#SynthID
Google hat einen unsichtbaren Wasserzeichen in jedes Bild integriert, das Gemini je generiert hat. Über 10 Milliarden Inhalte sind markiert. Ein arbeitsloser Ingenieur hat es gerade geknackt. Mit 200 schwarzen Bildern und bisschen Mathematik.
GitHub - aloshdenny/reverse-SynthID: reverse engineering Gemini's SynthID detection
reverse engineering Gemini's SynthID detection. Contribute to aloshdenny/reverse-SynthID development by creating an account on GitHub.
github.com
April 9, 2026 at 7:22 PM
AI generated, just so people seeing this are aware
September 19, 2026 at 9:25 PM
THREAD: I'll fact-check all the viral misinformation about the US military operation in Venezuela in this thread

This image, purporting to show the US military arresting Venezuelan President Nicolas Maduro, is AI-generated.

According to Google's SynthID detector, it was created using Google AI.
January 3, 2026 at 5:55 PM
It has an OpenAI SynthID. 100% GenAI.
September 27, 2026 at 1:35 AM
AI generated or at least edited
September 20, 2026 at 5:01 PM
Gemini can identify images that have the SynthID watermark
January 3, 2026 at 8:04 PM
i wonder if i should just make a labeler that checks for synthID in images on the firehose
my purpose in life seems to be pointing out gpt-image 2 outputs now
June 28, 2026 at 1:51 PM
in case you missed it, the new fun discourse today is that the AI labeler uses SynthID watermarks to label AI content. good, as it should!

the fun part: anti-AI folks getting content labeled as having AI-generated content are upset that the SynthID watermarks are being "hallucinated". lol.
August 22, 2026 at 9:26 PM
In unserer technischen Analyse ließ sich ein maschinenlesbares SynthID-Wasserzeichen nachweisen. www.mimikama.org/trump-mit-ro...
December 13, 2025 at 7:20 PM
It's actually quite easy to test. Go to SynthId (deepmind.google/models/synth...) , upload the video and ask it to test. I downloaded the video directly from Linkedin
SynthID
SynthID is a tool to watermark and identify AI-generated content, helping to foster transparency and trust in generative AI.
deepmind.google
September 26, 2026 at 8:30 PM
a screenshot (so, metadata stripped) triggers synthid too
September 15, 2026 at 6:46 PM
i have literally had gemini hallucinate synthid results at me, the complaints are not, in fact, insane.
August 23, 2026 at 4:11 AM
"SynthID détecté
Nous avons détecté un filigrane SynthID provenant d'OpenAI."
September 13, 2026 at 5:51 PM
Aww I wanted Larry to be real

Source: www.facebook.com/cat.overlord...

Confirmed as AI-generated using: openai.com/research/ver...
July 4, 2026 at 1:51 AM
This trips Google's SynthID, so this is probably not a photograph but an AI image of an AI image
We are all dogs locked in god’s hot car
August 21, 2026 at 4:15 PM
Most AI image detectors are unreliable but one reason we know this image is fake: it has Google's SynthID watermark.

Google rolled out the ability to ask Gemini to check for the watermark this week. It can't be seen by the naked ID and is embedded in the pixels. This image has the SynthID mark.
November 22, 2025 at 4:55 PM
dawg
June 28, 2026 at 6:41 PM
fair warning that burrito image is likely fake (it has a gemini hidden synthid which means that the image in the post itself, not the poster * were AI processed)

* if someone made the image of burrito, printed and photographed it, synthID would not survive
August 21, 2026 at 4:14 PM
🖥🤖 MAS TAVA DEMORANDO MESMO, ALÔ STF

Vídeos fakes sobre uma suposta paralisação nacional de caminhoneiros e motoboys em apoio a Bolsonaro começaram a circular no TikTok. Segundo o Projeto Comprova, o conteúdo foi gerado por inteligência artificial usando a ferramenta Veo 3, do Google.
June 16, 2025 at 3:29 PM
Just a note to ruin your #caturday

Larry the truffle hunting cat is AI.

Via @aiueo.ooo
July 4, 2026 at 2:49 PM
We know it’s AI and not Photoshop because it has Google’s SynthID watermark which can’t be seen by the naked eye but shows up if you run it through Gemini.
Damn I thought this was a Photoshop composite (background lighting doesn't match the people perfectly).
November 24, 2025 at 4:41 AM
For anyone curious, SynthID-Image is extremely accurate (much more so than text watermarking)
Then press the button and Deep Learning goes brrr and, as long as you have enough training examples, you can learn f and g which can do this very successfully. They report rather astonishing performance, like 100% TPR@FPR=0.001 on watermarked images, and 99.98 on transformed watermarked images.
August 22, 2026 at 9:52 PM
This image, viewed millions of times on Instagram, X and Telegram, purports to show an attack on a US base near Erbil International Airport.

The image is AI-generated. Google's SynthID finds watermarks in the image indicating it was made with Google AI.
March 3, 2026 at 6:19 AM
the level of brutal incuriosity to not understand that validating SynthID is not the same thing as using an LLM
August 22, 2026 at 8:13 PM