Thread by @blairasaservice

I've submitted a brief to Parliament on generative AI & its impacts on artists. It outlines 5 significant gaps in Canada's proposed Artificial Intelligence and Data Act and makes 5 recommendations for strengthening AIDA

Read on SSRN here: papers.ssrn.com/sol3/papers.cf…

Summary below🧵

Generative AI (GAI) systems like ChatGPT, Dall-E, and Stable Diffusion have been trained on an enormous volume of stolen text, images, audio and video. GAI has already caused innumerable physical, psychological, and economic harms to artists & creators.

The brief provides a high-level technical overview of how GAI systems function, focusing on four main phases of data preparation, model development, model deployment, and operation & monitoring.

Across those phases of the GAI lifeycle, Canada's proposed AI law (AIDA) contains 5 significant gaps in its ability to prevent GAI from having harmful impacts on artists & creators.

(1) AIDA contains no specific requirements including reqs for GAI, even the key term "high-impact system" is undefined. See @Maroussia_____ & @f_mb's piece on AIDA's "blank cheque" approach - specificity is to be provided only after AIDA is already law cigionline.org/articles/ai-re…

(2) AIDA was drafted without adequate public consultation/awareness of AI & AI regulation issues. @biancawylie provides an excellent analysis of how this particular gap in AIDA has created significant risks of regulatory and democratic process capture: biancawylie.medium.com/iseds-bill-c-2…

Important to note that the gov's own AI Public Awareness Working Group urged the gov to undertake awareness-building, consultation, & policy co-design activities, but AIDA was drafted and tabled in Parliament without observing their recommendations: ised-isde.canada.ca/site/advisory-…

(3) AIDA does not enable the deletion of IP-infringing datasets/AI models like the US FTC does through its practice of "algorithmic disgorgement", nor does AIDA enable targeted prohibitions/restrictions on infringing AI developers/service providers
protocol.com/policy/ftc-alg…

Additionally, AIDA does not enable adequate mechanisms for public oversight. The proposed AI & Data Commissioner is to be a ministerial assistant rather than an independent official. @christelletono & co. cover this concern at length in their report: cybersecurepolicy.ca/aida

(4) The federal gov lacks adequate capacities to conduct effective and timely research on AI policy. @GPAI_PMIA reports provide compelling evidence of the harmful environmental & labor impacts of AI, but these impacts are not addressed by AIDA gpai.ai/projects/futur…

(5) AIDA's jurisdictional basis relates to trade and commerce. This jurisdictional scope is far too narrow to effectively govern the wide range of impacts that GAI & other types of AI systems pose to artists, human rights, workers, and the environment.

To correct those 5 gaps, the brief makes 5 corresponding recommendations: (1) Remove AIDA from the legislative agenda until adequate public consultation has been conducted and draft regulations/standards have received adequate public feedback.

GAIs dev'd with scraped data should be pre-classified as "high-impact systems." Draft regs should provide specific requirements for sourcing of training/testing data and creator opt-in, attribution/documentation of sources, & quality assurance of data prep & model dev activities

(2) Launch national public info campaigns & consultations to gather more detailed info about GAI impacts & harms, including on artists & creators. Participation in policy co-design should be open, transparent, rigorous, & inclusive of a diverse range of interests & policy needs

(3) Amend a future version of AIDA to include stronger enforcement & oversight mechanisms, including powers & mandates for regulatory independence, proactive investigations, algorithmic disgorgement, targeted prohibitions/restrictions, and periodic public consultation

(4) Allocate resources to strengthen AI policy capacities across federal institutions, and involve a greater diversity of AI experts & stakeholders in policymaking, including arts & community orgs, social science/humanities researchers, labour unions, & professional associations

(5) Mandate federal departments and agencies to coordinate on AI regulation, and extend co-regulatory activities into other govs and civil society. A National AI Alliance would help to facilitate this kind of interdepartmental, intergovernmental, & intersectoral approach

Thank you to everyone who shared thoughts with me on this brief & on AIDA more generally. Convos I've had with artists & researchers from the #CreateDontScrape community and beyond about GAI & their policy needs for much of the last year have been vital in shaping this brief

This brief is a bit technical, but if you're unfamiliar with GAI/AIDA and looking for a more digestible explanation of all of this info, please take a look at the explainer doc we've organized: docs.google.com/document/d/1KR…

I'll continue to seek out thoughts on GAI & AIDA from folks in the community and update the explainer doc periodically, but will be mostly stepping back from AIDA for the summer until Parliament begins their committee study of the bill in the fall. Much more on AIDA from me then