Global Governance of Generative AI: The Right Way Forward
Download as PPTX, PDF0 likes1,452 views
The document discusses the global governance challenges of generative AI, including its definitions, risks such as misinformation and bias, and the regulatory landscape across different regions such as the EU, UK, and China. It highlights various proposed solutions for AI governance, ranging from strict regulations to voluntary guidelines, and underscores the importance of international harmonization in addressing AI-related risks. The document raises concerns about the power dynamics in AI regulation, emphasizing the need for inclusive governance that considers the voices of those affected by AI harms.
![Solution 3(a) – no new law
• UK “pro-innovation” “world-leading”
approach
• March 2023 White Paper
• Running against EU trend towards hard
regulation of AI
• Sectoral regulators eg ICO, CMA, FCA to
coordinate on regulation and apply
crosscutting “principles”
• No new powers, no new law.. So far
• But UK industry selling into EU will still
need to meet EU AIA standards!
• ?? AI Safety Institute = global hub for[what
type of?] regulation?! Law haven??](https://image.slidesharecdn.com/adelaidenov23-231107121203-be08a0b4/85/Global-Governance-of-Generative-AI-The-Right-Way-Forward-25-320.jpg)
Global Governance of Generative AI: The Right Way Forward
- 1. Global Governance of Generative AI : the Best Way Forward? Lilian Edwards Professor of Law, Innovation & Society Newcastle University Lilian.edwards@ncl.ac.uk Twitter : @lilianedwards Flinders, November 2023 (artwork by James Stewart, Edinburgh University, using MidJourney) Bender, Gebru, Mitchell : On the Dangers of Stochastic Parrots: Can Language Models Be Too Big? ACM FAccT, 2021
- 2. What this talk isn’t about.. (6/11/23)
- 3. Summary • What is generative AI? • Risks and harms • Global models for governance • Assessment • The future for Australia?
- 4. A.What is generative AI? Generative AI; general purpose AI”; “foundation models”; “frontier AI”
- 5. Types of large “foundation” models • GPT-3 (Open AI/Microsoft)(text prompt to text output)(June 2020) • “Large Language Model” or LLM • ChatGPT, November 2022 • GPT4, March 2023 • Multimodal, ChatGPT+ and via API • Integration into Bing • BARD (Google) • DALL-E 2, 3 (text to images – Google) • CoPilot (prompt generates computer code – GitHub/OpenAI) • Meta Make-me-A-Video (text to video - Meta) • Stable Diffusion (open source – text to image (T2I) from StabilityAI) • Midjourney (commercal API text to image (T2I() • HarmonAI – makes music from prompts • WuDao (LLM – China)
- 7. B. What’s the problem? (a) Disinfo/misinfo/deepfakes
- 12. P Hacker “The propensity of ChatGPT particularly to hallucinate when it does not find readymade answers can be exploited to generate text devoid of any connection to reality, but written in the style of utter confidence”
- 13. https://www.education.sa.gov.au/parents-and- families/curriculum-and-learning/ai#
- 14. nurse Stable Diffusion doctor DALL-E 2 ( c) Bias & stereotyping
- 15. (f) Competition/Anti-trust …and more (d)Labour issues (e)Environmental harms
- 16. New York Times Who are we regulating? Ecology of upstream providers and downstream deployers
- 17. C. Global AI Regulation & the rise of generative AI
- 18. Models for governance of generative AI • Comprehensive, mandatory regimes eg EU AI Act • Vertical new regimes of regulation – China • No mandatory regulation but… +++ • No new regulation – UK • Pre-existing relevant laws • Eg Copyright • Eg Data protection • Voluntary guidelines – “co-regulatory” instruments • Private ordering
- 19. Solution 1 : “comprehensive” - the EU AI Act – before gen AI AIA “risk based” approach • Unacceptable risk – ‘Complete’ prohibition, 4 examples – Article 5 • High-risk –Fixed categories of risky domains, based on intended use ; “essential requirements” including dataset quality, human oversight – • Limited risk – Transparency obligations for a few AI systems (chatbots, deepfakes, emotion ID, biometric categorisation) – Article 52 • Minimal risk – Codes of conduct – Article 69 Photo Source: European Commission, Digital Strategy Website https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai
- 20. High-Risk AI systems (annex III) • Annex III - Services intended for these uses.. • Biometric identification and categorisation of natural persons; • Management and operation of critical infrastructure; • Education and vocational training; • Employment, workers management and access to self-employment; • Access to and enjoyment of essential private services and public services and benefit; • Law enforcement; • Migration, asylum and border control management; • Administration of justice and democratic processes • Recommender algorithms… (EU Parl)
- 21. • Risk management system • Data and data governance ; (“data quality”) • Technical documentation • Transparency and provision of information to users; • Human oversight • “Human oversight shall aim at preventing or minimising the risks to health, safety or fundamental rights that may emerge when a high-risk AI system is used in accordance with its intended purpose or under conditions of reasonably foreseeable misuse” • Accuracy, robustness and cybersecurity (see Edwards AI Act explainer for Ada Lovelace Institute, 2022) High-Risk AI systems (Requirements) – Arts. 8 - 15
- 22. ChatGPT enters the chat… • No “intended use” !! • Definition of gen AI conflicted - OECD? Who regulates? • Should duties/liability lie on developers or deployers? • Practical power & profit benefit vs unforeseeable risk & restraint innovation • High risk, limited risk, the future of AI? • Copyright & creative industries special worries • Environmental special worries • Competition: peimarily a matter for the Digital Markets Act?? • Open vs proprietary models - risks & benefits
- 23. EU Parliament proposed “Foundation model” duties, May 2023 • Pre-audit of design, testing etc to assess, remove or mitigate foreseeable risks • Data quality duties – as with high risk AI • Performance on metrics to be checked throughout lifecycle - by independent experts • Environmental by design • Information sharing to downstream deployers– but while maintaining trade secrets (?) • Be transparent as to copyright works used • Public register of foundation models
- 24. Solutions 2 : New vertical rules - China Addiction? “prevent users' over-reliance upon or addiction to the AIGC” Censorship? “ensure AI generated content (AIGC) is legal, consistent with the core values of the socialist system and free from discrimination and fake information” Transparency+ “guide users to scientifically understand and rationally use the AIGC” License needed for providers from security services
- 25. Solution 3(a) – no new law • UK “pro-innovation” “world-leading” approach • March 2023 White Paper • Running against EU trend towards hard regulation of AI • Sectoral regulators eg ICO, CMA, FCA to coordinate on regulation and apply crosscutting “principles” • No new powers, no new law.. So far • But UK industry selling into EU will still need to meet EU AIA standards! • ?? AI Safety Institute = global hub for[what type of?] regulation?! Law haven??
- 26. • Discrimination/ equality law • Liability/ product liability • Product safety • Copyright • Andersen v. Stability AI • Content regulation, hate speech, libel, advertising, fake news • Privacy & data protection • Italy Garantie Chat GPT ban • Competition law • Labour law • Human rights Solution 3(b) : Pre-existing laws
- 27. Solution 3( c) – international voluntary principles • OECD Principles on AI, 2019 • “Four years on” review, 2023 • G7/GPAI/OECD “Hiroshima” Principles for “Advanced AI” • Guiding principles • Code of Conduct • “risk based approach” • Endorsed EU , 30 October 2023 • Identify, evaluate, and mitigate risks across the AI lifecycle • Publicly report advanced AI systems’ capabilities, limitations and domains of appropriate and inappropriate use • Information sharing across countries and reporting of incidents • Implement robust security controls across AI lifecycle
- 29. US Draft Blueprint for AI Bill of Rights, Oct 2022 & NIST standards work Solution 3 ( c) : National voluntary principles : US
- 30. • US Biden/Harris Executive Order Nov 2023 • Looks good but only actually binds public agencies where President had pre existing powers • US’s own AI Safety Institute • Practical, near term emphasis • generative misinformation ; • watermarking for transparency; • AI generated robocalls deceiving elderly • Content moderation • Will operationalize NIST technical framework for reducing AI risks • Using govt procurement tools
- 31. Private ordering: empirical work, Jan-Mar 2023 Copyright (eg service) Privacy & data protection Consumer protection& dispute resolution Illegal/harmful content; content moderation; misinfo Text to text (language models) Egs ChatGPT 4.0, Clova (SKorea) 6 models Text to image (T2I) Egs Lensa, DALL-E 2.0 4 models Text to video/audio Synthesia (T2Audio) 6 models • T&C routinely assign risk and liability for harmful or infringing outputs entirely to users • “Platformisation” • Positioned as intermediary services, like social media or search, even though actually publishers ? • But evade incoming regulation of “traditional”platforms
- 32. Australia? • “Safe and Responsible AI in Australia”, Dept of Industry, June 2023 • Domestic sectoral contributions eg eSafety Commissioner on deepfake CSAM • Takeaways? • Complex governance landscape • Key areas; competition, govt use of AI • “relatively small, open economy” • -> Global governance harmonisation important • Spectrum from voluntary to mandatory --leans towards former • “Risk management” toolkits for risk- based cases - clever pick n mix!
- 33. Assessment of global governance of genAI • Will the EU AIA become a “gold standard” global model, like the GDPR? • “General purpose AI” section rushed & fits badly w basic framework of AIA & EU digital acquis • Over-regulation? Too soon? Anti-innovation? • But we said that about social media & search – get in fast this time! • Voluntary / self regulation models • Do we need international harmonization on AI? If so, a race to the bottom – vague promises eg Bletchley - or the top – a Council of Europe Convention on AI model? FRIA? • Is debate on AI safety being being fanned by tech industry as self regulatory strategy - to avoid the detailed regulation of current harms the AIA envisages, in favour of future unlikely risks? • Self and co-regulation mainly negotiated between industry and politicians, for growth, not rights • Where is the voice of those most affected by AI harms?? UK AI Summit had zero UK civil society at it. Almost no Global South. • “Global” governance? • Rule by rich white technocratic Western men rather than states? • Tech industry as the new states? • Transnational conflicts loom: A “Beijing” effect as well as a “Brussels” effect?
Editor's Notes
- #5: Chatbots, journalism, idea generation, copywriting, coding assistance, summarizing, virtual meeting bot, AI friends
- #13: P Hacker “The propensity of ChatGPT particularly to hallucinate when it does not find readymade answers can be exploited to generate text devoid of any connection to reality, but written in the style of utter confidence”
- #16: Access to data, compute, fine tuning resources Computationally expensive and retraining slow -> large tech co dominance GPT-4 training cost >$100mn environmentally worrying
- #18: EU – soft power over US; China – keeping control over data, info 7 its tech industry; US – trad permissionless innovatin but now AGI
- #23: Issues : over-inclusive (no emphasis generality in “ability, task or output”); important for classification ->
- #25: Recommender algos; synthetic ai; generative ai
- #30: Note OECD Princs on AI 2019 endorsed by 46 countries, see 2022 follow up report
- #34: If the key US providers worried about x-risk are so worried (Open AI, Anthropic), why don’t they stop?