SynthID Detector goes global: sampling and coverage limits after Bubble bulk catalogs via BIAIG

Context

On 2026-10-07, Google opened SynthID Detector to everyone at synthid.com (English UI). You upload images, video, or audio; the tool tells you whether a SynthID watermark from supported partners is present—not whether the content is “AI” in general. Named partners: Google, OpenAI, NVIDIA, Kakao; Apple listed as coming soon. Google’s deployment scale stats (180B+ images/videos, 240k years of audio) do not mean every asset in your catalog is verifiable. A closer look: The ~500-row bulk-tools ceiling in. A closer look: The ~500-row bulk-tools ceiling in.

I run Bubble bulk catalogs after BIAIG, and I treat synthid.com as a complementary QA layer, not a stand-in for OpenAI content_provenance_checks. I automate OpenAI-origin signals through the API (C2PA + SynthID on images; SynthID on audio). I use synthid.com for manual spot samples when vendors might include non-OpenAI SynthID partners, or C2PA got stripped on export/thumbnail while a watermark might still be there. Worth reading next: OpenAI Decisions API (preview): classification. Worth reading next: OpenAI Decisions API (preview): classification.

What results mean

Official synthid.com FAQ:

• Watermark detected — likely SynthID-enabled generation/editing from in-scope partners.
• Watermark not detected — no confident signal; not proof of human origin and does not rule out non-partner AI.

Video and audio can show watermarked segments. Compression, editing, and low-quality exports degrade the signal; Google acknowledges rare false positives. Sign-in is required (Google, Apple, or OpenAI accounts per secondary reporting).

Quota: Gemini Help documents ~10 checks per modality per 24-hour rolling window with file size/duration caps; press (Ars Technica, citing Google) describes a similar ~10-check/day cap on synthid.com and tighter limits on near-duplicate uploads (anti–watermark-removal rationale attributed to engineers). The public web UI is simpler than the pro tool (no region-level highlighting per Ars).

API: Scout review found no Google-published self-serve HTTP API for synthid.com—I do not plan Bubble/backend batch integration on an official Google endpoint. OpenAI stays the programmatic path for OpenAI-origin assets.

Post-batch QA workflow in Bubble

1. Split automation vs manual sampling

Path	When	Scope
OpenAI POST /v1/content_provenance_checks	OpenAI in your BIAIG pipeline	OpenAI signals; not_detected is inconclusive per docs
synthid.com (manual)	Multi-vendor SynthID, spot QA, suspected stripped C2PA	Partner SynthID only

I keep Gemini in-app verification (docs emphasize Google-only recognition) separate from multi-partner synthid.com.

2. Which SKU rows to sample after BIAIG

At ~10 checks/account/day, exhaustive verification is not feasible—an editorial/ops choice, not a vendor rule. My practical minimum:

• First row of each new SKU (hero image + primary video/audio if any).
• Variant rows after prompt/model/post-processing changes (new batch chunk).
• Ad-variant export candidates—100% sample within the daily quota or prioritize SKUs by revenue/risk.
• After aggressive catalog compression—retry on the highest-fidelity source (Google recommends highest-quality uploads).
• Mixed human+AI workflows—I do not auto-block on detected alone (false positive risk); I set a manual review flag.

I maintain an internal pipeline map (expected SynthID vs C2PA-only vs neither per row)—I do not invent coverage percentages without that mapping (open question in Scout packet).

3. Recording pass / fail / retry in Bubble

I add to catalog rows (or a QA subtable):

• provenance_openai_status — API enum (detected / not_detected / error / skipped_not_openai).
• synthid_spot_status — pass_detected | fail_not_detected_inconclusive | retry_quality | retry_quota | manual_review | not_sampled.
• synthid_checked_at, synthid_checker_account_id, synthid_source_file_url (master, not thumbnail).
• synthid_notes (segment timestamps for A/V).

Gate semantics:

• not_detected — never auto-clear for publish; inconclusive by design (Google FAQ; OpenAI API).
• retry_quality — re-export without extra compression; re-check master.
• retry_quota — queue for the next 24h window or an alternate verifier account only under explicit ops policy (multi-account may conflict with anti-abuse intent).

4. When to block catalog publish or ad-variant export

I block (hard stop or “Needs review”):

• OpenAI API error on mandatory OpenAI-origin rows without a documented fallback.
• Brand/compliance policy requires provenance and the row is not_sampled in a high-risk cohort.
• Repeated inconclusive after retry_quality on master.

I do not auto-block solely on:

• watermark detected in legitimate mixed pipelines (false positive risk).
• not_detected read as “AI-free” (coverage gap).

Ad-variant export: I keep a gate of “last spot-check ≤ N days” for sampled cohorts; for unsampled rows, I defer export or document accepted risk.

Risks (short)

• Many AI systems lack SynthID—a negative result must not authorize publish by itself.
• Third-party synthiddetector.com is not Google’s synthid.com—I do not mix endpoints.
• I avoid retry loops that probe watermark removal (Google rate-limit rationale per press).
• Detectors do not prove human origin; they do not replace licensing, consent, or brand QA.

Takeaway

After BIAIG bulk: automate OpenAI where the stack is OpenAI; statistical synthid.com sampling for multi-vendor and C2PA-stripped cases; in Bubble, honest pass/fail/retry states with inconclusive negatives; I gate publish/export on policy + repeated inconclusive, not on a misleading “AI-free” not_detected. Worth reading next: gpt-image-1 sunsets 23 Oct 2026:. Worth reading next: Batch API up to 50k.

Sources: Google blog 2026-10-07, synthid.com, DeepMind SynthID, Gemini Help quotas, OpenAI Content Provenance, Ars Technica.