The Signal

Upload the File. A Miss Is Not Proof It Is Human.

SynthID Detector opened to everyone, in English, on 7 October 2026. Drag the file in. If the page says the watermark was not detected, that is not a clean bill.

PF

Pamela Flannery

Chief Marketing Officer

Before the post goes live, run the file through SynthID Detector. Drag it in. Read the label the page gives you. If that label says the watermark was not detected, leave "human" out of the caption. The page says that result does not rule AI out.

On 7 October 2026, Pushmeet Kohli at Google DeepMind wrote that Detector access is expanding to everyone. The tool is available globally, in English, starting that day.[1] The same post points the check at synthid.com.[1]

I fetched that post and the Detector page on 9 October 2026. I did not sign in. I did not upload a file. The steps below are the controls those pages name. If a button is missing in your session, that is the page you have, not a click I skipped.

What the check can see

Kohli's post says anyone can check whether an image, a video, or an audio file was made with AI from Google or from partners named on that page.[1] The partner names are in that sentence. Read them there before you brief a team. This piece is the check, not the roster.

The same post gives Google's own totals. It says Google has watermarked over 180 billion images and videos, and 240,000 years of audio.[1] It says built-in checks in Search, the Gemini app, and Chrome regularly handle over 1 million requests a day.[1] Those are Google's figures. I did not recount them, and I did not fetch a click path for Search, Gemini, or Chrome. The upload path I can teach is the Detector page.

The Detector page draws the line in one sentence. "This is not a general AI detector." It can only detect media from the companies shown above that line.[2] A file from a tool that is not on that list can miss, and the Detector is doing what it says.

You are not asking "is this AI?" You are asking whether this file still carries a SynthID watermark from a company this page can see.

Open the page, and deal with the terms if they are in the way

The address is synthid.com. The page I fetched was titled SynthID Detector. Under that, it said "Identify AI generated media," then "Upload a file for detection."[2]

The same fetch also showed a terms screen, "Welcome to the SynthID Detector." The choices on that screen are Agree and continue, and a No thanks link that goes to DeepMind's SynthID page.[2] I did not click Agree. I will not invent the screen that follows. If the terms are in front of the upload box, read them, then Agree and continue if you want the check.[2]

The fetch ended with "Not signed in."[2] The privacy notice says that if you authenticate with a third-party account, Google processes basic account identifiers from that provider, including an email address, a user ID, and the sign-in provider, to manage the session and to limit abuse.[2] The pages I fetched do not say whether a signed-out visit can finish a check. Do not brief that either way.

There is a Learn more link on the Detector page. It points at DeepMind's SynthID technology page.[2] I did not fetch that page. Do not treat it as a second set of steps.

Use the file you would publish

The Detector page asks you to upload the highest quality file you can.[2] It says heavy compression and editing affect detection, and that a high-quality file gives you the best chance of a read.[2]

If the asset in the queue is a crushed export, do not check that copy and call it the source. Check the file you would hand to the channel. If you still have the master, check that too. The page recommends more than one point of data before you decide.[2] The master and the export are two points you already have.

The formats on the page are specific. If your export is not on the list, export again before you blame the Detector.

Images:

  • JPG, JPEG, PNG, BMP, WEBP, AVIF, HEIC, HEIF, TIFF, TIF, GIF

Video:

  • MP4, MOV, WEBM

Audio:

  • WAV, MP3, OGG, FLAC, AAC, M4A[2]

The page I fetched does not name a file-size cap. Do not invent one. If the upload refuses a file, that refusal is the product, not a limit I measured.

The upload control is one line. "Drag & drop, or Select file."[2] That is the click path. There is no second menu in the text I fetched.

How the mark gets into the file

You do not need the math to use the page. You do need the plain version, because it changes what a miss means.

SynthID is Google DeepMind's watermark. The Detector FAQ says it embeds an imperceptible mark into AI-generated images, video, audio, and text, so the file can later be identified as AI-generated.[2] For images and video, the mark is in the pixels. For audio, it is in the frequency components.[2] The FAQ says the mark is meant to travel with the file when people share it, including after ordinary edits such as filters, color changes, and lossy saves.[2]

Stripping a caption, or wiping file metadata, is not the same move. The FAQ contrasts the watermark with visible logos and metadata, which can be lost.[2] A clean metadata panel does not tell you the watermark is gone. A missing watermark does not tell you a person shot the file.

The FAQ also says the mark is not infallible. If someone tries enough transformations, they may find one the Detector does not catch.[2] The page calls SynthID one layer in a wider approach to where a file came from, not a standalone guarantee.[2]

The Detector itself, on the page I fetched, checks three types. An image, for a watermark in the picture. A video, for segments that carry the watermark. An audio file, for segments that carry the watermark.[2] Text is in the FAQ's description of SynthID. It is not in the Detector's content-type list on that page. Do not upload a caption and expect this box to score it.

Read the two labels, then stop on the second one

The FAQ tells you how to read the result. Use its words in the caption. Do not soften them.

  • Watermark detected. A SynthID watermark is present. The file was likely edited or generated by a SynthID-enabled model.[2]
  • Watermark not detected. No confident signal was found, so it was not likely edited or generated by a SynthID-enabled model. The same note says this does not totally rule out AI generation or manipulation.[2]

A not-detected result can still be an AI file. The page names two reasons. The file was made by a model without SynthID, including a model from a company that has not adopted it, or a partner model from before they adopted the mark. Or heavy modifications degraded the signal.[2]

There is a rare miss in the other direction. The page says SynthID can, in very rare cases, falsely trigger on content that was not watermarked.[2] A detected label is a strong read. It is still one check. Collect another point of data if the post is going to make a hard claim about origin.[2]

Write the caption to match the label you got.

If the page says detected, you can say the file carries a SynthID watermark, and that the Detector read it as likely edited or generated by a SynthID-enabled model.[2] That is a disclosure. Say it if your brand rule says AI-assisted creative gets labeled.

If the page says not detected, you have a narrower sentence. This check did not find a confident SynthID signal. You do not have a sentence that says a person made the file.[2]

Do not point a script at the box

If the next idea is a publish job that drops a folder of ads on this page, stop. The additional terms forbid it.

  • Do not use the service, its outputs, or its feedback to run attacks meant to break, bypass, or degrade the watermark.[2]
  • Do not bypass the interface to submit requests, scrape the page, or use the frontend as an automated API for bulk checks.[2]
  • Do not open extra accounts, or dodge the rate limit, to get around the quota.[2]

Google says it can suspend access for those.[2] One editor, one file, before the post goes live. That is the use the upload box is built for. A job that scores the asset library is the use the terms name as off limits.

Know what you hand over. The privacy notice says media you upload is processed in real time, and the original files are deleted once the result comes back.[2] To enforce daily quotas, and to defend the service, Google stores a digital signature of the media, mapped to your user identifier, and deletes that signature after 24 hours.[2] That is the notice. I did not audit the deletion.

If the file is unreleased creative, read that notice before you upload the master. The page says the original is deleted after the result. It also says a signature of the file stays for a day, tied to the account identifier.[2] If that is more exposure than the brief allows, check a copy you are already willing to show, or wait until the file is public.

What to tell the person hitting publish

Give them the short version. It is all on the pages I fetched.

The Detector is open, in English, as of 7 October 2026.[1] Open synthid.com, agree if the terms are up, and drag in the file you would publish.[2] Detected means treat it as likely made or edited with a SynthID-enabled model, and label it if that is your rule.[2] Not detected means leave "human" out of the caption.[2] Do not automate the check.[2]

Then one more sentence for the person writing the brief, not for the caption. The partner list is on Kohli's post. Read it. A company that is not on that list is outside what this tool can see, and a miss on that file is the tool working.[1][2]

Check the file. Read the miss. Publish the sentence that matches the label you got.

Sources

[1] https://blog.google/innovation-and-ai/models-and-research/google-deepmind/synth-id-ai-content — Pushmeet Kohli, Google DeepMind, "We're making it easier to identify AI-generated content globally," 7 October 2026. Fetched 9 October 2026.

[2] https://synthid.com/ — SynthID Detector. Fetched 9 October 2026.