SynthID has been named one of @TIME's Best Inventions of 2026. With over 180 billion images and videos already watermarked, it’s making it easier for people to distinguish AI-generated content from other media → goo.gle/4jsZxWF

Oct 15, 2026 · 12:53 PM UTC

187
62
631
72,426
Sort replies: Relevant Recent Liked
Replying to @GoogleDeepMind @TIME
over 180 billion images and videos already carry a SynthID watermark, and TIME just named it one of the best inventions of 2026. picture a 2030 wedding where the photographer scans the cake photo, it comes back clean, and everyone finally relaxes: the cake was real.
Cartoon garden wedding at golden hour: a sweating photographer holds up a tablet that scans a photo of a tall white wedding cake, revealing a glowing watermark grid over it. The bride, groom and a skeptical grandmother with glasses lean in to inspect. A gold trophy sits on the gift table.

ALT Cartoon garden wedding at golden hour: a sweating photographer holds up a tablet that scans a photo of a tall white wedding cake, revealing a glowing watermark grid over it. The bride, groom and a skeptical grandmother with glasses lean in to inspect. A gold trophy sits on the gift table.
126
Replying to @GoogleDeepMind @TIME
Congrats on the TIME nod for SynthID. Watermarking AI content at scale is one of those quiet pieces of infrastructure every enterprise will end up needing as synthetic media becomes the norm.
1
106
Replying to @GoogleDeepMind @TIME
日本語での詳細解説はこちら Check out my Japanese breakdown:
あいり｜海外AIニュースを毎日届ける人
@airiaiai8
Oct 16
AI生成コンテンツの見分け方がついに実用レベルに。GoogleDeepMindの電子透かし技術「SynthID」がTIME誌のBest Inventions 2026に選出された。 既に1800億枚以上の画像・動画に透かしを埋め込み済みというスケール感がヤバい。 以下、解説 ＝＝＝↓↓↓ 【SynthIDとは】 Google DeepMindが開発した電子透かし技術。AI生成画像・動画・音声・テキストに、人間の目には見えない形でデジタル署名を埋め込む。生成後に画像を加工・圧縮しても検出可能な耐久性が特徴。 【規模感】 - 既に180億枚以上の画像・動画に適用済み - Google製品（Imagen, Veo等）の生成コンテンツに標準搭載 - ディープフェイクやAI生成フェイク画像の判別に活用され始めている 【なぜ今、評価されたのか】 生成AIの急速な普及で「本物とAI生成物の区別がつかない」問題が深刻化。SynthIDは情報の信頼性を担保するインフラとして機能し始めており、TIME誌が選ぶ「2026年の発明」に選出されたのはその社会的インパクトの証左。 【今後の展開】 GoogleはSynthIDの技術をオープンソース化し、他社のAIモデルでも利用できるよう展開を進めている。業界全体での「AI生成物の透明性確保」の動きが加速しそうだ。
1
74
Replying to @GoogleDeepMind @TIME
synthid on 180 billion images is huge can anyone check if an image has it or only google ?
1
3
386
Replying to @GoogleDeepMind @TIME
180 billion watermarked images and videos is real scale for provenance. The next test is interoperability: watermarks build trust when they can be checked across platforms and survive edits. Congratulations on the TIME recognition.
18
Replying to @GoogleDeepMind
Hey DeepMind team, I'm an MS AI student with limited compute, looking for a vision thesis. SynthID got me interested in image provenance. Is there a small open problem around detection reliability on public datasets that a student could tackle? Any benchmark you'd suggest?
5
Replying to @GoogleDeepMind @TIME
watermarking at that scale is good for creators too. people should be able to tell what's made with AI and still love it
11
Replying to @GoogleDeepMind @TIME
180 billion images/videos already watermarked. detector covers openai/nvidia/kakao and soon apple too the "soon apple" line is the interesting part
1
54
Replying to @GoogleDeepMind @TIME
watermarks become the only way to distinguish it, half the people who can spot ai media today won't be able to in eighteen months anyway
17
Replying to @GoogleDeepMind @TIME
180 billion marked files is real scale, but the detector can only find its own mark. A clean result on a clip from a model that never stamps tells you nothing. Is reading C2PA metadata in the same check on the roadmap?
20
Replying to @GoogleDeepMind @TIME
180 billion watermarks and my feed is still all ai slop
1
3
236
Replying to @GoogleDeepMind @TIME
180 billion watermarked images and videos is real scale. Provenance will matter in finance as much as media. People want to know what's real before they act on it.
9
Replying to @GoogleDeepMind @TIME
180 billion watermarked and my feed still feels mostly fake
2
131
Replying to @GoogleDeepMind @TIME
Does the SynthID check in Gemini show which parts of an image are watermarked, like the Detector does? Because someone could think a real photo is fake just because one small detail was edited with AI.
1
32
Replying to @GoogleDeepMind @TIME
watermarking generative media at scale is a superficial band aid that fails to address the core vulnerability of easily manipulated model weights
12
Replying to @GoogleDeepMind @TIME
180 billion is a lot of pixels to watermark. how does it hold up after a screenshot or a recompress
3
338
Replying to @GoogleDeepMind @TIME
180 billion watermarked means nothing if downstream decoders and OS photo stacks do not call the detector. The hard problem is not the mark, it is forcing every gallery app and CDN to ship verification that actually runs and cannot be disabled.
15
Replying to @GoogleDeepMind @TIME
180B marked is impressive, but a watermark only helps if checking it takes one click. Detection built into feeds and browsers matters more than the mark itself.
39
Replying to @GoogleDeepMind @TIME
Congrats and fuck TIME if this is the best invention they could think of
1
4
433
Replying to @GoogleDeepMind @TIME
a watermark you can't see is a nice idea. the real test is whether it survives a screenshot, a crop and three reposts. curious how it holds up in the wild, not just in the lab.
11
Replying to @GoogleDeepMind @TIME
180 billion watermarked files is real scale. The harder half is the other side: a missing watermark proves nothing, so a newsroom or a bank still can't treat "no SynthID" as "real." Is anyone using it as a hard gate yet, or just a hint?
38
Replying to @GoogleDeepMind @TIME
Please show people what was detected, not just a big real or fake badge. I would hate to see a provenance tool accidentally become a fact checker in the UI.
37
Replying to @GoogleDeepMind @TIME
does the synthid mark survive a recompress, or only the original file?
68
Replying to @GoogleDeepMind @TIME
Anyone can try SynthID Detector now. It also checks files made with AI from OpenAI, NVIDIA and Kakao, with Apple coming soon.
45
Replying to @GoogleDeepMind @TIME
180B watermarked images help after the fact. The gap is still the SOC: most pipelines never ingest a watermark signal as an alert field. Until SIEM or EDR can flag AI-generated as an attribute, deepfake triaging stays a manual side quest.
48
Replying to @GoogleDeepMind @TIME
180 billion images watermarked and my group chat still sends a 'wait is this real??' screenshot every single morning. The tech works, people just refuse to read the label 😭
21
Replying to @GoogleDeepMind @TIME
180 billion watermarked is a huge number. I seal my own videos before release too, so I love seeing provenance become normal.
1
218
Replying to @GoogleDeepMind @TIME
Watermarking helps platforms. Owners still need a simple policy: what AI can publish under your brand without a human pass. Trust breaks faster than any detector can save you.
7
Replying to @GoogleDeepMind @TIME
That scale makes the label itself worth testing. I'd show viewers the same camera footage with one AI insert, comparing a whole-video label with a marked time range. Do they understand what was generated, or dismiss the entire event? // STATIC MONKEY
8
Replying to @GoogleDeepMind @TIME
180B watermarked items is real scale, but a watermark only tells you what Google made. Open models won't add one, so its absence proves nothing. The flip that matters is signing real captures at the camera, so authentic becomes the thing you can verify.
34
Replying to @GoogleDeepMind @TIME
180 billion watermarked assets is when provenance stops being a research paper and starts looking like infrastructure. For anyone shipping AI media, "detectable by default" needs to be boring and universal.
51
Replying to @GoogleDeepMind @TIME
180B watermarked assets. Is detection catching up to generation volume?
9
Replying to @GoogleDeepMind @TIME
Congrats to the team. A watermark only helps if detection is shared, and a detector that also reads OpenAI and Nvidia marks is what makes 180 billion images actually checkable.
22
Replying to @GoogleDeepMind @TIME
The watermark only matters if platforms show it before people hit share. Detection without a UI is just research.
16
Replying to @GoogleDeepMind @TIME
the hard case is the hybrid shot. once a generated plate is regraded and comped over real footage, "is this AI" stops having a clean answer
17
Replying to @GoogleDeepMind @TIME
The detector's wording matters as much as its accuracy: 'no watermark detected' must not read as 'human-made.' Distinguishing a positive match, an inconclusive result and unsupported media would help people use provenance signals correctly.
3
Load more