Gemini tries to sniff out AI slop images while also making them easier to create
(2025/11/20)
- Reference: 1763674064
- News link: https://www.theregister.co.uk/2025/11/20/google_ai_image_detector/
- Source link:
hands on Google Gemini users can now use the AI's app and website to figure out whether an image is AI-generated, though with some considerable limitations.
Google today announced the general [1]release of SynthID Detector in the Gemini app and Gemini on the web, allowing users to upload any image and ask the bot whether the picture was created or modified using AI. Because the new feature is using SynthID, however, it's incredibly limited – it can only recognize images created by Gemini and tagged with Google-made SynthID watermarks.
[2]Developed in 2023, SynthID has been part of Google's various image-generating AI models for a couple of years now. The watermarking system is [3]designed to be imperceptible to humans and still detectable when images are cropped or modified, according to Google.
[4]
SynthID is open source, and Google has scored a few high-profile partners in [5]Hugging Face and [6]Nvidia , but aside from those examples, Gemini's SynthID Detector won't be able to actually determine whether an image is AI-created with any sort of certainty.
[7]
[8]
In our testing of AI images generated via ChatGPT, Gemini wasn't reliably able to tell if the picture was AI-generated or not, but it did reason correctly a few times based on small details that give AI-generated content away. Images created using Gemini, on the other hand were all flagged as containing SynthID watermarks.
Instead of relying on watermarking like Google, ChatGPT relies on a metadata system developed by the Coalition for Content Provenance and Authenticity (C2PA) to mark its content as AI-generated. Along with OpenAI, C2PA is used by companies including Microsoft, Adobe, Meta, and others.
[9]
A Google spokesperson told The Register that it plans to add support for C2PA to Gemini in the future "so you'll be able to check the original source of content created by models and products that exist outside of Google's ecosystem."
Even with C2PA support, experts have warned that detecting AI images based on watermarks and metadata isn't the most reliable way to ensure provenance.
As we reported in June, computer scientists at Canada's University of Waterloo [10]developed a method dubbed "UnMarker" they said can remove AI image watermarks - including SynthID - in just a few minutes using a 40 GB Nvidia A100 GPU, no internet connection required.
[11]
Google Deepmind researchers previously [12]concluded much the same, and that paper even called C2PA metadata into question, remarking that it was even less robust than watermarking.
[13]China announces plan to label all AI-generated content with watermarks and metadata
[14]Watermarks on AI art a futile game of digital hide-and-seek
[15]TikTok becomes first platform to require watermarking of AI content
[16]How 'AI watermarking' system pushed by Microsoft and Adobe will and won't work
In other words, as with any modern AI product, don't trust SynthID Detector to reliably tell fact from fiction - it can be tricked by bad input data just as well as your average LLM.
Oh, there's a Nano Banana Pro now, too
[17]
A Nano Banana Pro-generated image of a man generating an image of a man generating an image with Nano Banana Pro - Click to enlarge
Along with the public release of AI detection in Gemini, Google also today released Nano Banana Pro, an upgraded version of its Nano Banana image generation engine built using the [18]recently-released Gemini 3 Pro model mere months after releasing the initial version, which we were [19]quite impressed with.
The latest model, according to The Chocolate Factory, is specifically designed with clearer text in mind, and can also be used to generate infographics with actual readable content instead of the language-adjacent gibberish that has been typical of previous generations of AI image generation tools.
We used Nano Banana Pro to create the recursively-themed image above, as well as the refreshing-looking bottle of Reg-branded soda included below. Compared to ChatGPT, Nano Banana Pro created the image incredibly quickly, and with no need to be re-prompted for text corrections.
[20]
Aah, refreshing - created using Nano Banana Pro - Click to enlarge
Like all of Google's AI-generated images, those created in Nano Banana Pro include a visible Gemini sparkle watermark as well as SynthID content.
For those curious, we ran the Nano Banana Pro-created image of The Register soda through a tool designed to strip out SynthID watermarking, but Gemini still detected the presence of SynthID in the image. Even then, Gemini told us, it probably would have picked up that it wasn't real based on our reputation.
"[The Register] frequently uses humorous or custom-made graphics for their article headers," Gemini explained. "This image was likely generated to serve as a stock photo or illustration for one of their stories."
Well, drat - You've got us this time, Gemini. ®
Get our [21]Tech Resources
[1] https://blog.google/technology/ai/ai-image-verification-gemini-app/
[2] https://www.theregister.com/2023/09/11/google_think_tanks/
[3] https://ai.google.dev/responsible/docs/safeguards/synthid
[4] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_software/aiml&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=2&c=2aR-dipKtlylGDLC1lGLTlgAAANA&t=ct%3Dns%26unitnum%3D2%26raptor%3Dcondor%26pos%3Dtop%26test%3D0
[5] https://huggingface.co/spaces/google/synthid-text
[6] https://nvidianews.nvidia.com/news/nvidia-alphabet-and-google-collaborate-on-the-future-of-agentic-and-physical-ai
[7] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_software/aiml&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=4&c=44aR-dipKtlylGDLC1lGLTlgAAANA&t=ct%3Dns%26unitnum%3D4%26raptor%3Dfalcon%26pos%3Dmid%26test%3D0
[8] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_software/aiml&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=3&c=33aR-dipKtlylGDLC1lGLTlgAAANA&t=ct%3Dns%26unitnum%3D3%26raptor%3Deagle%26pos%3Dmid%26test%3D0
[9] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_software/aiml&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=4&c=44aR-dipKtlylGDLC1lGLTlgAAANA&t=ct%3Dns%26unitnum%3D4%26raptor%3Dfalcon%26pos%3Dmid%26test%3D0
[10] https://www.theregister.com/2025/07/24/ai_watermarks_unmarker/
[11] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_software/aiml&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=3&c=33aR-dipKtlylGDLC1lGLTlgAAANA&t=ct%3Dns%26unitnum%3D3%26raptor%3Deagle%26pos%3Dmid%26test%3D0
[12] https://arxiv.org/html/2502.04901v1
[13] https://www.theregister.com/2025/03/17/asia_tech_news_roundup/
[14] https://www.theregister.com/2024/01/08/ai_in_brief/
[15] https://www.theregister.com/2024/05/10/tiktok_ai_watermarking/
[16] https://www.theregister.com/2023/10/15/microsoft_adobe_ai_watermark/
[17] https://regmedia.co.uk/2025/11/20/nano-banana-pro.jpg
[18] https://blog.google/products/gemini/gemini-3/#gemini-3
[19] https://www.theregister.com/2025/08/26/google_gemini_ai_images/
[20] https://regmedia.co.uk/2025/11/20/the-register-soda.jpg
[21] https://whitepapers.theregister.com/
Google today announced the general [1]release of SynthID Detector in the Gemini app and Gemini on the web, allowing users to upload any image and ask the bot whether the picture was created or modified using AI. Because the new feature is using SynthID, however, it's incredibly limited – it can only recognize images created by Gemini and tagged with Google-made SynthID watermarks.
[2]Developed in 2023, SynthID has been part of Google's various image-generating AI models for a couple of years now. The watermarking system is [3]designed to be imperceptible to humans and still detectable when images are cropped or modified, according to Google.
[4]
SynthID is open source, and Google has scored a few high-profile partners in [5]Hugging Face and [6]Nvidia , but aside from those examples, Gemini's SynthID Detector won't be able to actually determine whether an image is AI-created with any sort of certainty.
[7]
[8]
In our testing of AI images generated via ChatGPT, Gemini wasn't reliably able to tell if the picture was AI-generated or not, but it did reason correctly a few times based on small details that give AI-generated content away. Images created using Gemini, on the other hand were all flagged as containing SynthID watermarks.
Instead of relying on watermarking like Google, ChatGPT relies on a metadata system developed by the Coalition for Content Provenance and Authenticity (C2PA) to mark its content as AI-generated. Along with OpenAI, C2PA is used by companies including Microsoft, Adobe, Meta, and others.
[9]
A Google spokesperson told The Register that it plans to add support for C2PA to Gemini in the future "so you'll be able to check the original source of content created by models and products that exist outside of Google's ecosystem."
Even with C2PA support, experts have warned that detecting AI images based on watermarks and metadata isn't the most reliable way to ensure provenance.
As we reported in June, computer scientists at Canada's University of Waterloo [10]developed a method dubbed "UnMarker" they said can remove AI image watermarks - including SynthID - in just a few minutes using a 40 GB Nvidia A100 GPU, no internet connection required.
[11]
Google Deepmind researchers previously [12]concluded much the same, and that paper even called C2PA metadata into question, remarking that it was even less robust than watermarking.
[13]China announces plan to label all AI-generated content with watermarks and metadata
[14]Watermarks on AI art a futile game of digital hide-and-seek
[15]TikTok becomes first platform to require watermarking of AI content
[16]How 'AI watermarking' system pushed by Microsoft and Adobe will and won't work
In other words, as with any modern AI product, don't trust SynthID Detector to reliably tell fact from fiction - it can be tricked by bad input data just as well as your average LLM.
Oh, there's a Nano Banana Pro now, too
[17]
A Nano Banana Pro-generated image of a man generating an image of a man generating an image with Nano Banana Pro - Click to enlarge
Along with the public release of AI detection in Gemini, Google also today released Nano Banana Pro, an upgraded version of its Nano Banana image generation engine built using the [18]recently-released Gemini 3 Pro model mere months after releasing the initial version, which we were [19]quite impressed with.
The latest model, according to The Chocolate Factory, is specifically designed with clearer text in mind, and can also be used to generate infographics with actual readable content instead of the language-adjacent gibberish that has been typical of previous generations of AI image generation tools.
We used Nano Banana Pro to create the recursively-themed image above, as well as the refreshing-looking bottle of Reg-branded soda included below. Compared to ChatGPT, Nano Banana Pro created the image incredibly quickly, and with no need to be re-prompted for text corrections.
[20]
Aah, refreshing - created using Nano Banana Pro - Click to enlarge
Like all of Google's AI-generated images, those created in Nano Banana Pro include a visible Gemini sparkle watermark as well as SynthID content.
For those curious, we ran the Nano Banana Pro-created image of The Register soda through a tool designed to strip out SynthID watermarking, but Gemini still detected the presence of SynthID in the image. Even then, Gemini told us, it probably would have picked up that it wasn't real based on our reputation.
"[The Register] frequently uses humorous or custom-made graphics for their article headers," Gemini explained. "This image was likely generated to serve as a stock photo or illustration for one of their stories."
Well, drat - You've got us this time, Gemini. ®
Get our [21]Tech Resources
[1] https://blog.google/technology/ai/ai-image-verification-gemini-app/
[2] https://www.theregister.com/2023/09/11/google_think_tanks/
[3] https://ai.google.dev/responsible/docs/safeguards/synthid
[4] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_software/aiml&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=2&c=2aR-dipKtlylGDLC1lGLTlgAAANA&t=ct%3Dns%26unitnum%3D2%26raptor%3Dcondor%26pos%3Dtop%26test%3D0
[5] https://huggingface.co/spaces/google/synthid-text
[6] https://nvidianews.nvidia.com/news/nvidia-alphabet-and-google-collaborate-on-the-future-of-agentic-and-physical-ai
[7] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_software/aiml&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=4&c=44aR-dipKtlylGDLC1lGLTlgAAANA&t=ct%3Dns%26unitnum%3D4%26raptor%3Dfalcon%26pos%3Dmid%26test%3D0
[8] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_software/aiml&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=3&c=33aR-dipKtlylGDLC1lGLTlgAAANA&t=ct%3Dns%26unitnum%3D3%26raptor%3Deagle%26pos%3Dmid%26test%3D0
[9] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_software/aiml&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=4&c=44aR-dipKtlylGDLC1lGLTlgAAANA&t=ct%3Dns%26unitnum%3D4%26raptor%3Dfalcon%26pos%3Dmid%26test%3D0
[10] https://www.theregister.com/2025/07/24/ai_watermarks_unmarker/
[11] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_software/aiml&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=3&c=33aR-dipKtlylGDLC1lGLTlgAAANA&t=ct%3Dns%26unitnum%3D3%26raptor%3Deagle%26pos%3Dmid%26test%3D0
[12] https://arxiv.org/html/2502.04901v1
[13] https://www.theregister.com/2025/03/17/asia_tech_news_roundup/
[14] https://www.theregister.com/2024/01/08/ai_in_brief/
[15] https://www.theregister.com/2024/05/10/tiktok_ai_watermarking/
[16] https://www.theregister.com/2023/10/15/microsoft_adobe_ai_watermark/
[17] https://regmedia.co.uk/2025/11/20/nano-banana-pro.jpg
[18] https://blog.google/products/gemini/gemini-3/#gemini-3
[19] https://www.theregister.com/2025/08/26/google_gemini_ai_images/
[20] https://regmedia.co.uk/2025/11/20/the-register-soda.jpg
[21] https://whitepapers.theregister.com/