Before we get too upset... can we verify this is MSFT's official position? I suspect this may be hyperbole. It could be Sulyman was constructing a hypothetical point that didn't survive translation into click-bait. That being said... MSFT has a history of chicanery. I'm off to try to find original sources. If anyone else has any, please provide a link.
FWIW... I found a few videos related to Endicott's story:
* This is a quick 5 minute video where Suleyman talks about how indeterminacy is good. So... you know... it's a good think that Co-Pilot can't tell you why it thinks it needs to dump 800 line of java code into your hello world program. At around 3:44, he confuses LLMs (with a surface understanding of syntax married with a markov chain on steroids) with people (who as best we can tell have a different understanding of the thing represented.) Corporate management confusing the the map with the territory? Who could have forseen such a thing: https://youtu.be/GsGFYoIx1YM
* This one seems to be the longer version, but I'm still looking for where Endicott's quote comes from, but around the 14minute mark is where the conversation turns towards "who owns the ip" used to train LLMs and the terms "Fair Use" and "Freeware" are used around the 14m50s mark: https://youtu.be/lPvqvt55l3A
[EDIT: So... yes... get out the pitch-forks... Microsoft is saying anything on the web is inherently freeware or subject to fair use even if you think you remember putting a copyright notice on it (or, as is mentioned in US copyright law, the creator automatically receives copyright protections upon creation of the work.)]
It’s a pro-AI position but not really controversial?
My reading is he is saying content that is not under an explicit license for usage, that is made available publicly and freely, is fair game for training.
> In his remarks, Suleyman claimed that all content shared on the web is available to be used for AI training unless a content producer says otherwise specifically.
> "With respect to content that is already on the open web, the social contract of that content since the 90s has been that it is fair use. Anyone can copy it, recreate with it, reproduce with it. That has been freeware, if you like. That's been the understanding," said Suleyman.
> "There's a separate category where a website or a publisher or a news organization had explicitly said, 'do not scrape or crawl me for any other reason than indexing me so that other people can find that content.' That's a gray area and I think that's going to work its way through the courts."
> My reading is he is saying content that is not under an explicit license for usage, that is made available publicly and freely, is fair game for training.
In the absence of an explicit copyright, the default copyright is "all rights reserved". Putting something in a public space is not license to reproduce it, however much I might disagree with this position
I wish someone would define what scraping "open web" means since physical media has been dead for a while. YouTube content? Doxxing people? Archiving books and old video games? Bank transactions? VoIP calls?
Why is it potentially ok for AI to copy these things, but if a human does it they get into trouble?
We need an exhaustive list and that's probably impossible to define.
The courts ruled that content can be scraped from the open web legally. From your examples
1. YouTube content: ok because the scraped content becomes derivative works which falls under free use
2. Doxxing people: if the data was already on the web and a human found it then the person was already doxed
3. Bank transactions: are they behind a login wall? Not Ok to scrap
4. VoIP calls: again behind a wall? Not Ok to scrape
If a human can find it online without doing anything beyond accessing a public web link then it is Ok to scrape it. It's not the complicated and the law has a series of frameworks that for existing cases. It's not particularly complicated.
Derivative works are not automatically fair use. Otherwise I could release an album with a Prince song and a Taylor Swift song, claim it's a derivative work and covered under fair use. Again. Try to do this on YouTube (even though YouTube is not the arbiter of what is and isn't fair use, they're just aggressive about removing content.)
Doxxing people can give cause to restraint if it's inciting speech. Consider a situation where you say "Hey. I just found Joe Schmo's Address. He lives at 123 Main St., Anytown, DK 10010." Now compare it to when you say "Hey. I just found Joe Schmo's Address. He lives at 123 Main St., Anytown, DK 10010. Joe's a child molester. Let's go burn down his house." or if you call up the local police department and said "Hey. I saw Joe Schmo with Goody Parker in their house at 123 Main St., Anytown, DK 10010. They're molesting a child right now! Hurry, if you're fast you might stop them!"
Yeah... I don't remember saying bank transactions were copyrighted or in the public domain. Probably someone elses' comment.
VoIP calls? uh... you may want to review the Electronic Communications Privacy Act of 1986, 18USC2510-2522 before assuming VoIP calls can be intercepted for the purpose of training an AI.
You are free to believe anything you can access is fair use. (I mean, maybe in Iceland or Sweden.) But give me a call if you get caught doing this, I can refer you to a decent attorney. (Again... this all assumes you're in the states. International copyright is sort of a wild west. But you would expect MSFT to open up a fully owned offshore subsidiary in a country like Vanuatu or Iceland to evade continental copyright (what is it with islands and copyright???)
It’s very complicated, and if a lawyer tells you it’s not you need to get a new lawyer. Whether something is fair use is determined based on a number of different factors that all are weighed. There’s not some simple if-then algorithm to tell you the answer. There is case law which sets up some known points in the space, but if you walk off that manifold, no one knows how things will be weighed.
LLMs are off the manifold in a lot of ways. We’ve never had a form of information digestion/encoding quite like them.
Politically, both content creators and AI companies see this decision as existential. Ruling one way or the other is actually almost unthinkable. The only solution is going to be some compromise that comes from Congress, because the sorts of things that would be in such a compromise can’t be imposed by the courts with existing legislation.
No, the courts ruled that in the case of a searchengine, scraping was ok because the results are displayed to the user who then clicks on a link that brings them back to the original source. Seems like many in the web development world like to ignore that second part.
AI responses do not bring the user back to the original source/site, the therefore the data use isn’t in line with that reasoning, and thus is a copyright violation.
Fair use is not granted by a social contract, as he hopes. It's granted by courts. Suleyman's interpretation is the most self-serving possible opinion of the circumstances.
As Joshua Topolski put it, "Only someone utterly divorced from creation and completely insulated in a bubble of entitlement would have such a distorted view of other people's work like this" [0]
What nonsense. You can download it, view it, create a search engine index with it. You cannot "recreate with it", "reproduce with it".
Interesting that Microsoft openly rejects the DMCA. So we now can all reverse engineer and make available for free all Windows versions, since they have been on the web at some point. I guess Suleyman's statement can be used in court against Microsoft in the next IP trial.
Microsoft is head-quartered in Seattle, which is part of the United States, so it's not unreasonable for harmed parties to bring action under US Copyright law. US Copyright law is explicit about this. Copyright exists upon creation of the work unless specifically disclaimed.
You can reasonably assume that by posting content to the intarwebs, the author provided an implicit license to view the content (for the community that could reasonably be identified as being able to view the work.) But this does not imply an automatic "freeware" or "fair use" exemption.
How can Suleyman say someone who published something on the web in 1995 intended for it to be used to train their LLM in 2020? It's unreasonable to assume that.
There's a reason you see people put explicit licenses on things, it's because they don't exist unless the work's creator releases the work under a specific license. And license for a human to read something on the web is not a license to give your content to a large company for them to remix it into... whatever the heck it is they're remixing it into.
But yeah. It's probably worth it to brush up on robots.txt definitions to exclude spiders from scraping your site if they're going to use it as LLM fodder.
It's ironic that Microsoft used copyright protection and IP law for years to secure a dominant market position, and now they don't need to play by the same rules because "something something AI".
Truthfully, I agree with this new stance of theirs and believe it's always been the case, but there's no doubt that they're only adopting it now because it's become advantageous.
Sure. But there's a fair amount of US Copyright law that says this has never been the case. To test this hypothesis, upload a video to YouTube with music by Prince playing in the background.
Copyright law is more of an ideal, and if it had "never been the case", there wouldn't have been any content to DMCA off of YouTube in the first place. US copyright law is fighting the tide, not conveying a natural order.
Someone's reasons for sharing information are coloured by the situation at the time of sharing it, amongst many other factors.
Two years ago (say) no one predicted the meteoric rise of LLMs and their voracious appetite for data sets for training. These beasties are not simply search engines that are better direction pointers to your stuff (with a frisson of ads) but insist on being the final word and keep you out. To be blunt: It is stealing.
The implied contract for publishing on the web has changed again, just as it has several times in the past. The worst thing here is the use of the term "freeware". Describing original content, displayed for all to see as -ware is outrageous.
They might as well describe the content on Spotify and co as freeware ... bear with me: you could scrape wifi connections through your publicly available APs or even do some more broadband funky spectrum capture analysis and claim that is what an internet search engine does in its spare time and all is fine (lol).
LLMs and GenAI are quite interesting things but I do not think that they are the last word in ... AI. Anyway the latest cool thingie cannot be allowed to break whatever the current unspoken and somewhat undefined social contract is in place.
This bloke from MS seems to have forgotten that there really is a social contract of some sort and that if you say: "fuck you lot, omnomnom ... mmmm data ... ... laters (lol)" there might be some come back.
I make an http _REQUEST_, the server voluntarily fulfills the request.
Why is it okay for a person to view your content, memorize it, and use it as a base for new content while it's not okay for an AI? at the end of the day it is the same thing.
Because it's never been okay to reproduce something as the base of "new" content that competes with the original work. It's pretty much why there is copyright. I merely remembered your painting with my chemical film memory machine and then sold prints of my memories.
> Why is it okay for a person to view your content, memorize it, and use it as a base for new content while it's not okay for an AI? at the end of the day it is the same thing.
Because it is not really AI. We can have this comparison when it actually can think. Until then this 'training' is just big data processing and that is not what person can do by reading/watching.
You don't even have to make a request to hear a radio broadcast or see a tv broadcast, they are actively broadcast and bathe you whether you wanted it or not.
Now imagine setting up a service, even a free one where you don't make any commercial gain from it, where people ask you for stuff, and you give them bits of the stuff you've collected that was just literally falling from the sky. You give them the stuff not in the form of quotes for discussion, but presented as content itself, devoid of any prior context.
The user asks for some sci fi, and you give them 15 minutes of Star Wars which you got from a tv broadcast. But you don't tell the user "This is Star Wars, written by George Lucas" You just give them the content and the user thinks you created it.
Or worse, the user thinks THEY created it, merely using your tool, just a fancier version of using a spell checker, they are still the author since they directed the tool.
The user asks for a deep insight into human nature, you give them something that some famous author said, but not the authors name. The user then publishes a book containing this gem of deep insight, and other users think they are deeply insightful and buy their book.
This is not what happens when a person learns from life, including reading famous books, and then starts producing their own content.
The fact that it's possible to construct some scenarios where it seems like all the essential elements are the same does not actually make the two things equivalent.
Here's my own little insight into human nature I'll offer: It's pretty common for some people who pride themselves on being smart, for some mysterious reason to only be selectively smart. They will construct good valid logic that serves one purpose readily, and yet somehow fail to construct or even percieve equally valid logic that serves some other purpose.
If one is smart enough to try to present this logical argument that it's ok to steal other peiple's work, I submit that they are then also smart enough to be able to work out how and why that is in fact stealing.
We don't even have to get into the obvious, how MS themselves "broadcast" Windows to everyone by getting manufacturers to preinstall it on every machine. That wasn't even an http request. You just wanted to buy a piece of hardware, and it just came with Windows whether you wanted it or not. Even if you remove it, it was already given to you before you opened the box, and even after you wipe it, it's licence key is still there literally burned into the hardware. Therefor, since MS put it out there themselves, put it into your hands themselves, you are free to reproduce Windows. Oh, just pieces? Pieces are fair use? So, just a handy dll then. That must be ok.
There is no end to the ways this argument doesn't hold water.
28 comments
[ 2.8 ms ] story [ 47.7 ms ] threadFWIW... I found a few videos related to Endicott's story:
* This is a quick 5 minute video where Suleyman talks about how indeterminacy is good. So... you know... it's a good think that Co-Pilot can't tell you why it thinks it needs to dump 800 line of java code into your hello world program. At around 3:44, he confuses LLMs (with a surface understanding of syntax married with a markov chain on steroids) with people (who as best we can tell have a different understanding of the thing represented.) Corporate management confusing the the map with the territory? Who could have forseen such a thing: https://youtu.be/GsGFYoIx1YM
* This one seems to be the longer version, but I'm still looking for where Endicott's quote comes from, but around the 14minute mark is where the conversation turns towards "who owns the ip" used to train LLMs and the terms "Fair Use" and "Freeware" are used around the 14m50s mark: https://youtu.be/lPvqvt55l3A
[EDIT: So... yes... get out the pitch-forks... Microsoft is saying anything on the web is inherently freeware or subject to fair use even if you think you remember putting a copyright notice on it (or, as is mentioned in US copyright law, the creator automatically receives copyright protections upon creation of the work.)]
Obviously Microsoft is a huge company so he doesn't represent everyone, that would be impossible. But it's a massive red flag.
My reading is he is saying content that is not under an explicit license for usage, that is made available publicly and freely, is fair game for training.
> In his remarks, Suleyman claimed that all content shared on the web is available to be used for AI training unless a content producer says otherwise specifically.
> "With respect to content that is already on the open web, the social contract of that content since the 90s has been that it is fair use. Anyone can copy it, recreate with it, reproduce with it. That has been freeware, if you like. That's been the understanding," said Suleyman.
> "There's a separate category where a website or a publisher or a news organization had explicitly said, 'do not scrape or crawl me for any other reason than indexing me so that other people can find that content.' That's a gray area and I think that's going to work its way through the courts."
In the absence of an explicit copyright, the default copyright is "all rights reserved". Putting something in a public space is not license to reproduce it, however much I might disagree with this position
Why is it potentially ok for AI to copy these things, but if a human does it they get into trouble?
We need an exhaustive list and that's probably impossible to define.
If a human can find it online without doing anything beyond accessing a public web link then it is Ok to scrape it. It's not the complicated and the law has a series of frameworks that for existing cases. It's not particularly complicated.
Doxxing people can give cause to restraint if it's inciting speech. Consider a situation where you say "Hey. I just found Joe Schmo's Address. He lives at 123 Main St., Anytown, DK 10010." Now compare it to when you say "Hey. I just found Joe Schmo's Address. He lives at 123 Main St., Anytown, DK 10010. Joe's a child molester. Let's go burn down his house." or if you call up the local police department and said "Hey. I saw Joe Schmo with Goody Parker in their house at 123 Main St., Anytown, DK 10010. They're molesting a child right now! Hurry, if you're fast you might stop them!"
Yeah... I don't remember saying bank transactions were copyrighted or in the public domain. Probably someone elses' comment.
VoIP calls? uh... you may want to review the Electronic Communications Privacy Act of 1986, 18USC2510-2522 before assuming VoIP calls can be intercepted for the purpose of training an AI.
You are free to believe anything you can access is fair use. (I mean, maybe in Iceland or Sweden.) But give me a call if you get caught doing this, I can refer you to a decent attorney. (Again... this all assumes you're in the states. International copyright is sort of a wild west. But you would expect MSFT to open up a fully owned offshore subsidiary in a country like Vanuatu or Iceland to evade continental copyright (what is it with islands and copyright???)
LLMs are off the manifold in a lot of ways. We’ve never had a form of information digestion/encoding quite like them.
Politically, both content creators and AI companies see this decision as existential. Ruling one way or the other is actually almost unthinkable. The only solution is going to be some compromise that comes from Congress, because the sorts of things that would be in such a compromise can’t be imposed by the courts with existing legislation.
AI responses do not bring the user back to the original source/site, the therefore the data use isn’t in line with that reasoning, and thus is a copyright violation.
As Joshua Topolski put it, "Only someone utterly divorced from creation and completely insulated in a bubble of entitlement would have such a distorted view of other people's work like this" [0]
[0] https://x.com/joshuatopolsky/status/1806796270402699553?s=19
Interesting that Microsoft openly rejects the DMCA. So we now can all reverse engineer and make available for free all Windows versions, since they have been on the web at some point. I guess Suleyman's statement can be used in court against Microsoft in the next IP trial.
You can reasonably assume that by posting content to the intarwebs, the author provided an implicit license to view the content (for the community that could reasonably be identified as being able to view the work.) But this does not imply an automatic "freeware" or "fair use" exemption.
How can Suleyman say someone who published something on the web in 1995 intended for it to be used to train their LLM in 2020? It's unreasonable to assume that.
There's a reason you see people put explicit licenses on things, it's because they don't exist unless the work's creator releases the work under a specific license. And license for a human to read something on the web is not a license to give your content to a large company for them to remix it into... whatever the heck it is they're remixing it into.
But yeah. It's probably worth it to brush up on robots.txt definitions to exclude spiders from scraping your site if they're going to use it as LLM fodder.
Two years ago (say) no one predicted the meteoric rise of LLMs and their voracious appetite for data sets for training. These beasties are not simply search engines that are better direction pointers to your stuff (with a frisson of ads) but insist on being the final word and keep you out. To be blunt: It is stealing.
The implied contract for publishing on the web has changed again, just as it has several times in the past. The worst thing here is the use of the term "freeware". Describing original content, displayed for all to see as -ware is outrageous.
They might as well describe the content on Spotify and co as freeware ... bear with me: you could scrape wifi connections through your publicly available APs or even do some more broadband funky spectrum capture analysis and claim that is what an internet search engine does in its spare time and all is fine (lol).
LLMs and GenAI are quite interesting things but I do not think that they are the last word in ... AI. Anyway the latest cool thingie cannot be allowed to break whatever the current unspoken and somewhat undefined social contract is in place.
This bloke from MS seems to have forgotten that there really is a social contract of some sort and that if you say: "fuck you lot, omnomnom ... mmmm data ... ... laters (lol)" there might be some come back.
I make an http _REQUEST_, the server voluntarily fulfills the request.
Why is it okay for a person to view your content, memorize it, and use it as a base for new content while it's not okay for an AI? at the end of the day it is the same thing.
Because it is not really AI. We can have this comparison when it actually can think. Until then this 'training' is just big data processing and that is not what person can do by reading/watching.
It's never been ok for a person to do that.
You don't even have to make a request to hear a radio broadcast or see a tv broadcast, they are actively broadcast and bathe you whether you wanted it or not.
Now imagine setting up a service, even a free one where you don't make any commercial gain from it, where people ask you for stuff, and you give them bits of the stuff you've collected that was just literally falling from the sky. You give them the stuff not in the form of quotes for discussion, but presented as content itself, devoid of any prior context.
The user asks for some sci fi, and you give them 15 minutes of Star Wars which you got from a tv broadcast. But you don't tell the user "This is Star Wars, written by George Lucas" You just give them the content and the user thinks you created it.
Or worse, the user thinks THEY created it, merely using your tool, just a fancier version of using a spell checker, they are still the author since they directed the tool.
The user asks for a deep insight into human nature, you give them something that some famous author said, but not the authors name. The user then publishes a book containing this gem of deep insight, and other users think they are deeply insightful and buy their book.
This is not what happens when a person learns from life, including reading famous books, and then starts producing their own content.
The fact that it's possible to construct some scenarios where it seems like all the essential elements are the same does not actually make the two things equivalent.
Here's my own little insight into human nature I'll offer: It's pretty common for some people who pride themselves on being smart, for some mysterious reason to only be selectively smart. They will construct good valid logic that serves one purpose readily, and yet somehow fail to construct or even percieve equally valid logic that serves some other purpose.
If one is smart enough to try to present this logical argument that it's ok to steal other peiple's work, I submit that they are then also smart enough to be able to work out how and why that is in fact stealing.
We don't even have to get into the obvious, how MS themselves "broadcast" Windows to everyone by getting manufacturers to preinstall it on every machine. That wasn't even an http request. You just wanted to buy a piece of hardware, and it just came with Windows whether you wanted it or not. Even if you remove it, it was already given to you before you opened the box, and even after you wipe it, it's licence key is still there literally burned into the hardware. Therefor, since MS put it out there themselves, put it into your hands themselves, you are free to reproduce Windows. Oh, just pieces? Pieces are fair use? So, just a handy dll then. That must be ok.
There is no end to the ways this argument doesn't hold water.