Fortunately, it's pretty simple in real life. We have certain publications and sources we trust, whether they're the NYT or a respected industry blog. We know they take accurate reporting seriously, fire journalists who are caught fabricating things, etc.
If we see a clip on YouTube from the BBC, we can trust it's almost certainly legit. If it's some crazy claim from a rando and you care whether it's real, it's easy to look it up to see if anyone credible has confirmed it.
So no, no worry at all about the past being erased.
This goes all the way back to yellow news with newspapers: https://en.wikipedia.org/wiki/Yellow_journalism
Imagine facebook decides to subtly change every public post and comment to show some particular person or cause in a better light.
We have tons of credible archived sources owned by different institutions. And these sources are successful in large part due to their credibility and trustworthiness.
It's just not economically rational for any of them to start "altering the past", and if they did, they'd be caught basically immediately and their reputation would be ruined.
This isn't an ML/tooling question, it's a question of humans and reputation and economic incentives.
Second, people take screenshots of Facebook posts all the time. They're everywhere. If you suddenly have a ton of people with timestamped screenshots from their phones that show Facebook has changed content, that's exactly the kind of story journalists will pounce on and verify.
The idea that Facebook could or would engage in widespread manipulation of past content and not get caught is just not realistic.
Maybe it is improbable, but there now is the technical possibility which was not there before.
It is valuable to explore that possibility and maybe even work to prevent such a use.
I would be interested in a ledger of cryptographically signed records of important public information such as newspapers, government communication and intellectual discourse.
Your argument that large social media will behave rationally is not backed up by reality. Consider Musk and Twitter.
Detection doesn't really matter, because people are too lazy to validate the facts, and reporters are not interested in reporting them. AI is simply another tool to manipulate people, like Wikipedia, Reddit.com, Twitter, or any other BS psuedo-authority. Think someone will actually crack open a book to prove the AI wrong? Not a chance.
You really think that if the NYT started altering its past stories, other publications would just... ignore it?
It would be a front-page scandal that the WaPo would be delighted to report on. As well as a hundred other news publications.
Thankfully.
If you can't alter world news headlines, you can still alter the tone of the article. If you can't alter front page news, you still can alter the remaining 95% of news.
Influencing public opinion is more subtle than the one important headline per day.
You are also ignoring the fact that news sites regularly edit published articles already, from fixed typos to corrections to large re-editings.
This isn't about a small percentage of stories, it's not about tone, it's the fact that if the NYT ever did this even once with the intention to truly "alter the past" it would be a major scandal.
And obviously things like corrections or taking down libelous content aren't included.
So no, I'm not constructing any kind of straw man here. I'm saying that the threat of subtly nefariously "altering the past" isn't realistic because it would be caught and exposed and there's no financial motivation to do it in the first place.
This is already happening without generative AI, and this new stuff is only going to speed things up exponentially.
The difference is that the floodgates are being opened.
We have lots of tools to fight spam, and there's no reason to believe they won't continue to evolve and work well.
Might not be possible on platforms - only if it's posted on a trusted domain.
That's the entire point of having trusted sources. Regular people can post whatever fake things they want on their own accounts; they can't post to the BBC's YouTube channel or to the NYT's website.
We haven't been able to generate 1,000 different forged variants of the same speech in a day before.
> We have certain publications and sources we trust, whether they're the NYT or a respected industry blog.
We can't even be sure that most of these aren't changing old stories, unless we notice and check archive.org, and they haven't had them deleted from the archive. The NYT has blockchain verification, but the reason nobody else does is because no one else wants to. They want to be free to change old stories.
You're wildly assuming a motive with zero evidence.
No, the reason companies aren't building blockchain verification of their stories is simply because it's expensive and complicated to do, for literally zero commercial benefit.
Archive.org already will prove any difference to you, and it's much easier to use/verify than any blockchain technology.
I'm saying this is going to become increasingly important fast, and we may miss the window where now almost everything not properly indexed by a large media organization is invalidated as there is no way to verify it.
I have a picture of Frank Sinatra at Disney World riding the tea cups. Who is the Frank Sinatra media authority that can tell me if this ever happened or not? A very small example to extrapolate from. It's going to get worse when everyone can create audio/video/pictures/text of anything they can dream.
The past may very well become a fictional dream, mythology, most of it impossible to verify.
Maybe it's not a big deal to 'lose' the past, maybe landfills will be mined for authentic content.
We'll likely be able to verify whether an entity is a real human, using some kind of "proof of humanity" system.
We will have cameras/mics with private keys built-in. The content can be signed as it's produced. But in this case, what's stopping me from recording a fake recording?
Maybe it's a non-issue. We used text to record history and we've been able to manipulate that since, well, forever.
*my favorite is always the nightclub scene that goes real quiet when the actors act using their voices (which are real, but may be dubbed in afterwards).
Yeah this is not true. Sota Text, Image generation is well above average baselines. You can certainly generate professional level art on Midjourney