Ethical challenges when #AI are trained on #opensource #CCBY licenses? Very concrete cases where #OpenData have been reused in ways authors would probably sanction. #PostOpenSource era pushes for ethical uses only -- will be interesting to see how this plays out. #OAI14 #YanivBenhamou
#YanivBenhamou #OAI14 - takeaways:
1. Upstream SA terms propagate to training data sets but copyright change might brake once the model is created (depends on German vs. UK court)
2. AI vendors incentivized to use SA models.
#YanivBenhamou #OAI14 - Trained model does NOT contain/reproduce its training data, fully or partially (but there might be "traces"). SA terms thus arguably DO NOT propagate to the trained model - as the latter is not a software derivative of its training data (but legal landscape shifting quickly)
#YanivBenhamou #OAI14 - how does #ShareAlike clause (SA) works. See Apache/MIT terms propagating terms to all derivatives from original data source. Training dataset directly contains and reproduces SA-protected material. SA terms arguably propagate to the entire dataset. #opensource #openscience
#YanivBenhamou #OAI14 #opensourceAI doesn't necessarily mean open training data. Data contained non-copyrighted elements and leads to fragmented/conflicted legal regimes
#openscience #opensource
#YanivBenhamou #OAI14 authoritative refs from EU AI Act: #opensource license = access, usage, modification, and distribution of the model - parameters including weights and model architecture usage are made publicly available (Article 53 (2)) #openscience
#YanivBenhamou #OAI14: Spectrum of openness in license- permissive <-> restricted. OA are accessible but not fully #opensource vs Mistral. #openWashing - AI vendors promoting models as OS when they're really not - unitarily apply restricting licensing term (Llama 2) but terms contain restrictions
#YanivBenhamou #OAI14 What is #OpenSource #AI? OS License = grants access to content, usually respecting "four essential freedoms" (run, study, modify, and redistribute - FOSS); permissive license (like MIT/apache) or CCBY NC SA.
The question of what restrictions go where is key #YanivBenhamou #OAI14: Weighting factors, architecture, math calculations are non copyrightable. Does copyright license apply to such non copyrightable elements? #AI #openscience
#YanivBenhamou #OAI14 - legal landscape moving quickly - Getty Images rights holders just lost to Stability AI. #openAI use of music however lost in German courts.
#YanivBenhamou next: #opensource #AI propogation of OS licenses in Age of AI: Prof of #copyright #digitalLaw at #UniversityofGeneva. Recent paper from his team: link.springer.com/article/10.1... #OAI14
