3 ms·

The courts don't come to decisions like this, but it's pretty straightforward to want to prevent AIs from being able to invent patents (even as an aid) because

by Osmose 3y ago

The courts don't come to decisions like this, but it's pretty straightforward to want to prevent AIs from being able to invent patents (even as an aid) because the obvious consequence of allowing it is companies automating the generation of patents as much as possible and flooding the patent system. Anyone who's worked on spam prevention knows that as long as the profit is higher than the cost of impersonating a human, spammers will continue to hammer you until you can make the costs high enough to be no longer profitable.

sgt101 3y ago
There is an intersting wrinkle here, I haven't seen it mentioned so far. In the UK a patent is invalid if you put an inventor on it who didn't make a significant and definable contribution to the invention. If a patent is AI generated, and then labelled with a human inventor it may be argued that it is invalidated by the humans lack of inventive input.
- jonplackett 3y ago
  If the human knows this though surely than can work backwards and do some busywork that makes it look like they did it all.
- pbhjpbhj 3y ago
  That sounds great to me, it means companies can create with ML but can't lock up the fruits of that. So they can't ditch their R&D people if they want to use patents -- that looks, inadvertently, like a good balance?
- alexisread 3y ago
  That's a great observation. I guess the problem here is one of enforcement- with so many patents coming in, how can you a) prove that a human didn't contribute meaningfully, an act of transcription obfuscates the origin here. b) Devote enough manpower to doing this for every patent? If the contribution is opt-in then it's marginally better, but I can see that the lion's share of patents would go to whoever has the biggest LLM, thus pushing the wealth gap larger. Patents are about novelty and effort (we want to reward the inventor), who's to say that these inventions would be novel, when anyone with a large LLM can find them?
nonameiguess 3y ago
I suppose this will vary from jurisdiction to jurisdiction, but for the most part, the idea of assigning a patent to a machine system seem precluded by the more general inability of machine systems to have legal rights at all other than the right of property to not be defaced or destroyed by anyone other than the property owner. Rights to ownership and control in general are typically limited to humans and collective entities such as governments and private corporations created to disaggregate the interests of communities of humans from the specific individuals making up that community at one particular time. There have been some exceptions, such as Leona Helmsley bequeathing part of her estate to her dog, but just as with infants or mentally unfit humans who may possess rights of ownership but not understand them well enough to make appropriate decisions regarding disposition, that money was put into a trust. Putting aside the philosophical and legal theory questions regarding what kind of entities can or should have rights and the ability to own property, there are more practical questions that would need to be answered. If you really wanted to assign a patent to an automated system that generated an idea, how would you do it? What is the boundary of this system separating it from its environment? We have a fairly clear answer to this when the system in question is a biological organism. But a software system? Is it the process ID of the first process spawned from the executable file you used to instantiate the system before querying it? Presumably that was killed at some point well before you ever filed the application. The server or servers it ran on? Those also may be ephemeral. Some content-addressing tagging system identifying a released version of the software? Well then the specific instances you ran to generate the idea shares ownership with a bunch of other instances that didn't do jack. The source code? Now it's shared with even more potential instances that don't exist yet and may never exist. Do you snapshot the entire state machine consisting of register and memory contents across all hardware servers as they happened to exist at the moment of creation? Then what? As long as it is possible to physically realize this system by restoring the snapshot to currently powered on hardware with a sufficient similar processor architecture and operating system the system lives forever? But if that operating system or architecture disappears then it dies and the rights expire? Does the snapshot itself retain some kind of legal rights even if it is never physically realized and only exists as a file? Seemingly no, as even a potentially intelligent system that is in a state of hibernation has no interests and can't make decisions. So now we have a further question. Even if we can somehow define an acceptable boundary of what even constitutes the entity that has rights, we have to define at what point it is still considered to be alive. It's not clear for even a single process on a single server that is getting pre-empted and potentially paged out anywhere from a few times a day to billions of times per second, but much less clear for a distributed system that relies upon replaying log-structured files shared across an unreliable network. Humans get death certificates when a physician determines the heart has been stopped long enough that it is unlikely to ever come back. But for a distributed system that relies on log consistency, it can always potentially come back as long as some combo of hardware and software exists that can interpret the state transition commands to restore its state and someone with the power to plug servers together, turn them on, and hit enter makes the decision to do it. The lifetime of such a system is arbitrary.

by Osmose 3y ago

sgt101 3y ago

There is an intersting wrinkle here, I haven't seen it mentioned so far. In the UK a patent is invalid if you put an inventor on it who didn't make a significant and definable contribution to the invention. If a patent is AI generated, and then labelled with a human inventor it may be argued that it is invalidated by the humans lack of inventive input.

jonplackett 3y ago
If the human knows this though surely than can work backwards and do some busywork that makes it look like they did it all.
pbhjpbhj 3y ago
That sounds great to me, it means companies can create with ML but can't lock up the fruits of that. So they can't ditch their R&D people if they want to use patents -- that looks, inadvertently, like a good balance?
alexisread 3y ago
That's a great observation. I guess the problem here is one of enforcement- with so many patents coming in, how can you a) prove that a human didn't contribute meaningfully, an act of transcription obfuscates the origin here. b) Devote enough manpower to doing this for every patent? If the contribution is opt-in then it's marginally better, but I can see that the lion's share of patents would go to whoever has the biggest LLM, thus pushing the wealth gap larger. Patents are about novelty and effort (we want to reward the inventor), who's to say that these inventions would be novel, when anyone with a large LLM can find them?

nonameiguess 3y ago

I suppose this will vary from jurisdiction to jurisdiction, but for the most part, the idea of assigning a patent to a machine system seem precluded by the more general inability of machine systems to have legal rights at all other than the right of property to not be defaced or destroyed by anyone other than the property owner. Rights to ownership and control in general are typically limited to humans and collective entities such as governments and private corporations created to disaggregate the interests of communities of humans from the specific individuals making up that community at one particular time. There have been some exceptions, such as Leona Helmsley bequeathing part of her estate to her dog, but just as with infants or mentally unfit humans who may possess rights of ownership but not understand them well enough to make appropriate decisions regarding disposition, that money was put into a trust. Putting aside the philosophical and legal theory questions regarding what kind of entities can or should have rights and the ability to own property, there are more practical questions that would need to be answered. If you really wanted to assign a patent to an automated system that generated an idea, how would you do it? What is the boundary of this system separating it from its environment? We have a fairly clear answer to this when the system in question is a biological organism. But a software system? Is it the process ID of the first process spawned from the executable file you used to instantiate the system before querying it? Presumably that was killed at some point well before you ever filed the application. The server or servers it ran on? Those also may be ephemeral. Some content-addressing tagging system identifying a released version of the software? Well then the specific instances you ran to generate the idea shares ownership with a bunch of other instances that didn't do jack. The source code? Now it's shared with even more potential instances that don't exist yet and may never exist. Do you snapshot the entire state machine consisting of register and memory contents across all hardware servers as they happened to exist at the moment of creation? Then what? As long as it is possible to physically realize this system by restoring the snapshot to currently powered on hardware with a sufficient similar processor architecture and operating system the system lives forever? But if that operating system or architecture disappears then it dies and the rights expire? Does the snapshot itself retain some kind of legal rights even if it is never physically realized and only exists as a file? Seemingly no, as even a potentially intelligent system that is in a state of hibernation has no interests and can't make decisions. So now we have a further question. Even if we can somehow define an acceptable boundary of what even constitutes the entity that has rights, we have to define at what point it is still considered to be alive. It's not clear for even a single process on a single server that is getting pre-empted and potentially paged out anywhere from a few times a day to billions of times per second, but much less clear for a distributed system that relies upon replaying log-structured files shared across an unreliable network. Humans get death certificates when a physician determines the heart has been stopped long enough that it is unlikely to ever come back. But for a distributed system that relies on log consistency, it can always potentially come back as long as some combo of hardware and software exists that can interpret the state transition commands to restore its state and someone with the power to plug servers together, turn them on, and hit enter makes the decision to do it. The lifetime of such a system is arbitrary.