Hacker News Reader

Comment by bluepuma77 | Hacker News Reader

Interesting how everyone says „But I can google you“ instead of thinking about the issue.

Companies are building and selling GPT-3 with 6 billion parameters and one of those „parameters“ seems to be OP‘s username and his „strange“ two word last name.

If models grow bigger, they will potentially contain personal information about everyone of us.

If you can get yourself removed from search indices, shouldn’t there be a way for AI models, too?

Another thought: do we need new licenses (GPL, MIT, etc.) which disallow the use for (for-profit) AI training?

View on HN

The FTC has a method for dealing with this: they have in the past year or two ordered companies with ML models built from the personal information of minors to completely delete their models.

I asked this question a few days ago; I just wanted to say thanks for answering. Some companies can find themselves without a business model if they handle that badly.

I agree. There is no reason ML tech should perform worse than traditional software, allowing privacy as per GDPR and CA regulations, at the very least.

The input datasets should be managed as per GDPR/CA regulations, with clear flags protecting privacy of EU citizens and CA residents. And any derived models should propagate these labels and not allow querying information violating these regulations.

If GitHub Colab implementation or and GPT-3/4 models were developed without these regulations in mind these models should be retrained.

Yes, it is a hard research problem. Yet, there is no reason these models should be allowed to violate privacy in worse ways than traditional software.

Is it really that different than a search engine? Take away the AI specific language and you have two products that when given his username return results with his real name.

Another commenter pointed out that a lot of these models aren't publicly accessible, but will still be used to retrieve information about you - by say employers contracting with a ML company

But they can only be used to retrieve information that is already out there. This is still just using GPT-3 as a search engine, it's just a weird search engine that isn't made to purpose and most of the time produces nice-looking nonsense instead of valid data.

It can also be used to retrieve deleted/delisted information. It's not like an search engine, more like an indexed database.

That's a different issue. And something that can be easier done by purpose-built non-GPT systems.

Nobody said it's easier or more optimal with GPT-3, the problem is that it's possible at all.

If that's the case, it means that GPT-3 doesn't just raise ethical questions, but legal ones as well: several jurisdictions around the world currently require that search engines allow for the erasure of private information upon request.

With classic search engine indexing you can find and remove exact matches from the index, but with neural networks it's harder to make sure you removed every representation of a specific information from the parameters. For example you remove somehow the exact username-name from the model parameters ( that doesn't seems to hard at first) but then it may still return the information if somebody ask the model differently.

So if you try to remove the information from a neural network model then it can still have it in different forms you may not even think of, for example in language models the same thing described with different words.

And on the other hand removing one thing may affect the models performance on other unrelated things too.

well, probably it's time to tag pieces of data, so it's possible to block certain results based on where the data originated.

> Another thought: do we need new licenses (GPL, MIT, etc.) which disallow the use for (for-profit) AI training?

I don't think that we need new licenses, but probably open source projects need a better way to enforce them.

E.g. Copilot just ignores the licensing issues although I can imagine that there could be a solution with a few different models that return code for different purposes. (Like one model returns everything and the code can be used safely only for learning or hobby projects. Another model returns code for GPL code. And a third model returns code compatible with commercial or permissive open source projects.)

Or the model spits out also the licence(s) of the code, but not sure if this is technically possible.

The information is embedded in the weights of various layers in the network. Trying to remove that information by editing weights would be like trying to alter someone’s memory by tinkering with synapses.

The only way to be completely sure of removing information would be to re-train the model without that data.

> If you can get yourself removed from search indices, shouldn’t there be a way for AI models, too?

Absolutely yes!

Reply on news.ycombinator.com