No, not what I said at all. If you’re trying to say I’m making this argument I’d urge you (ironically) to actually analyze what I said rather than putting words in my mouth ;) (Or just, you know, ask me to clarify)
Copyright infringement (or plagiarism) in it’s simplest form, as in just taking the material as is, is devoid of any analysis. The point is to avoid having to do that analysis and just get right to the end result that has value.
But that’s not what AI technology does. None of the material used to train it ends up in the model. It looks at the training data and extracts patterns. For text, that is the sentence structure, the likelihood of words being followed by another, the paragraph/line length, the relationship between words when used together, and more. It can do all of this without even ‘knowing’ what these things are, because they are simply patterns that show up in large amounts of data, and machine learning as a technology is made to be able to detect and extract those patterns. That detection is synonymous with how humans do analysis. What it detects are empirical, factual observations about the material it is shown, which cannot be copyrighted.
The resulting data when fed back to the AI can be used to have it extrapolate on incomplete data, which it could not do without such analysis. You can see this quite easily by asking an AI to refer to you by a specific name, or talk in a specific manner, such as a pirate. It ‘understands’ that certain words are placeholders for names, and that text can be ‘pirateitfied’ by adding filler words or pre/suffixing other words. It could not do so without analysis, unless that exact text was already in the data to begin with, which is doubtful.
No, not what I said at all. If you’re trying to say I’m making this argument I’d urge you (ironically) to actually analyze what I said rather than putting words in my mouth ;) (Or just, you know, ask me to clarify)
That was your implied argument regardless of intent.
Copyright infringement (or plagiarism) in it’s simplest form, as in just taking the material as is, is devoid of any analysis. The point is to avoid having to do that analysis and just get right to the end result that has value.
Completely wrong, which invalidates the point you want to make. “Analysis” and “as is” have no place in the definition of copyright infringement. A derivative work can be very different from the original material, and how you created the derivative work, including whether you performed whatever you think “analysis” means, is generally irrelevant.
What it detects are empirical, factual observations about the material it is shown, which cannot be copyrighted.
No it detects patterns. You already said it correctly above. And the problem is that some patterns can be copyrighted. That’s exactly the problem highlighted here and here. For copyright law, it doesn’t matter if, for example, that particular image of Mario is copied verbatim from the training data. The character likeness, which is encoded in the model because it is in fact a discernible pattern, is an infringement.
That was your implied argument regardless of intent.
I decide what my argument is, thank you very much. Your interpretation of it is outside of my control, and while I might try to avoid it from going astray, I cannot stop it from doing so, that’s on you.
Completely wrong, which invalidates the point you want to make. “Analysis” and “as is” have no place in the definition of copyright infringement. A derivative work can be very different from the original material, and how you created the derivative work, including whether you performed whatever you think “analysis” means, is generally irrelevant.
I wasn’t giving a definition of copyright infringement, since that depends on the jurisdiction, and since you and I aren’t in the same one most likely, that’s nothing I would argue for to begin with. In the most basic form of plagiarism, people do so to avoid doing the effort of transformation. More complex forms of plagiarism might involve some transformation, but still try to capture the expression of the original, instead of the ideas. Analysis is definitely relevant, since to create a work that does not infringe on copyright, you generally can take ideas from a copyrighted work, but not the expression of those ideas. If a new work is based on just those ideas (and preferably mixes it with new ideas), it generally doesn’t infringe on copyright. It’s why there are so many copycat products of everything you can think of, that aren’t copyright infringing.
No it detects patterns. You already said it correctly above. And the problem is that some patterns can be copyrighted. That’s exactly the problem highlighted here and here. For copyright law, it doesn’t matter if, for example, that particular image of Mario is copied verbatim from the training data.
While depending on your definition Mario could be a sufficiently complex pattern, that’s not the definition I’m using. Mario isn’t a pattern, it’s an expression of multiple patterns. Patterns like “an italian man”, “a big moustache”, “a red rounded hat with the letter ‘M’ in a white circle”, “overalls”. You can use any of those patterns in a new non-infringing work, Nintendo has no copyright on any of those patterns. But bring them all together in one place again without adding new patterns, and you will have infringed on the expression of Mario. If you give many images of Mario to the AI it might be able to understand that those patterns together are some sort of “Mario-ness” pattern, but it can still separate them from each other since you aren’t just showing it Mario, but also other images that have these same patterns in different expressions.
Mario’s likeness isn’t in the model, but it’s patterns are. And if an unethical user of the AI wants to prompt it for those specific patterns to be surprised they get Mario, or something close enough to be substantially similar, that’s on them, and it will be infringing just like drawing and selling a copy of Mario without Nintendo’s approval is now.
The character likeness, which is encoded in the model because it is in fact a discernible pattern, is an infringement.
You have absolutely no legal basis to claim they are infringement, as these things simply have not been settled in court. You can be of the opinion that they are infringement, but your opinion isn’t the same as law. The articles you showed are also simply reporting and speculating on the lawsuits that are pending.
Plagiarism is not the same as copyright infringement. Why you think people probably plagiarize is doubly irrelevant then.
Analysis is definitely relevant, since to create a work that does not infringe on copyright
Show me literally any example of the defendant’s use of “analysis” having any impact whatsoever in a copyright infringement case or a law that explicitly talks about it, or just stop repeating that it is in any way relevant to copyright.
But bring them all together in one place again without adding new patterns
Wrong. The “all together” and “without adding new patterns” are not legal requirements. You are constantly trying to push the definition of copyright infringement to be more extreme to make it easier for you to argue.
you generally can take ideas from a copyrighted work, but not the expression of those ideas
Unfortunately, an AI has no concept of ideas, and it simply encodes patterns, whatever they might happen to be. Again, you’re morphing the discussion to make an argument.
Mario’s likeness isn’t in the model, but it’s patterns are.
Mario’s likeness has to be encoded into the model in some way. Otherwise, this would not have been the image generated for “draw an italian plumber from a video game”. There is absolutely nothing in the prompt to push GPT-4 to combine those elements. There are also no “new” patterns, as you put it. That’s exactly the point of the article. As they put it:
Clearly, these models did not just learn abstract facts about plumbers—for example, that they wear overalls and carry wrenches. They learned facts about a specific fictional Italian plumber who wears white gloves, blue overalls with yellow buttons, and a red hat with an “M” on the front.
These are not facts about the world that lie beyond the reach of copyright. Rather, the creative choices that define Mario are likely covered by copyrights held by Nintendo.
This is contradictory to how you present it as “taking ideas”.
You have absolutely no legal basis to claim they are infringement
You’re mixing up different things. I’m saying that the image contains infringing material, which is hopefully not something you have to be convinced about. The production of an obviously infringing image, without the infringing elements having been provided in the prompt, is used to show how this information is encoded inside the model in some form. Whether this copyright-protected material exists in some form inside the model is not an equivalent question to whether this is copyright infringement. You are right that the courts have not decided on the latter, but we have been talking about the former. I repeat your position which I was directly responding to before:
What it detects are empirical, factual observations about the material it is shown, which cannot be copyrighted.
Plagiarism is not the same as copyright infringement. Why you think people probably plagiarize is doubly irrelevant then.
I never claimed it was, but as I said before, it is irrelevant because copyright infringement differs in places depending on the local laws, but plagiarism is usually the concept that guides the ethical position from which those laws are produced, which is why yes, it’s relevant.
Show me literally any example of the defendant’s use of “analysis” having any impact whatsoever in a copyright infringement case or a law that explicitly talks about it, or just stop repeating that it is in any way relevant to copyright.
This is an unreasonable request, and you know it to be. Again, we don’t share the same laws and different jurisdictions provide different exceptions like fair use, fair dealing, or just straight up exclusion from copyright for their use. But it is wholly besides my argument. You can look at any piece of modern media that exists in the same space and see ideas the two share, while not sharing the same expression of that idea. How some characters fulfill the same purpose, dress the same way, or have similar personalities. You are free to make a book with a plumber, a mustached man, someone wearing a red hat with the letter M on it, and someone that goes to save a princess from a castle, but if they’re not the same person they are most likely not considered to be the protected expression of Mario. Same ideas that make up Mario, one infringing, the other not.
Nobody goes to court over this because EVERYONE takes each others ideas, “Good artists copy, great artists steal”. It’s only when you step on the specific expression of an idea that it becomes realistically actionable, and at that point transformativeness is definitely discussed almost every single time, because it is critical to determining the copyright was actually infringed, or if not.
Wrong. The “all together” and “without adding new patterns” are not legal requirements. You are constantly trying to push the definition of copyright infringement to be more extreme to make it easier for you to argue.
I’m sorry but, are you really being this dishonest? I’ve mentioned EXPLICITLY in my last comment that I wasn’t giving a definition of copyright infringement, because it’s besides the point, and not what I’m claiming. Yet here you are saying I am “trying to push” a definition. We are not lawyers or law scholars speaking to each other, I am having a discussion with you as another anonymous person on a message board.
Unfortunately, an AI has no concept of ideas, and it simply encodes patterns, whatever they might happen to be.
You are just arguing semantics and linguistics, it’s meaningless. We are not talking technical specifics, not even a specific model, nor a specific technique to specific exactly how the information is encoded. It’s a rough concept of “ideas” / “data” / “patterns”: information. And AI definitely has that.
Again, you’re morphing the discussion to make an argument.
You mean, I’m making an argument. Because yes. I am. I don’t see why this negative framing is necessary nor why this is noteworthy enough to bring up, unless you really just want to make me look bad for no apparent reason.
Mario’s likeness has to be encoded into the model in some way. Otherwise, this would not have been the image generated for “draw an italian plumber from a video game”. There is absolutely nothing in the prompt to push GPT-4 to combine those elements. There are also no “new” patterns, as you put it. That’s exactly the point of the article. As they put it:
Yes, there is some idea/pattern of “Mario-ness” in the model, I said that. This was not me trying to say no material of Mario was used in training, but that it’s not like someone pasted direct images of Mario in there, but that AI models makes logical connections between concepts and even for things we cannot put a good name to does it make those connections, and will allow you to prompt for them, but that does not mean you should.
Clearly, these models did not just learn abstract facts about plumbers—for example, that they wear overalls and carry wrenches. They learned facts about a specific fictional Italian plumber who wears white gloves, blue overalls with yellow buttons, and a red hat with an “M” on the front.
These are not facts about the world that lie beyond the reach of copyright. Rather, the creative choices that define Mario are likely covered by copyrights held by Nintendo.
I sort of already explained this without mentioning this specific example, but I’ll make it extra clear.
In the article they prompted the AI for a “video game Italian plumber”.
What person, if you asked them, to think of an “Italian video game plumber”, would not think of Mario? Maybe Luigi?
I’ll tell you, because there are very damn few famous Italian video game plumbers. The prompt is already locked in on Mario, and even humans make the logical connection to Mario. It might have had billions of images and texts to use, but any time a relation to an “Italian video game plumber” showed up, there’s Mario.
So this whole point the article makes about it not learning abstract facts about plumbers, is complete moot because they completely biased the outputs towards receiving what they want to receive. If you ask for just a plumber, for which it does have many, many results. It will make more generalizations and become less specific. Because there are more than 2 examples of plumbers in other types of situations. Humans do this exact same thing in the same task, yet somehow the AI must be infallible to this despite being artificial versions of the biological thing. And that is why analysis is protected, because humans simply cannot stop doing it and everyone is tainted by their knowledge of Mario, even though for whatever reason we might need to use one of the ideas Mario is built upon. And this is why AIs use this same defense. I can say this regardless of the jurisdiction because unless you live in some kind of dictatorship this is generally true.
Sadly, this kind of deceptive framing of AI output is common, particularly among those that are biased against AI. Sometimes it’s unintentional, but frequently specific parameters are used that will just generate specific bad results, ignoring that this may not even represent 0.001% of what the model can generate in normal situations.
This is contradictory to how you present it as “taking ideas”.
It is not. You can use the idea of Mario, you cannot use the totality of Mario. For the AI to be able to use the idea of Mario, it will also ‘learn’ the totality of Mario in the process, as Mario is a collection of ideas that are extracted. But those ideas are stored separately so they can be individually prompted for. You can prompt it to make Mario, because like literally almost every person in society, they know what ideas make up Mario better than I can put to words here. If I hire a human artist to make me a “video game Italian plumber”, their first question to me would be “Oh, something like Mario?” and their second response will be “Oh I can’t do that, and you should not want to, because you don’t own Mario.”. Humans use AI, so they need to be the ones to give that second response.
Just like a kitchen knife can be used to stab someone, doesn’t mean we produce kitchen knives for stabbing people. Just because an AI can be used to infringe, does not mean that they are produced to infringe. Which is evidence by the vast majority of other ways that it can be used that don’t infringe, which is self evident after just tinkering around with it for a little while.
You’re mixing up different things. I’m saying that the image contains infringing material, which is hopefully not something you have to be convinced about. The production of an obviously infringing image, without the infringing elements having been provided in the prompt, is used to show how this information is encoded inside the model in some form. Whether this copyright-protected material exists in some form inside the model is not an equivalent question to whether this is copyright infringement. You are right that the courts have not decided on the latter, but we have been talking about the former. I repeat your position which I was directly responding to before:
If it’s anything like the examples before, then the AI has definitely been prompted by the user to make infringing elements.
But anyways, to the question, you just don’t seem to grasp that collections of ideas can communicate copyright infringing material without being infringing on their own. It’s like arguing that if Paint or Photoshop knows about the color red that this is copyright infringing because it’s the same red that Mario uses. None of the ideas that make up Mario are infringing, and cannot be copyrighted. They are what the AI is designed to extract, not Mario as a totality.
You can definitely use AI to make an infringement machine by making it less likely to make leaps in ideas and just only combine the ideas it’s been taught on, which we as humans can do as well in the form of plagiarism and forgery. But if you’re going to be unethical why use an AI when you might as well just take the easy route directly with print screen or a photo. Two other technologies we didn’t ban for having this ability to capture copyrighted material, even if they far more blatantly copy the material.
This is where good AI usage deviates, because it instead tries to MAXIMIZE the amount of leaps and connections the AI makes for as little possibility to make something infringing. Even honest people trying to make new creative works sometimes have to change things because they might be too close to being infringing.
No, not what I said at all. If you’re trying to say I’m making this argument I’d urge you (ironically) to actually analyze what I said rather than putting words in my mouth ;) (Or just, you know, ask me to clarify)
Copyright infringement (or plagiarism) in it’s simplest form, as in just taking the material as is, is devoid of any analysis. The point is to avoid having to do that analysis and just get right to the end result that has value.
But that’s not what AI technology does. None of the material used to train it ends up in the model. It looks at the training data and extracts patterns. For text, that is the sentence structure, the likelihood of words being followed by another, the paragraph/line length, the relationship between words when used together, and more. It can do all of this without even ‘knowing’ what these things are, because they are simply patterns that show up in large amounts of data, and machine learning as a technology is made to be able to detect and extract those patterns. That detection is synonymous with how humans do analysis. What it detects are empirical, factual observations about the material it is shown, which cannot be copyrighted.
The resulting data when fed back to the AI can be used to have it extrapolate on incomplete data, which it could not do without such analysis. You can see this quite easily by asking an AI to refer to you by a specific name, or talk in a specific manner, such as a pirate. It ‘understands’ that certain words are placeholders for names, and that text can be ‘pirateitfied’ by adding filler words or pre/suffixing other words. It could not do so without analysis, unless that exact text was already in the data to begin with, which is doubtful.
That was your implied argument regardless of intent.
Completely wrong, which invalidates the point you want to make. “Analysis” and “as is” have no place in the definition of copyright infringement. A derivative work can be very different from the original material, and how you created the derivative work, including whether you performed whatever you think “analysis” means, is generally irrelevant.
No it detects patterns. You already said it correctly above. And the problem is that some patterns can be copyrighted. That’s exactly the problem highlighted here and here. For copyright law, it doesn’t matter if, for example, that particular image of Mario is copied verbatim from the training data. The character likeness, which is encoded in the model because it is in fact a discernible pattern, is an infringement.
I decide what my argument is, thank you very much. Your interpretation of it is outside of my control, and while I might try to avoid it from going astray, I cannot stop it from doing so, that’s on you.
I wasn’t giving a definition of copyright infringement, since that depends on the jurisdiction, and since you and I aren’t in the same one most likely, that’s nothing I would argue for to begin with. In the most basic form of plagiarism, people do so to avoid doing the effort of transformation. More complex forms of plagiarism might involve some transformation, but still try to capture the expression of the original, instead of the ideas. Analysis is definitely relevant, since to create a work that does not infringe on copyright, you generally can take ideas from a copyrighted work, but not the expression of those ideas. If a new work is based on just those ideas (and preferably mixes it with new ideas), it generally doesn’t infringe on copyright. It’s why there are so many copycat products of everything you can think of, that aren’t copyright infringing.
While depending on your definition Mario could be a sufficiently complex pattern, that’s not the definition I’m using. Mario isn’t a pattern, it’s an expression of multiple patterns. Patterns like “an italian man”, “a big moustache”, “a red rounded hat with the letter ‘M’ in a white circle”, “overalls”. You can use any of those patterns in a new non-infringing work, Nintendo has no copyright on any of those patterns. But bring them all together in one place again without adding new patterns, and you will have infringed on the expression of Mario. If you give many images of Mario to the AI it might be able to understand that those patterns together are some sort of “Mario-ness” pattern, but it can still separate them from each other since you aren’t just showing it Mario, but also other images that have these same patterns in different expressions.
Mario’s likeness isn’t in the model, but it’s patterns are. And if an unethical user of the AI wants to prompt it for those specific patterns to be surprised they get Mario, or something close enough to be substantially similar, that’s on them, and it will be infringing just like drawing and selling a copy of Mario without Nintendo’s approval is now.
You have absolutely no legal basis to claim they are infringement, as these things simply have not been settled in court. You can be of the opinion that they are infringement, but your opinion isn’t the same as law. The articles you showed are also simply reporting and speculating on the lawsuits that are pending.
Plagiarism is not the same as copyright infringement. Why you think people probably plagiarize is doubly irrelevant then.
Show me literally any example of the defendant’s use of “analysis” having any impact whatsoever in a copyright infringement case or a law that explicitly talks about it, or just stop repeating that it is in any way relevant to copyright.
Wrong. The “all together” and “without adding new patterns” are not legal requirements. You are constantly trying to push the definition of copyright infringement to be more extreme to make it easier for you to argue.
Unfortunately, an AI has no concept of ideas, and it simply encodes patterns, whatever they might happen to be. Again, you’re morphing the discussion to make an argument.
Mario’s likeness has to be encoded into the model in some way. Otherwise, this would not have been the image generated for “draw an italian plumber from a video game”. There is absolutely nothing in the prompt to push GPT-4 to combine those elements. There are also no “new” patterns, as you put it. That’s exactly the point of the article. As they put it:
This is contradictory to how you present it as “taking ideas”.
You’re mixing up different things. I’m saying that the image contains infringing material, which is hopefully not something you have to be convinced about. The production of an obviously infringing image, without the infringing elements having been provided in the prompt, is used to show how this information is encoded inside the model in some form. Whether this copyright-protected material exists in some form inside the model is not an equivalent question to whether this is copyright infringement. You are right that the courts have not decided on the latter, but we have been talking about the former. I repeat your position which I was directly responding to before:
I never claimed it was, but as I said before, it is irrelevant because copyright infringement differs in places depending on the local laws, but plagiarism is usually the concept that guides the ethical position from which those laws are produced, which is why yes, it’s relevant.
This is an unreasonable request, and you know it to be. Again, we don’t share the same laws and different jurisdictions provide different exceptions like fair use, fair dealing, or just straight up exclusion from copyright for their use. But it is wholly besides my argument. You can look at any piece of modern media that exists in the same space and see ideas the two share, while not sharing the same expression of that idea. How some characters fulfill the same purpose, dress the same way, or have similar personalities. You are free to make a book with a plumber, a mustached man, someone wearing a red hat with the letter M on it, and someone that goes to save a princess from a castle, but if they’re not the same person they are most likely not considered to be the protected expression of Mario. Same ideas that make up Mario, one infringing, the other not.
Nobody goes to court over this because EVERYONE takes each others ideas, “Good artists copy, great artists steal”. It’s only when you step on the specific expression of an idea that it becomes realistically actionable, and at that point transformativeness is definitely discussed almost every single time, because it is critical to determining the copyright was actually infringed, or if not.
I’m sorry but, are you really being this dishonest? I’ve mentioned EXPLICITLY in my last comment that I wasn’t giving a definition of copyright infringement, because it’s besides the point, and not what I’m claiming. Yet here you are saying I am “trying to push” a definition. We are not lawyers or law scholars speaking to each other, I am having a discussion with you as another anonymous person on a message board.
You are just arguing semantics and linguistics, it’s meaningless. We are not talking technical specifics, not even a specific model, nor a specific technique to specific exactly how the information is encoded. It’s a rough concept of “ideas” / “data” / “patterns”: information. And AI definitely has that.
You mean, I’m making an argument. Because yes. I am. I don’t see why this negative framing is necessary nor why this is noteworthy enough to bring up, unless you really just want to make me look bad for no apparent reason.
Yes, there is some idea/pattern of “Mario-ness” in the model, I said that. This was not me trying to say no material of Mario was used in training, but that it’s not like someone pasted direct images of Mario in there, but that AI models makes logical connections between concepts and even for things we cannot put a good name to does it make those connections, and will allow you to prompt for them, but that does not mean you should.
I sort of already explained this without mentioning this specific example, but I’ll make it extra clear.
In the article they prompted the AI for a “video game Italian plumber”. What person, if you asked them, to think of an “Italian video game plumber”, would not think of Mario? Maybe Luigi? I’ll tell you, because there are very damn few famous Italian video game plumbers. The prompt is already locked in on Mario, and even humans make the logical connection to Mario. It might have had billions of images and texts to use, but any time a relation to an “Italian video game plumber” showed up, there’s Mario.
So this whole point the article makes about it not learning abstract facts about plumbers, is complete moot because they completely biased the outputs towards receiving what they want to receive. If you ask for just a plumber, for which it does have many, many results. It will make more generalizations and become less specific. Because there are more than 2 examples of plumbers in other types of situations. Humans do this exact same thing in the same task, yet somehow the AI must be infallible to this despite being artificial versions of the biological thing. And that is why analysis is protected, because humans simply cannot stop doing it and everyone is tainted by their knowledge of Mario, even though for whatever reason we might need to use one of the ideas Mario is built upon. And this is why AIs use this same defense. I can say this regardless of the jurisdiction because unless you live in some kind of dictatorship this is generally true.
Sadly, this kind of deceptive framing of AI output is common, particularly among those that are biased against AI. Sometimes it’s unintentional, but frequently specific parameters are used that will just generate specific bad results, ignoring that this may not even represent 0.001% of what the model can generate in normal situations.
It is not. You can use the idea of Mario, you cannot use the totality of Mario. For the AI to be able to use the idea of Mario, it will also ‘learn’ the totality of Mario in the process, as Mario is a collection of ideas that are extracted. But those ideas are stored separately so they can be individually prompted for. You can prompt it to make Mario, because like literally almost every person in society, they know what ideas make up Mario better than I can put to words here. If I hire a human artist to make me a “video game Italian plumber”, their first question to me would be “Oh, something like Mario?” and their second response will be “Oh I can’t do that, and you should not want to, because you don’t own Mario.”. Humans use AI, so they need to be the ones to give that second response.
Just like a kitchen knife can be used to stab someone, doesn’t mean we produce kitchen knives for stabbing people. Just because an AI can be used to infringe, does not mean that they are produced to infringe. Which is evidence by the vast majority of other ways that it can be used that don’t infringe, which is self evident after just tinkering around with it for a little while.
If it’s anything like the examples before, then the AI has definitely been prompted by the user to make infringing elements.
But anyways, to the question, you just don’t seem to grasp that collections of ideas can communicate copyright infringing material without being infringing on their own. It’s like arguing that if Paint or Photoshop knows about the color red that this is copyright infringing because it’s the same red that Mario uses. None of the ideas that make up Mario are infringing, and cannot be copyrighted. They are what the AI is designed to extract, not Mario as a totality.
You can definitely use AI to make an infringement machine by making it less likely to make leaps in ideas and just only combine the ideas it’s been taught on, which we as humans can do as well in the form of plagiarism and forgery. But if you’re going to be unethical why use an AI when you might as well just take the easy route directly with print screen or a photo. Two other technologies we didn’t ban for having this ability to capture copyrighted material, even if they far more blatantly copy the material.
This is where good AI usage deviates, because it instead tries to MAXIMIZE the amount of leaps and connections the AI makes for as little possibility to make something infringing. Even honest people trying to make new creative works sometimes have to change things because they might be too close to being infringing.