4 ms·

I don't mean to be a vexatious asshole (I suppose my problem is that I can't stop myself), but if content that's created before 2021 is more valuable, what's to

I don't mean to be a vexatious asshole (I suppose my problem is that I can't stop myself), but if content that's created before 2021 is more valuable, what's to stop AIs from creating more of it via the power of lying? (Or call it a hallucination or whatever.)

snapcaster 3y ago
The whole source of the value increase is that it's verifiably non-AI generated content
- silverquiet 3y ago
  Ok, how do you verify that?
  - CoastalCoder 3y ago
    I agree with your skepticism. I think you're being realistic about the underlying adversarial dynamics. I.e., it's an arms race regarding recognizing AI authorship.
  - snapcaster 3y ago
    Because the capabilities didn't exist before then? Which part do you think is difficult to verify?
    - staindk 3y ago
      I can get an AI to write 50k bad articles and publish them on my site, with the publish dates set back to before 2021.
      - vik0 3y ago
        The first sentence of the article reads: >*Google Books is indexing low quality, AI-generated books that will turn up in search results* This was never about articles, it is about books, though. You can't retroactively publish a book in 2021 if we're not in 2021. It's bewildering to see how people here are so confident in their tone while arguing about things they think are said in the linked post, but don't know for sure because they never even opened it to read it.
        staindk 3y ago
        Topics of discussion change, in this thread I believe we're talking about AIs "hallucinating" "content". Not books specifically. You make a valid point though, which is actually more scary now that I think about it... it is becoming less and less possible for non-google (/other huge) companies to train models on verifiably human-generated content, as only Google (/other huge companies) have enough accurately timestamped data to train on. Hmm.
    - silverquiet 3y ago
      What capabilities didn't exist? The AIs I've interacted with create artifacts that are either text files or digital images, either of which can be printed out into physical manifestations. It's not like this is exactly a new phenomenon - Plato's famous "Allegory of the Cave" was a thought experiment in how the media we consume can create what we perceive as reality - it's just that we're discussing an uncomfortably recent intersection of philosophy with computer science. I said it a few times in the threads that blew up over Google's diverse Nazi bungle; I didn't get into computers to discuss anthropology and how humanity constructs meaning; I actually did it to avoid that and do work I considered purely practical after washing out of a liberal arts degree. That said, the whole thing ironically felt like very good art in the sense that it got people to discuss the nature of these generative AI systems. For me especially, the image of the Black Pope was very thought provoking as a (lapsed) Catholic; why is it that there has never been a black pope?
      - vik0 3y ago
        >why is it that there has never been a black pope? There were a few popes of African descent, though; don't know if their skin color happened to be black Still, I don't get why it would matter to any Catholic (I'm not Catholic), except to (some) Catholics in the US of A maybe, since race seems to reign supreme over there, even more than God Himself to (some, supposed) believers
        silverquiet 3y ago
        Well, this is a different rabbithole than I would have expected to go down, but why not? I have no idea of the Catholic experience outside the US and if that image means anything to them. I can get into my personal feelings if it matters; I don't think they matter much to anyone except me. The quick summary is that I dug the grave of a great-great Aunt of mine by hand at a small cemetery attached to the Catholic church in the tiny, forgettable town she was born in. One quirk of this rural, very white (I assume the entire population is a fairly close cousin of one type or another to me), Catholic town is that the priest is an African immigrant. My aunt was quite racist and so I thought it was a sort of ironic that the last group of people who saw her out of this world were twenty-five percent black. So an AI-generated image of a black pope made me stop and think for a minute about what Catholic identity means; a concept that as someone who was raised in that tradition but is unable to myself believe in the divine has always been problematic. Basically, it was good art, though I don't think that was the intent of the corporation that created it.
ADeerAppeared 3y ago
It's not the AI lying about the date, it's about the people using the content. One big problem is that Google also lies about dates when you do a 'before' search, so AI spam still shows up even if you filter it out. Scraping a new dataset without contaminated content is impossible now. There's no authoritative source on the age of any given website. (You can try places like the Internet Archive, but most such archivers aren't interested in handing their data over to AI firms because of the bad PR.) The only way to ensure you don't have AI content is to use a dataset created before ~2021-2022. This too is hard because all the big ones are illegal to posses as they contain CSAM. The real answer is that AI developers played themselves. Humanity was creating exponentially more data each year, and by rushing these AI tools out the door, all that data is useless now. By the end of this year there'll be more new "AI-polluted" data on the web made after 2021 than the total amount of "clean" data made up to 2021.
- Solvency 3y ago
  I don't buy this take for a second. You honestly don't think Google has a VAST and complete trove of time-stamped data scraped and indexed from decades of its existence. You don't have access to it. But they do.
  - lupire 3y ago
    People care a lot more about current events than past events.
  - withinboredom 3y ago
    In my experience, indicies tend to trust the date in the article more than the date it was discovered. That’s also how some people fight bad press: copy an article verbatim, set the date to the day before, file takedown notice for theft of content. By the time the actual content comes back up, it won’t matter.
- ancientworldnow 3y ago
  AI training doesn't care anymore. Huge amounts of it now are intentionally created synthetic data made by LLM's to generate data to train bigger LLM's. The larger the model, the less it matters.
burkaman 3y ago
A dishonest steel salesperson can also lie to you about when their steel is from. If you don't have a way of measuring the radiation, then you have to rely on trusted sources that you don't believe will lie to you, and not accept any steel from someone you don't know. It's the same here, we have many trusted sources that we don't believe would lie about when content is from - Internet Archive, newspaper archives, widely used web crawls, Wikipedia exports, etc. Obviously you don't use an AI (or a black box search engine) to get non-AI content, you get it from people and systems you trust.
- silverquiet 3y ago
  Yes, I see no other way than trusted humans; my problem is that I've worked with salespeople and there are vanishingly few that I trust.
  - wussboy 3y ago
    But at least we have evolution-provided mental models for dealing with trusts for humans.
    - lupire 3y ago
      Those modes that salespeople and PR people professionally exploit? We've been yelling about Web of Trust for decades by now, and never built it and now it's too late.

snapcaster 3y ago

The whole source of the value increase is that it's verifiably non-AI generated content

silverquiet 3y ago
Ok, how do you verify that?
- CoastalCoder 3y ago
  I agree with your skepticism. I think you're being realistic about the underlying adversarial dynamics. I.e., it's an arms race regarding recognizing AI authorship.
- snapcaster 3y ago
  Because the capabilities didn't exist before then? Which part do you think is difficult to verify?
  - staindk 3y ago
    I can get an AI to write 50k bad articles and publish them on my site, with the publish dates set back to before 2021.
    - vik0 3y ago
      The first sentence of the article reads: >*Google Books is indexing low quality, AI-generated books that will turn up in search results* This was never about articles, it is about books, though. You can't retroactively publish a book in 2021 if we're not in 2021. It's bewildering to see how people here are so confident in their tone while arguing about things they think are said in the linked post, but don't know for sure because they never even opened it to read it.
      - staindk 3y ago
        Topics of discussion change, in this thread I believe we're talking about AIs "hallucinating" "content". Not books specifically. You make a valid point though, which is actually more scary now that I think about it... it is becoming less and less possible for non-google (/other huge) companies to train models on verifiably human-generated content, as only Google (/other huge companies) have enough accurately timestamped data to train on. Hmm.
  - silverquiet 3y ago
    What capabilities didn't exist? The AIs I've interacted with create artifacts that are either text files or digital images, either of which can be printed out into physical manifestations. It's not like this is exactly a new phenomenon - Plato's famous "Allegory of the Cave" was a thought experiment in how the media we consume can create what we perceive as reality - it's just that we're discussing an uncomfortably recent intersection of philosophy with computer science. I said it a few times in the threads that blew up over Google's diverse Nazi bungle; I didn't get into computers to discuss anthropology and how humanity constructs meaning; I actually did it to avoid that and do work I considered purely practical after washing out of a liberal arts degree. That said, the whole thing ironically felt like very good art in the sense that it got people to discuss the nature of these generative AI systems. For me especially, the image of the Black Pope was very thought provoking as a (lapsed) Catholic; why is it that there has never been a black pope?
    - vik0 3y ago
      >why is it that there has never been a black pope? There were a few popes of African descent, though; don't know if their skin color happened to be black Still, I don't get why it would matter to any Catholic (I'm not Catholic), except to (some) Catholics in the US of A maybe, since race seems to reign supreme over there, even more than God Himself to (some, supposed) believers
      - silverquiet 3y ago
        Well, this is a different rabbithole than I would have expected to go down, but why not? I have no idea of the Catholic experience outside the US and if that image means anything to them. I can get into my personal feelings if it matters; I don't think they matter much to anyone except me. The quick summary is that I dug the grave of a great-great Aunt of mine by hand at a small cemetery attached to the Catholic church in the tiny, forgettable town she was born in. One quirk of this rural, very white (I assume the entire population is a fairly close cousin of one type or another to me), Catholic town is that the priest is an African immigrant. My aunt was quite racist and so I thought it was a sort of ironic that the last group of people who saw her out of this world were twenty-five percent black. So an AI-generated image of a black pope made me stop and think for a minute about what Catholic identity means; a concept that as someone who was raised in that tradition but is unable to myself believe in the divine has always been problematic. Basically, it was good art, though I don't think that was the intent of the corporation that created it.

ADeerAppeared 3y ago

It's not the AI lying about the date, it's about the people using the content. One big problem is that Google also lies about dates when you do a 'before' search, so AI spam still shows up even if you filter it out. Scraping a new dataset without contaminated content is impossible now. There's no authoritative source on the age of any given website. (You can try places like the Internet Archive, but most such archivers aren't interested in handing their data over to AI firms because of the bad PR.) The only way to ensure you don't have AI content is to use a dataset created before ~2021-2022. This too is hard because all the big ones are illegal to posses as they contain CSAM. The real answer is that AI developers played themselves. Humanity was creating exponentially more data each year, and by rushing these AI tools out the door, all that data is useless now. By the end of this year there'll be more new "AI-polluted" data on the web made after 2021 than the total amount of "clean" data made up to 2021.

Solvency 3y ago
I don't buy this take for a second. You honestly don't think Google has a VAST and complete trove of time-stamped data scraped and indexed from decades of its existence. You don't have access to it. But they do.
- lupire 3y ago
  People care a lot more about current events than past events.
- withinboredom 3y ago
  In my experience, indicies tend to trust the date in the article more than the date it was discovered. That’s also how some people fight bad press: copy an article verbatim, set the date to the day before, file takedown notice for theft of content. By the time the actual content comes back up, it won’t matter.
ancientworldnow 3y ago
AI training doesn't care anymore. Huge amounts of it now are intentionally created synthetic data made by LLM's to generate data to train bigger LLM's. The larger the model, the less it matters.

burkaman 3y ago

A dishonest steel salesperson can also lie to you about when their steel is from. If you don't have a way of measuring the radiation, then you have to rely on trusted sources that you don't believe will lie to you, and not accept any steel from someone you don't know. It's the same here, we have many trusted sources that we don't believe would lie about when content is from - Internet Archive, newspaper archives, widely used web crawls, Wikipedia exports, etc. Obviously you don't use an AI (or a black box search engine) to get non-AI content, you get it from people and systems you trust.

silverquiet 3y ago
Yes, I see no other way than trusted humans; my problem is that I've worked with salespeople and there are vanishingly few that I trust.
- wussboy 3y ago
  But at least we have evolution-provided mental models for dealing with trusts for humans.
  - lupire 3y ago
    Those modes that salespeople and PR people professionally exploit? We've been yelling about Web of Trust for decades by now, and never built it and now it's too late.