What Is Google SynthID? Can You Remove the Invisible AI Watermark?

There Are Two Kinds of AI Watermark — Most People Only Know About One

When people talk about removing watermarks from NotebookLM or Gemini, they usually mean the visible badge — the "Made with NotebookLM" logo in the corner of a video, or the sparkle glyph on a Gemini image. That kind of watermark is a graphic element sitting on top of your content. It can be cropped, painted over, or removed with the right tool.

But Google also deploys a second, completely different watermarking system: SynthID. This one is invisible. You can't see it, you can't crop it out, and no current public tool can reliably strip it. If you've ever wondered why Google seems relaxed about people removing visible watermarks, SynthID is a big part of the answer.

This article explains what SynthID actually is, how it works, why it's fundamentally different from the visible badge, and what it means for anyone using Google's AI tools.

What Is SynthID?

SynthID is an invisible watermarking technology developed by Google DeepMind. It was introduced in 2023 and has since been expanded across Google's AI product line. Its purpose is to mark AI-generated content so it can be identified later — by Google, by other platforms, or by detection tools — without being visible to humans.

SynthID works across multiple media types:

• Images — modifies pixel values at a level imperceptible to the human eye
• Audio — embeds a signal in the audio waveform that survives compression and format conversion
• Video — marks individual frames
• Text — adjusts token probabilities during generation so the statistical pattern of the output carries a detectable signature (called "watermarking at the logit level")

The key property of SynthID is robustness: the watermark is designed to survive common transformations like cropping, resizing, compression, screenshots, re-encoding, and format conversion. It's not a metadata tag that can be stripped with an EXIF cleaner — it's woven into the content itself.

SynthID vs. the Visible Watermark: A Complete Comparison

Property	Visible Watermark	SynthID (Invisible)
Can you see it?	Yes — logo, text, or sparkle overlay	No — imperceptible to humans
Purpose	Brand promotion + upsell to paid tier	AI provenance tracking + content authentication
Where it lives	On top of content (overlay, shape, annotation)	Inside the signal (pixel values, audio waveform, token distribution)
Can it be removed?	Yes — crop, paint, or use a removal tool	Not practically — designed to survive all common edits
Survives screenshot?	Only if the badge area is captured	Yes (for images and video)
Survives compression?	Yes (it's a visual element)	Yes (designed for robustness)
Removed by Ultra plan?	Yes	No — SynthID persists even on paid-tier exports

That last row is important: even NotebookLM Ultra subscribers get SynthID-watermarked content. Paying $250/month removes the visible badge but not the invisible fingerprint. SynthID is not a free-tier penalty — it's applied to all AI-generated content regardless of your plan.

Gemini Images Have Both Watermarks

This is where the confusion peaks. When you generate an image with Google Gemini, the output carries two separate watermarks simultaneously:

1. The visible sparkle (✦) — an alpha-blended overlay in the bottom-right corner. This is the one our Gemini image tool removes via alpha-channel reversal. The math is straightforward: original = (watermarked - α×255) / (1-α), and the result is near-lossless.
2. SynthID — an invisible signal embedded directly into the pixel data during image generation. This remains after you remove the sparkle. It remains after you crop, resize, compress, or screenshot the image.

So when you use our tool to remove the Gemini watermark, you're removing the visible sparkle. The SynthID fingerprint stays. For most users this doesn't matter — SynthID is invisible and doesn't affect how the image looks or functions. But it does mean the image can still be identified as AI-generated by anyone with a SynthID detector.

Can You Actually Remove SynthID?

The short answer: not reliably, and not without potentially degrading your content.

SynthID is designed by one of the world's top AI research labs specifically to resist removal. Here's why it's so hard:

• No public specification — the exact embedding algorithm is not published. You can't reverse what you can't fully describe.
• Signal-level embedding — for images, the watermark is spread across pixel values in a way that's statistically detectable but visually invisible. Simply adding noise or adjusting colors isn't guaranteed to remove it, and heavy-handed approaches degrade image quality.
• Redundancy — the signal is embedded redundantly, so partial destruction (cropping, lossy compression) doesn't eliminate it.
• Adversarial robustness — Google explicitly designs SynthID to resist adversarial removal attempts. It's an ongoing arms race, and Google has vastly more resources than individual attackers.

Some academic research has explored attacks on invisible watermarks in general, and heavily degrading the image (extreme compression, adding substantial noise, regenerating through another AI model) can sometimes reduce detection confidence. But these approaches damage your content in the process — you'd be trading an invisible watermark for a visibly worse image.

Why Does This Matter for NotebookLM Users?

For most practical purposes, SynthID doesn't affect you. If your goal is to use a NotebookLM video in a presentation, or share a Gemini image on social media, or submit a slide deck to a client — the invisible watermark has zero impact on any of those use cases. Nobody looking at your content can see it.

Where it could matter:

• AI content detection — platforms that deploy SynthID detectors (Google Search, YouTube, social networks) can flag your content as AI-generated. If you're submitting content to a platform that penalizes or labels AI content, SynthID may trigger that.
• Academic integrity tools — as universities adopt AI detection, SynthID could be one signal used to flag AI-assisted work.
• Legal and regulatory — emerging AI transparency regulations (like the EU AI Act) may require AI-generated content to be identifiable. SynthID is how Google complies.

For these edge cases, the practical advice is: don't rely on watermark removal (visible or invisible) to disguise AI-generated content. Use AI tools openly and transparently where appropriate, and remove the visible badge only for aesthetic and professional presentation reasons — which is exactly what our tools are designed for.

What You Can Remove (and How)

To be clear about what's actionable: the visible watermarks across all Google AI products are fully removable, and NotebookLM Remover handles all of them:

• NotebookLM videos — logo overlay + "Made with Google" end card → FFmpeg delogo + tail trim
• NotebookLM PDFs & slides — text badge in the corner → detected and gradient-filled
• NotebookLM PPTX — embedded watermark image → unpacked, cleaned, repacked
• NotebookLM infographics — connected-component detection + gradient fill
• Gemini images — alpha-channel reversal, mathematically near-lossless

All processing happens in your browser. No upload, no account, no cost. The visible watermarks disappear; SynthID stays but doesn't affect how your content looks or functions.

Frequently Asked Questions

Does paying for NotebookLM Ultra remove SynthID?

No. Ultra removes the visible badge only. SynthID is applied to all AI-generated content regardless of your subscription tier — it's a provenance-tracking system, not a free-tier limitation. Google applies it to comply with emerging AI transparency norms.

Can I tell if my content has a SynthID watermark?

Not with consumer tools. SynthID detection requires specialized software. Google offers detection through some of its products and APIs, and there are a few academic research tools, but there's no simple "SynthID checker" available to the public. If the content was generated by a Google AI product (Gemini, NotebookLM, Imagen), it almost certainly has SynthID embedded.

Is SynthID the same as C2PA metadata?

No. C2PA (Coalition for Content Provenance and Authenticity) is a metadata-based standard — it stores provenance information in the file's metadata, which can be stripped with any EXIF/metadata cleaner. SynthID is embedded in the content signal itself and can't be removed by stripping metadata. Some Google products use both: C2PA metadata for voluntary provenance, SynthID for tamper-resistant identification. They're complementary systems solving the same problem in different ways.