Re: [OT] Why governments must limit AI violations of copyright

On 5/27/2025 3:20 PM, Rhino wrote:

 On 2025-05-27 2:17 PM, BTR1701 wrote:

 On May 27, 2025 at 9:06:34 AM PDT, "Rhino"
 <no_offline_contact@example.com>
 wrote:
 

 Mary Spender presents a relatively brief but, I think, compelling
 argument for why governments need to reject the tech firms claims that
 using existing works to train AIs is fair use and does not need to be
 paid for.
 
 https://www.youtube.com/watch?v=R5pFE85oAnA [8 minutes]
 
 The tech bros are wallowing in almost unimagineable wealth: they can
 definitely afford to compensate copyright holders for using their work
 as training data. Alternatively, they can let copyright holders exclude
 their works from use in training data and compensate them for what they
 have used without permission.
 
 I don't believe the tech companies have some kind of natural right to
 generate new works that are closely modelled on existing works without
 paying for their use of those works.

 
 If you can show that the AI produces a copy of the work it was trained
 on, or
 one substantially similar enough as to be confusing to the reasonable
 man,
 then yes, I agree.
 
 E.g., if you ask it to generate a story about a young girl who finds
 herself
 lost in a fantasy world and it spits out the plot to Alice in Wonderland.
 
 But if you ask it that same question and it produces a totally
 different story
 that isn't Alice in Wonderland in any recognizable way but it learned
 how to
 do that from 'reading' Alice in Wonderland, then I don't see how you
 have a
 copyright violation under existing law or even under the philosophical
 framework on which existing law has been built. At that point, it's no
 different from a human reading Alice in Wonderland and figuring out
 how to use
 the elements and techniques employed by Carroll in his story to produce a
 different story of his own. No one would suggest copyright violation if a
 human did it, so how can it suddenly be one if a computer algorithm
 does it?
 

 The new works generated by humans are already pretty derivative in
 too many
 cases: we don't need AIs
 generating still more of the same.

 
 Well therein lies the rub. At least in America. We call it the Bill of
 Rights,
 not the Bill of Needs, for a reason.
 

 There's a wealth of art (whether music, visual art, or literature)
 freely available in the public domain. Let them use that if they need
 large quantities of art to train their models.

 
 
 

 Your points are well taken. Yes, if the AI-generated material isn't
 recognizable to someone familiar with Alice in Wonderland, it's hard to
 make a case for copyright infringement. And yes, even if *I* don't see a
 need for yet more derivative works, it's not illegal, even if it is
 annoying.
 
 The challenge is going to come with deciding if an AI-generated work is
 "too similar" to something it trained on. I expect that similarity, like
 beauty, is in the eye (or ear) of the beholder. Maybe a committee will
 have to do the deciding and only if a majority of its members thinks the
 similarity is too close will the AI work be labelled a copyright
 infringement. Of course selection of this committee will be challenging
 since the tech companies are going to favour people that don't ever see
 similarities even of identical things and the human creators will tend
 to see similarity in everything because its in their financial interest
 to find similarity.

 
Two ancillary thoughts:  Afaics, we're already within reach of such a
pilfering AI-agent that can be dialed to a desired degree of "distance"
from the original work it's copying.