Training AI
lmnt.me
lmnt.me
As someone who is both a technologist but also has an art degree, I feel out of place in this conversation. Because, while I totally understand the economic anxiety and philosophical discomfort of this new technology – I also see it as a kind of Library of Alexandria. A synthesizer built on the collective knowledge output of all of humanity. It's incredible!
But, things didn't turn out so good for the first Library of Alexandria. Beware of mobs!
I feel frustration and in some situations, contempt: with both sides, with the situation, with the collective decisions, and lack thereof that got us here, with the hoi polloi getting distracted,yet again, by their corporate masters (1), by a system eternally being ransacked by executives; executives so inept and profoundly sociopathic that they aren't even aware there's a quiet part not to say out loud.
(The contempt isn't evenly distributed, fwiw)
(1) https://x.com/doctorow/status/1804240404986184072?t=2z7-WEwB...
Tech industry today is pretty much 1980s greed-is-good Wall Street screwing over everyone they can, and the main difference is that today we don't realize we're behaving like the stereotypical coked-up bad guys.
We do tend to have some socially progressive ideas among tech workers, which is great. But not great is if we anchor our self-perceptions to that do-gooder side, and then give ourselves a blank check for society-destroying greed and obliviousness in other regards. (Can we please respect people's pronouns and also not screw over their economics, privacy, thinking, government, and general wellbeing?)
I recently came to realization that the 80s were able to be fueled by cocaine because we lacked the tools to identify a coked out maniac. People must have just seen these people as "engaged, enthusiastic, and passionate".
Now, we have the tools to see the current snake oil salesmen (this is what? our nth cycle? VR, GigEconomy, Blockchain, AR, ...) but people are still defending them. It's difficult to watch again and again.
It would be incredible if it existed in an egalitarian, post-capitalistic climate rather than one where it deprives people of their livelihood.
Unsure how "beware of mobs!" fits in there.
It's as if the comment is intentionally written to be inflammatory. I suppose that is unfortunate.
US Copyright law (which assigns that very copyright) disagrees with this. "Fair use" means you can use it, in certain transformative ways, not merely "access" it.
I think it is still very unclear though whether this could have any effect, or if is going to be a dud.
One weakness is that it does not specify how. There are a number of opting-out protocols out there, and with variety it follows of course that while some site might use some protocol/s, a robot might support another. For example I could find that Cara.app uses DeviantArt's "noai" tag only in its HTML headers, but when I read through the source of Spawning.ai's "datadiligence" library [1], I found that it checks for that tag only in HTTP headers. So even in those cases when the company scraping the web for content has intent to comply with EU law, there is still the risk that they could be violating it.
I'm in the latter camp, so I don't see why anyone would be OK with their content being seen by humans but not used to train an AI.
The question is whether it should be illegal even if you never profit from it, or somehow dilute the IP owner's brand.
I dont think this is clear to others given the parent comment answer Yes instead of No.
How much of the work you use in your transformations is a major factor for derivative works. So you probably could make a single Star Wars scene from memory using Blender. But not the whole movie.
[1] https://www.americanbar.org/groups/intellectual_property_law...
Morally, legally, socially, these are entirely different.
None of this is about the science. All of it is about the money.
The difference between taking inspiration from other work and copying that work relies on the actions and intent, but in the case of AI both of those are obfuscated and we only see the end result.
I think this makes it very difficult to gauge whether particular AI work is problematic.
It absolutely does not. Bunch of companies with pretty low ethical standards are scraping all they can find. People who wrote all this content never consented to this, and many of them explicitly opted-out.
- 'We own what we make the moment we make it.' ... you hold a copyright.
- 'Publishing text or images on the web does not make it fair game to train AI on'. ... postulate without further grounding elsewhere in the text.
- ' The “public” in “public web” means free to access; it does not mean it's free to use.' . Actually it is, barring restrictions on copying and trademarks etc as per (implementations of) the Berne convention.
- 'Someone who publishes my work as their own (theft)'. No, this is a) a copyright violation or b) plagiarism.
- 'or republishes my work (like quoting or linking back)' This depends, but linking is not (re)publication of your work. Quoting counts as fair use copying.
- 'doesn’t have the right to make the choice for me to let my content be used for training AI'. A link is not your content, you have no rights to it period. If someone makes a fair use quote of your content, then they may do so under fair use as a quote. I don't really see "quoting" as reposting, since that implies a substantial amount of copying. If it is not substantial, an AI would be permitted to use the quote in a similar manner. However, since an AI training system does not typically retain a copy of your quote, copyright probably doesn't apply at all (exceptions excepted).
- "Whether reposting my content elsewhere is in good faith or not". Previous examples were not of reposting (probably) but if they count as reposting, then you have standing to sue the person doing so either way. AI training is a red herring here.
- "To add insult to injury, that person may not have the knowledge—or even the power—to do so if they’re posting content they don’t own on a site they also don’t own, like social media.". Let's make a difference between ownership and holding copyrights. If you post things, you must hold the copyright, or have permission from the copyright holder. If you do not hold the copyright holder, posting content to a site you do not own is not permitted by law, and very likely a violation of the TOS of the site where you posted.
- "I can play whac-a-mole with those bots on servers I control—which I don’t like doing, for the record—but I have none of that control anywhere else." ... sure, but you seem to be pulling together disparate concepts and several misunderstandings of law here to begin with. Your conclusion(s) do not follow from the premises, given real world laws, treaties, and customs.
Fair enough, but what does that have to do with the default value of robots.txt? Opt-in or opt-out, they’re still making the choice for you.
"Ownership" is completely opposite to a natural state of things, and can only be enforced with force. Mix that abstract concept with the abstract of information for extra complexity. Add to it that Information can't be "owned" in the same sense that physical property can, and the fact that (if not copyrighed), it's a non-excludable, non-rivalrous good (it can benefit everyone, equally, at the same time).
On the other hand, it's nice to incentivize innovation, competition and creativity, or at least, not hinder it.
It would be very interesting to go to 2124 and check if these concepts are still alive, or if somehow we have evolved our social norms to be closer to the natural world. Once you convey information to any entity outside of your brain/body, it naturally belongs to the world.
A right is not ownership.
> "Someone who publishes my work as their own (theft) or republishes my work (like quoting or linking back) doesn’t have the right to make the choice for me to let my content be used for educating a student. This is where I struggle the most with the “opt-out” style of student education on the web."
Imagine a student who gets into a top-ranked research university, studies lots of material from proprietary textbooks, takes in hours of instruction from professors who have strict IP contracts with their university, and then graduates - can they then go produce for-profit content built on that knowledge base without paying a fraction of their earnings to the university that trained them?
What a joke.
Right, right, in 1989 Tim Berners-Lee and team all said to eachother, "We build this thing to learn from, even if the learner isn't human."
And lo, the world wide web was born.
But Zoom was about using videos for AI training, everything else is about text. And it looks like nobody really cares about text, except for maybe NY Times which is suing OpenAI.
What is the deal with this? It pisses me off that a few major companies are about to gobble up the entire internet and capitalise on it solely for themselves. Not going to give back anything to anyone, they even want you to pay a monthly fee just to get access to your own data.
It benefits millions of people who use these models every day, even the most powerful ones like GPT-4o and Claude 3.5 Sonnet are available for free to the public (with some rate limit).
Not saying I didn't pirate stuff or wouldn't do it today, or calling out this particular author. Just recognizing how people by and large are simply self-interested while wearing a cloak of morality.
I personally believe large corporations wont be able to control the productive AI market without regulatory capture. Creating models is just too easy once you have the data and the compute. I see all these pushes for regulation as red herrings that will likely result in cementing the current leaders in place rather than protecting the people that created the IP that went into their products.
The unfortunate truth is that the large corporations were going to do this with or without the legislation - that's a separate largely independently issue.
At the core, scale and control over data ultimately make or break these technologies, and large players (hiring the few talents capable of properly utilizing these technologies) are simply better equipped to leverage the markets.
The difference being, intent and public commons. At least with the legislation, public commons do not become fair game for corporate plundering, just as the public cannot (legally) plunder the "corporate" data.
Without the legislation, the corporate simply own and control all your data, without much explicit work on your part both technically and legally to defend your rights.
> I'm part of the limewire and BitTorrent generation
Myself as well - however this is where intent comes into play. Bringing data into the "public" commons as quickly and often as possible is most useful for the public, not corporations. Obviously, some of those public are in the employ of the corporate, however generally the intent of said generation was to "pirate, then buy". All studies on the matter showed this was true - free advertising and increased sales (for large scale ventures at least). Smaller vendors had trouble recouping costs, it is true, however attempts were made to build in compensation mechanisms for these cases as well.
In both cases, what is common is the intent to most broadly benefit the public, either through legislation (which is intended to serve the public interests) or through broadening the commons (which is intended to serve the public interests). In both cases, the technology employed suffered(suffers) gaping harmful holes that may not be possible to address except by some form of legislation/recompense.
The fact that corporations can and will exploit legal frameworks should not however, be a reason to think of all such efforts as "red herrings".
But time moves on, the internet skews younger, and Gen Z is all the rage now, and starting to set talking points. I think when all your music has always been streamed for free or like $8/mo, playing by the rules feels different.
But I am curious how many former "intellectual property isn't property" folks there are that are now on the side of "scraping web content to train a model is immoral".
I strongly disagree with this. The public web is as much "free to publish" as "free to use." If you don't want people to have unrestricted access to it, either put it behind a paywall, throttle access server side, or just don't publish at all.
The whole copyright system is already a step too far in the modern age when the marginal cost of personal copying is zero, I refuse to further give ground on freedom of association, which is what the internet is at its core. If someone republishes your copyright content, you can use the existing legal mechanisms to get it taken down and recover damages, but if you don't want people to use the content, don't publish it openly on the internet.
They usually have a clause around getting a license to all the content posted on their site and being able to sub-license/modify it/create derivative work.
Doesn't that basically mean once content is posted there. They can do what ever they want with it? What happens in those cases when the creator has given permission(although unknowingly perhaps)?
Can we expect any changes in this area at all? Can anyone point me to some resources?
Standardisation: Not really. There exist a bunch of novel protocols, but apart from robots.txt (which can only name robots you know), there isn't any established standard. The Open Future think-tank has posted a couple overviews of these [3] [4].
[1] https://www.europarl.europa.eu/doceo/document/TA-9-2024-0138...
[2] https://eur-lex.europa.eu/eli/dir/2019/790/oj
[3] https://openfuture.eu/publication/defining-best-practices-fo...
[4] https://openfuture.pubpub.org/pub/considerations-for-impleme...
This statement is false in so many circumstances: ‘The “public” in “public web” means free to access; it does not mean it's free to use.’ What is access without use even?
Because there are plenty of ways I can use that public content without the copyright holder having any say. It could be satire, quoting sizable sections of text, making a parody, teaching, news, etc.
Maybe you don’t think AI training should count as fair use, but there are countless ways already to create transformative/derivative works of from others copyrighted works without their consent.
This isn’t anywhere close to the black and white problem so many make it out to be.