TIME put Google DeepMind’s SynthID on its Best Inventions list for 2026. Pushmeet Kohli’s line is the thesis: neural networks will stop being able to tell real images from synthetic ones, so the company stopped trying to build a checker that guesses and started stamping pixels instead. In October, TIME says, Google launched a site where anyone can upload a file and look for that stamp. That is a watermark product. It is not a ranking of image models.
CNET on October 12 is still in the other business. Best generators. Clarifying prompts. Speed. Canva at 7.5, free, then $13 a month. Start with OpenAI or Canva if you want a free plan. Atlas Cloud the same day ran 22 built-in image requests on one free ChatGPT desktop account and got 23 images back with no limit banner. OpenAI still will not print a number.
We already filed EU labels against SynthID and Images 2.5 as an editor. This week the watermark got a magazine cover and the generators kept their scoreboard.
The October site is not a universal detector
synthid.com is explicit. This is not a general AI detector. It can only speak to media from companies that adopted SynthID. Today that list, in TIME and on the detector FAQ, includes Google, Nvidia, OpenAI, and Kakao, with Apple described as soon. If the file came from a model outside that list, “watermark not detected” is the expected result. It does not mean a human shot the photo.
The detector’s own language: watermark detected means a stamp is present and the file was likely edited or generated by a SynthID-enabled model. Not detected means no confident signal. Heavy edits can degrade the stamp. Pre-adoption models never had one. Treat a negative as a negative for this watermark, not as a certificate of authenticity.
DeepMind’s page says the Detector portal launched earlier this year for journalists and media testers. TIME’s October line is the public, easy-access version. Those can both be true. Do not collapse them into “invented this month.” The invention TIME is celebrating is the stamp. The October object is a URL you can actually use.
If you make images for a living, the useful test is: does your pipeline sit on a partner model. If it does, the October site is a self-check. If it does not, you still need whatever label the EU file already asked you for. A TIME badge does not fill that gap.
Small perturbations are the technical claim
TIME describes the image and video stamp as small perturbations in pixels that cannot be cropped, compressed, or cut after the fact. DeepMind’s wording is milder and more operational: added at creation, designed to stand up to cropping, filters, frame-rate changes, and lossy compression, without a quality hit you would notice. Use DeepMind for the robustness list. Use TIME for the magazine sentence. If a crop actually destroyed the signal, the detector FAQ already told you that heavy modifications can do that. The word “cannot” in TIME is the invention voice. The FAQ is the product voice.
Audio and text are in the same family and are not the same mechanism. Text, on DeepMind’s page, rides token probabilities in the Gemini app and web experience. The model already assigns odds to the next word. SynthID nudges those odds. You will not see it. A pixel checker will not catch it. If your workflow is images, stay on the image path. Do not quote the text method as if it watermarks a PNG.
Kohli’s decade-old decision, in TIME, was that a downstream classifier would lose. Veo3 eating spaghetti is the colorful proof that generation got good enough to make guessing embarrassing. The stamp is an admission. Google would rather mark its own output than play forensic catch-up on everyone else’s.
Audio watermarks, on the same DeepMind page, are a third modality. If your shop ships video with a generated soundtrack, the file can carry more than one stamp. The October site accepts image, video, or audio. Test the export you actually send, not the frame you liked in the editor. A soundtrack swap is an edit. The FAQ already warned that heavy edits can drop the signal.
CNET is still scoring generators
CNET’s October 12 piece is a lab habit the industry has not dropped. How many clarifying prompts until the picture matches. Whether you can follow up with an edit. Most tools finish in under two minutes; 10 to 30 seconds is an edge. Free and freemium exist. The page recommends OpenAI or Canva to start. Canva: 7.5, free, upgrades from $13 a month, “can handle a lot” for beginners and non-professionals.
None of that paragraph mentions SynthID. That is the increment. The invention TIME put on a list does not change the scoreboard people use to pick a tool. You can generate a Canva image in 20 seconds and still have nothing for a detector to find if Canva is not in the partner set. Confirm partners before you tell a client the file is “checkable.”
We already compared GPT Image, Midjourney, and the 2026 field. Do not rerun that matrix here. Steal one question from CNET’s method: if the tool cannot take an edit pass, you will regenerate, and regeneration is how you blow a meter. The watermark does not care. The meter does.
Canva at 7.5 is a beginner score, not a newsroom score. If your desk needs a follow-up edit on a face, CNET’s “clarifying prompts” column is the one that matters, and a tool that cannot edit will burn the free plan faster than a tool that can. TIME’s invention will not sit in that column until CNET decides watermarks are a feature. They have not, in the October 12 piece.
Twenty-three images is a snapshot, not a quota
Atlas Cloud, October 12, 2026: one free ChatGPT account, desktop app, 22 built-in image requests, 23 images in the library, no gray banner. They went looking for the wall because OpenAI does not publish a number for any plan. Plus FAQ, in that write-up, says usage limits may vary based on system conditions. Every blog number, including theirs, is a snapshot.
The app meter and the API meter are separate. The help article on image generation rate limits, they note, is about the API and depends on model plus usage tier. If you are quoting a Twitter screenshot of “I hit 50,” you are mixing products.
Their commercial aside: through the Atlas Cloud plugin, Seedream 5.0 Pro was $0.036 at 1.5K and $0.072 at 2K, 20 percent off list in October 2026. At those rates a $20 Plus month covers about 555 images at 1.5K or 277 at 2K, if the only goal is more pictures and you pay per image. That is a vendor’s arithmetic. It is not OpenAI’s quota. It is useful as a reminder that “free plan mystery cap” and “priced per image” are different contracts.
A watermark does not tell you when the banner will land. A TIME invention does not add requests. If you are producing a campaign, budget the meter as unknown and the stamp as optional depending on the model.
A watermark is not a disclosure
The EU story we already ran was about labels people can see. SynthID is a label machines can see, and only some machines, and only for some vendors. TIME calling it an invention does not make it a policy. You can stamp every Google image and still ship an unmarked file from a non-partner model in the same folder.
For an art team, the operational split is boring and correct. Partner models: run the October checker on a sample of outputs so you know the stamp survives your own crop-and-export path. Non-partner models: do not promise a SynthID result. Client-facing captions are a separate decision. The detector FAQ will not write your caption.
Apple as “soon,” in TIME, is a partnership claim without a ship date in the extract. Do not schedule a workflow on it. Nvidia, OpenAI, Kakao, Google are the present tense list. Re-read the FAQ the week you care, because the partner set is the product.
Kohli is probably right that pure detection was a losing game. That does not mean the winning game is finished. A stamp that four companies use is an island. CNET will keep scoring islands by prompt adherence and speed. Atlas Cloud will keep hitting refresh on a free account to see when the banner shows up. Your job is to know which island you are on before you tell anyone the file can be verified.
Gemini, on DeepMind’s page, is a second checker for people already in that chat. Upload the image, video, or audio clip and ask whether Google AI made or altered it. Gemini looks for SynthID and says if it finds one. That path is not the October website. It is a chat that already has a Google account attached. If your art desk does not live in Gemini, use the upload portal. Do not make “ask the chatbot” the compliance step.
CNET’s prompt advice is still the generator half of the week. Style, aesthetic, dimension, specific elements. Start specific, then cut back if the model chokes. That paragraph belongs next to the meter, not next to the watermark. A specific prompt that takes three regenerations is three hits on a cap OpenAI will not print. SynthID will stamp the third image if the model is a partner. It will not refund the second.
The 23-image run used the desktop app’s built-in image tool, not the API. Atlas Cloud is careful about that. If your production path is an API script with a usage tier, the help article on API image rate limits is the document, and it still depends on model and tier. Mixing a free-app anecdote into an API SLO is how a campaign misses Friday.
Seedream 5.0 Pro at $0.036 and $0.072 is a third path, priced per image, discounted 20 percent on that plugin page in October 2026. It is not ChatGPT. It is not SynthID. It is the reminder that when a vendor will not print a cap, another vendor will print a price. Pick the contract you can explain to a producer.
If you only remember one sentence from TIME, remember the website. Upload. Check. Then read the negative result as “not this watermark.” If you only remember one sentence from CNET, remember 7.5 and $13. If you only remember one sentence from Atlas Cloud, remember that 23 images on October 12 is not a quota you can put in a contract. The invention is real. The meter is still a shrug.