There is no single “off” switch
There are plenty of ways to opt out of AI training, but they’re scattered, each covering one slice:
| Mechanism | Controls | Catch |
|---|---|---|
| robots.txt (AI UAs) | Which crawler may fetch which paths | Voluntary; small crawlers ignore it |
AIPREF Content-Usage | Standardized “train vs search” preference | Delivery draft expired 2026/5; no RFC yet |
noai / noimageai meta | Page-level “don’t train on this” | Non-standard; major vendors mostly honor it |
| W3C TDMRep | /.well-known training reservation, EU legal basis | Still a CG report |
trust.txt datatrainingallowed | Site-level training reservation | Not enforced by mainstream AI |
| IPTC best practices | Bundles the above into a publisher kit | Guidance, not new power |
| CC Signals | Reciprocity / attribution ask | Pilot, no teeth |
The shared catch: it’s all voluntary
None of these has legal or technical enforcement — they all rely on crawlers honoring them. Big vendors usually respect robots and noai; small crawlers scrape anyway. Even AIPREF, the one most trying to become a real standard, saw its deployable draft expire by May 2026. This layer is far from settled.
The biggest trap: over-blocking backfires
“Block training” and “keep being cited” are two different things. Many people, blocking training, also block the search/citation bots and vanish from AI answers without knowing (see Block the wrong bot). Before you act, decide: are you blocking “training,” or “everything”?
Pragmatic advice
If you really want to block training, do it well with the mechanisms big vendors honor (robots AI UAs + noai) and don’t pile on the rest; always allow the search bots. Treat the newer standards (TDMRep, CC Signals, IPTC) as good to know, not a reason to rebuild. Over-engineered opt-out is costly to maintain and easy to get wrong — which is exactly why it’s worth checking regularly rather than setting once and forgetting.
