field	Type	Default	Description
text	string	—	One pasted passage
texts	string[]	—	Extra pasted passages
files	string[]	—	Uploaded file URLs (PDF, DOCX, or text)
maxItems	integer	5	Cap on pasted samples + files this run
maxCharsPerItem	integer	24000	Max characters sent to the detector per item (hard max 100,000)

---

Output

Each dataset item is one check. The headline field is isAiGenerated. model is the detector family (synthid-text), not a specific chat model.

{
  "id": "a1b2c3d4e5f67890",
  "sourceType": "text",
  "sourceName": "pasted-text-1",
  "sourceUrl": null,
  "textPreview": "Four score and seven years ago...",
  "textLength": 312,
  "wordCount": 58,
  "truncated": false,
  "isAiGenerated": false,
  "watermarkDetected": false,
  "watermarkType": "none",
  "matchesTargetProvider": false,
  "likelySource": "human",
  "likelihood": "low",
  "confidence": 0.86,
  "detectionMethod": "hybrid",
  "rationale": "No SynthID-Text signal; wording matches a well-known human speech.",
  "limitations": null,
  "inputTokens": 912,
  "outputTokens": 180,
  "billedTokenUnits": 1,
  "model": "synthid-text",
  "targetProvider": "google-gemini",
  "error": null,
  "checkedAt": "2026-08-26T12:00:00.000Z"
}

---

Pricing

Long files use more tokens than a tweet. This Actor passes that cost through with a margin instead of a flat fee that would lose money on large uploads.

Event	When	Price (FREE)
text-checked	Once per successful analysis	$0.005
input-1k-tokens	Per 1,000 input tokens of the sample actually sent (rounded up)	$0.001
Actor start	Per run (× memory GB)	platform default

Failed downloads are not billed. Failed detector calls are not billed for these events.

Examples (excluding start fee):

• ~500-word paste (~800 tokens) → 1 × $0.005 + 1 × $0.001 ≈ $0.006
• 24,000-character file sample (~6,000+ tokens with the prompt) → 1 × $0.005 + 7 × $0.001 ≈ $0.012

Leave maxCharsPerItem at 24,000 unless you have a reason to send more. Raising it toward 100,000 sends more tokens and costs more.

---

Files and token safety

• Text is extracted locally from PDF, DOCX, and text formats. Images, audio, and zip archives are rejected.
• Files over 15 MB are rejected.
• Longer documents are sampled (start / middle / end) up to maxCharsPerItem.
• Short samples (under ~50 words) are weak for watermarking — the JSON will say so.

A missing SynthID-Text watermark does not prove a human wrote the text. This Actor currently recognizes Google’s SynthID-Text key. ChatGPT text is not SynthID-watermarked. For ChatGPT images and audio, use OpenAI ChatGPT Watermark Checker. For Claude-style copy, use Anthropic Claude Watermark Checker.

---

Is it legal to check AI-generated text?

This Actor only analyzes content you provide. It does not log into anyone’s account. Use it on text you have a right to process, and do not treat the score as legal proof of authorship. Watermarks estimate whether Gemini was involved; they do not identify a person.

---

Tips

• Prefer longer passages. SynthID-Text checks need room to work.
• Upload text-based PDFs, not scanned image-only pages (those have no extractable text).
• Watch billedTokenUnits in the dataset if a run costs more than you expected.
• Start with maxItems 1–3 while you learn the output shape.

---

Image Credit

Image credit: Google