Another researcher says OpenAI trained on conversations, then claimed breakthrou

Also discussed here: https://news.ycombinator.com/item?id=49638353

Original posts

https://mathstodon.xyz/@andreasthom/117240535270608201

https://mathstodon.xyz/@andreasthom/117240536885387540

https://mathstodon.xyz/@andreasthom/117240537520615623

This story is someone on bluesky screenshotting someone on twitter without a link. The person on twitter screenshotted the source on the fediverse without a link. This is getting daft.

I pay for the Pro ChatGPT plan, and if you go to settings > data controls this is the first setting:

> Improve the model for everyone

> Allow your content to be used to train our models, which makes ChatGPT better for you and everyone who uses it. We take steps to protect your privacy. Learn more.

It's on by default. We can debate whether or not it should be opt in or opt out, but no one should be surprised by this.

Here's the original Mastodon post:

https://mathstodon.xyz/@andreasthom/117240535270608201

So people genuinely believe that toggling that "Improve the model for everyone" button makes their data safe from being used for training?

How do people become that trusting?

The phrasing itself is guilt tripping

The author of the original mastodon post, Andreas Thom, acknowledged that he had not opted his data out of being used for training until June 29 of this year. He spends most of the post lashing out at OpenAI for not being transparent about whether his data was trained on (when the answer is obviously yes).

People need to understand how all these AI company policies around training data work before working with them, because it seems that people have no clue. Some things you should internalize:

  1. Opt your data out of training with the AI companies. There are multiple reasons why this isnt an airtight solution (see the following)

  2. Never press the feedback button. Once you do, your entire conversation will get slurped up, retained, and used in training data. This is especially important with coding agents because they can sometimes be too trigger-happy with a root directory find command, which can expose a *ton* of your personal data without you even knowing.

  3. Understand ai lab-specific policies. For instance, Anthropic / Claude Code has data opt-outs, but commits to keeping (for 7 years) and training on any of your chats that trigger their safety classifiers, even if they're false positives! Anyone remotely familiar with CC over the years understands how easy it is to trigger their safety classifiers.

  4. Providers of open models will not be any more charitable with the use of your data than the large US labs. For some reason, I've noticed here that people have a fairly loose security/IP posture around open-model providers because "I'm not doing anything important." It's very difficult to properly judge the importance of your data, and whether or not it can or will be used against you. The best posture is to always be more paranoid than less.

Another post that made it on the front page presented as fact that OpenAI "stole" the proof from Thom. There's no excuse to use one's own ignorance as a reason to fan the flames of anger towards AI companies. Like, we need to pump the brakes here because things are getting unnecessarily nasty, and it's not hard to imagine a mentally unwell person who sees stuff like this feeling motivated to do bad things.

If it is found that OpenAI and other labs are not respecting the training opt out then, I agree, there is reason to raise a commotion. But, with Thom and Buckmaster, accusations of malice are more better explained by incompetence (naivete).

edit: i'm sure i'm going to be accused of being some bot shill of the AI labs again but, people, this stuff all falls under the umbrella of common sense opsec.

Hopefully OpenAI can solve good problems where no one can make a claim that they stole their idea where the methodologies used and the techniques used are so out of the ordinary that the achievement is respected. Furthermore they should work on new novel solution on the old problems to lay these issues rest, thus by improving their models they can avoid issues of academics accusing them of using their work additionally the academics should also demonstrate their unpublished work is significant enough to have solved the problem . This is a bit subjective but it is also objective for the person with domain knowledge.

Under current understanding of the law, anything produced purely by LLMs (with no substantive human input, which is what OpenAI claimed in their post) is firmly in the public domain. So OpenAI can "claim" anything they want, it doesn't make it reality. In fact if I were the original authors I would just take their 400k lines of lean proof and relicense it under their own names/terms.

I think the heart of this issue is: people assume they have anonymity in numbers, but we have the tools to make it easy to scoop your data if it's interesting to the company.

Reminder that there are degrees of "trained on conversations". From John Schulman:

> pretrain on user data, with users' tokens as prediction targets: high regurgitation risk, improper

> use user prompts to distill large models into small ones: low regurg. risk, some companies probably do this

> use user traces to construct RL tasks: low regurg. risk, because RL has low memorization abilities, but can extract customer IP, depending on how it's done. Ranges from benign "use explicit user feedback in reward model training" to invasive "upload user's coding environment and commit history to turn into rl envs"

source: https://x.com/johnschulman2/status/2097440545853637108

How can OpenAI figure out how to be trustworthy?

What happens if you use a different harness?? Does opting out online suffice?

There are so many naive academics. They still believe an "opt-out" button.

Navier Stokes was solved by an internal model, so good luck proving it wasn't trained on Buckmaster/Lepöge or other chats.

Academics don't get that AI is a dirty tech bro industry that stole IP via torrents and runs after every surveillance contract it can get.

"Another researcher[/artist/writer/musician/programmer/doctor/director/etc] says OpenAI trained on conversations[/imagery/books/songs/code/classifications/videos/etc], then claimed breakthrou[gh/original art/bestselling books/chart-topping songs/unique applications/medical advice/free special effects/etc]"

Welcome to the party, with the rest of humanity.

the fix is boring and known: BIG-bench shipped a canary GUID for exactly this, and you publish your decontam n-gram threshold (gpt-3 used 13-grams). no threshold disclosed, no claim.

But who are you going to believe? Multiple independent academic researchers or the CEO who was fired two years ago for gross dishonesty?

Considering how Apple today announced that its products will contain a spying-on-conversation anti-feature by default, it seems reasonable to assume that all this spying is primarily done to train their AI model; and secondarily also to gather information about The People.

AI is becoming more evil by the day.

I think this is stupid, for three reasons:

1. The researches didn't actually have the breakthroughs. In the Navier-Stokes case they didn't solve the full problem, in this case too they didn't actually have the solution, they were experimenting with the methods.

2. Different OpenAI employees have come to out to say the only reason they can't definitively say no is that for privacy reasons they can't go see whether they actually did get any data out of a given user.

3. In any case, nobody at any point has suggested that opted-out user data was used for training. The author of the new tweet explicitly said they only opted out in late June, which is well after any RL on Sol would've ended (AFAICT OpenAI used 5.6 sol for those solutions)