HN Explorer

Positive Text Origin Verification

2 points by Miguel_FierroM ↗ HN
Is anyone working on a positive text origin verification standard?

Everyday it gets harder to distinguish between AI generated content and human generated content.

This will never really work on the side of the labs but for people looking to actively prove they're the ones that wrote something there's an incentive to voluntarily collect that evidence.

Is anyone working on this?

8 comments

[ 0.70 ms ] story [ 32.2 ms ] thread
Juliate
Why should it be to the people to prove that their own work is not AI, and not the other way around?

EDIT: I mean, the effort has not to come from people that do the work by themselves. A new technology is not beneficial if it imposes more effort than before to do the same work.

Miguel_FierroM
I'm working under the assumption that there's still some value to real human interaction / media, which I do believe (not in a "end AI" way, exactly the opposite). There will always be an incentive to pass anything as human produced not only by the labs but by the people using these tools. As soon as Anthropic announced their watermarking strategy, there were people reverse engineering it and opensourcing tools to remove it.

In practice what has ended up happening is that in almost every surface people should assume that there's an increasing level of probability that that specific piece of media is produced by AI.

bbor
I think "standard" needs a lot more detail, here. To me that says "protocol", and I don't see a clear way a protocol could prove text provenance, as any protocol is ultimately only as good as what one side passes into it. Even if you were to somehow seal and stream keyboard events without invading privacy, that's just raising the bar for fraud -- a frontier model could whip up a python intermediary to simulate human keystroke speeds in 30 minutes or less.

I suppose the very hardest evidence to fake would be a video of the author typing, verified by AI (we'll soon lose the ability to distinguish on our own, if we haven't already).

Perhaps even better would be a keyboard with physical yubikey-esque technology... shit, that one might actually work. If someone has a robot dextrous enough to type, text provenance will be the least of our concerns, anyway!

I took a shot at something like what you're talking about, tho it's less "verification" and more "uncovering" information the creator is freely giving away. Undertone[1]: a local Python indexer plus a Firefox/Chrome extension that tints text by evidenced git provenance.[2] I've been too overwhelmed with the job search to finish it in any meaningful sense, but there's freely licensed scraps there you're/anyone is free to use if you'd like :)

But again, to be clear: it relies on good faith alone, and is intended for websites exclusively. Would never coherently work for non-VC'd prose (read: most prose).

[1]: What does it mean? I'm not sure. But it sounds provocative! It get's the people going!

[2]: https://gitlab.com/doering-ai/apps/undertone

Miguel_FierroM
Yes! This is interesting. I'll have a look.

And for sure it would need either the yubikey-esque keyboard and/or the adoption of a standardized origin system that publishing surfaces can verify against.

Seems like this will at some point need to be solved, or the way we interact with media will have to fundamentally change. It is an option at the end of the day. Certainly will be fun to see what happens.