No, multimodal ChatGPT is not going to “trivially” solve Generative AI's copyright problems
We recorded 9 claims from this post. 3 were checked against evidence. 6 can’t be checked. 4 more are held. Read the post at its source.
Claims from this post (9)
In the order they appear in the post. Each quotation is the speaker’s words in this post. A label such as “Supported” or “Can’t be checked” is the claim’s current assessment; its page has the evidence and the reasons.
Assessment: Can’t be checkedSaid
“Fundamentally, Southen and I had argued that our images were evidence not only that some popular GenAI systems had trained on copyrighted images (nobody is really disputing this) but also that users might accidentally find themselves in trouble, because the GenAI systems don’t list sources, and don’t warn users about possible infringment.”
What we tested: Users of popular GenAI systems might accidentally get into legal trouble because the systems do not list sources or warn about possible infringement.
Assessment: Can’t be checkedSaid
“We further speculated that smaller artists might be particularly vulnerable to being infringed; if the system made Mario, everyone might recognize that; if the system ripped off a lesser-known photographer, the user might fail to recognize the original source.”
What we tested: Marcus and Southen speculated that smaller artists might be particularly vulnerable to being infringed by GenAI, because a user might recognize a copy of Mario but might fail to recognize the original source when the system rips off a lesser-known photographer.
Assessment: Can’t be checkedSaid
“Output similarity and how well can it be resolved may well be central in tons of upcoming lawsuits, and as such the outcome could have a huge impact on the future economics of generative AI.”
What we tested: Output similarity may be central in many upcoming lawsuits, and their outcome could greatly affect the economics of generative AI.
Assessment: UnresolvedSaid
“Reid spent about 45 minutes looking into this earlier today (why not more? see below) with images generated by Midjourney, and it quickly became apparent that using ChatGPT as intern to trace down sources is, well, pretty intermittent.”
What we tested: In a 45-minute test, multimodal ChatGPT was unreliable at identifying the copyrighted sources of Midjourney-generated images.
Assessment: UnresolvedSaid
“It might be trivial to do it for the most popular characters, presented in the most canonical ways, but getting it right for all copyright materials is an entirely different matter.”
What we tested: Building a classifier that detects copyrighted material might be trivial for the most popular characters presented in the most canonical ways, but getting it right for all copyrighted materials is an entirely different matter.
Assessment: Can’t be checkedSaid
“(We also speculate that any such classifier would sometimes “false alarm” and misidentify some genuinely new materials as copyrighted.)”
What we tested: Marcus and Southen speculate that any such classifier for detecting copyrighted generated images would sometimes false-alarm and misidentify some genuinely new materials as copyrighted.
Assessment: Can’t be checkedSaid
“The moral of this story, like the moral of so many others I have told, is that almost nothing in AI is truly trivial.”
What we tested: Almost nothing in AI is truly trivial.
Assessment: SupportedSaid
“More than that, GenAI developers won’t get off so easy, and lawyers looking at copyright infringement will still have plenty of work to do.”
What we tested: GenAI developers will not easily escape copyright problems, and copyright infringement lawyers will have plenty of work.
What happened: The amended-complaint ruling kept Copyright Act claims alive against the image-generation defendants after an earlier dismissal round. It establishes continued contested litigation, not ultimate infringement liability. (U.S. District Court, Northern District of California, )
Assessment: Can’t be checkedSaid
“Gary Marcus still thinks that licensing is the best and only way to go.”
What we tested: Gary Marcus still thinks that licensing is the best and only way to go in dealing with generative AI's copyright problems.
Held from this post (4)
These records did not pass every check, so nothing about them is published. Held is not a finding.
- The criteria are ambiguous. Our reading, not assessed: It has not been shown that ChatGPT or reverse image search can reliably identify sources of generated images, especially for recent and less-known materials.
- The criteria are ambiguous. Our reading, not assessed: Identifying sources of generated images is not solved, and building a classifier to detect and suppress images containing copyrighted characters is not trivial.
- The record does not meet the current rules. Our reading, not assessed: 2024 is on track to be the year of generative AI litigation.
- The criteria are ambiguous. Our reading, not assessed: Some popular GenAI systems have trained on copyrighted images.