AI Is Training on Your Old Posts: What You Can and Cannot Do

If you have posted publicly for years, assume your text has been collected. Large-scale crawls of public web and social content have been standard practice, and the collected data is generally not re-crawled to remove content that later disappears. That does not mean nothing can be done — it means the useful actions are narrower than most advice suggests.

What "training on your posts" actually means

Two separate things get conflated:

Collection — a crawler copies your public text into a dataset. This already happened.
Influence — the model's weights are affected by that text. This is diffuse; no one can point at your tweet inside a model.

Because the influence is diffuse, individual opt-out requests cannot surgically remove your contribution. But collection is the part that keeps happening, and that you can affect going forward.

What you can actually do

1. Control future crawling

If you own a site, a robots.txt disallow for AI crawlers is respected by the major ones that publish their user agents. For social posts you do not control this — the platform decides.

2. Reduce the public surface

The less of your text that stays public, the less there is to collect next quarter. This is the part the audit checklist is for. It is preventative, not corrective.

3. Submit opt-out forms where they exist

Several major labs publish opt-out mechanisms for personal data. They are worth submitting, with realistic expectations: they affect future collection and sometimes future model generations, not the already-shipped ones.

4. Request removal of identifiable output

If a model directly outputs your phone number, address or private content, that is a different and much stronger claim than "my opinion was in the training set". Use the provider's removal process for that specific case.

What does not work

Deleting posts to "remove them from the model" — the copy is already made.
Adding a copyright notice to your bio — it is not a technical control.
Assuming a private account retroactively hides old public posts from existing datasets.

The realistic frame

Treat public posting the way you treat speaking at a conference: it is on the record, and you cannot un-say it. The goal is not to remove what is already out there, but to make sure the next ten years of your public writing is something you would be comfortable having quoted.

常见问题

Can I get my data removed from a trained model?

Not in any reliable way. Machine unlearning is an active research problem, not a service you can request today.

Does deleting my account help?

It stops future collection from that account. It does not undo past crawls.

Is this a reason to stop posting?

No — it is a reason to post with the same judgement you would apply to any permanent, public medium.

结论

Collection is already done; influence is diffuse; future crawling is the part you can still steer. Spend your effort there instead of chasing the impossible reversal.