AI impact, futures stakes, and governance: the analysis of the French AI commission

A short presentation in the context of the World Artificial Intelligence Conference 2024, in Shanghai, in an event on International AI governance.

I summarize the work of the French AI commission related to the topic of the day: analyzing the impact, stakes, and desirable governance around AI. I touch upon promises and risks, grounding the discussion on an overview of the current AI landscape and scientific evidence. I briefly touch upon some recommendations for collective action.

Transcript

1. AI impact, futures stakes, and governance The analysis of the French AI commission Ga¨ el Varoquaux
2. The French AI commission 15 experts: 8 tech executives, 4 academics, 3 civil society Philippe Aghion Anne Bouverot Gilles Babinet Bernard Charl` es C´ edric O Jo¨ elle Barral Luc Julia Isabelle Ryl Alexandra Bensamoun Yann Le Cun Arthur Mensch Nozha Boujemaa Franca Salis-Madinier Martin Tisn´ e Ga¨ el Varoquaux G Varoquaux 1
3. This talk 1 A short assessment of AI’s impact 2 Some dangers 3 Recommendations for collective actions G Varoquaux 2
4. Inventions bring hopes and fears In modern times, our societies have been reshaped by electricity, cars, digital technologies... Transformations brought progress, and new dangers some fantasized G Varoquaux 3
5. A driver of a better world AI is, in itself, neither good nor bad It depends what we, as a society, do of it G Varoquaux 4
6. A driver of a better world Evidence in AI for health Radiology screening Mammography reading (FDA-approveda) Health-economics rational: - Early-detected cancer curable - Ultra-sound is cheap - Doctor’s reading is the bottleneck aFDA report on “Mammoscreen”, 2020 https://fda.report/PMN/K192854 G Varoquaux 5
7. A driver of a better world Evidence in AI for health Hospital resource allocation Early warning, forecasting... - Length-of-stay prediction Stone et al, PLOS digit health, 2022 - Screening germs Yu et al, Microbiol Spectr, 2022 - Screening for heart attack Chang et al, European Heart Journal, 2021 - Screening for artery stenosis Hsu et al, Comput Biol Med. 2020 - Hospital discharge check-up Jiang et al, Nature 2023 Beneficial if helps organizing G Varoquaux 6
8. A driver of a better world Micro-economic evidence in a customer service department Adoption of GenAI Brynjolfsson et al, 2023, NBER G Varoquaux 7
9. But let us be mindful of inflated promises https://www.theverge.com/2016/8/23/12603624/delphi-mobileye-self-driving-autonomous-car-2019 G Varoquaux 8
10. But let us be mindful of inflated promises https://www.verdict.co.uk/fully-self-driving-cars-unlikely-before-2035-experts-predict/?cf-view&cf-closed G Varoquaux 9
11. Promises Generative AI startups = $ 22 billions investments in 2023 There is a bubble in AI G Varoquaux 10
12. Between overselling and underestimating AI = A gradual evolution in technology, a long term revolution in everyday experience AI techniques are already ubiquitous During the industrial revolutions, productivity gains where visible only 20 years after a technology was invented We must create the best conditions for collective appropriation G Varoquaux 11
13. A need for trust Norms and regulations G Varoquaux 12
14. A risk analysis Risks due to imperfections Discrimination and stereotyping Misinformation Threats to privacy Production of harmful content Intellectual property violation Risks due to malicious use Cyber criminality Cyber terrorism Biosecurity Mass surveillance Systemic risks Concentration of power Disruption of labor market Weakening of cultural diversity Electricity and carbon footprint Critical emergent behavior G Varoquaux 13
15. Mis-information Quality access to information is the cement of democratic societies Threats from AI Filter bubbles Fake content generation Disrupting the economics of news G Varoquaux 14
16. Threats to privacy 1. Centralized AI scoop up data 2. AI leak training data G Varoquaux 15
17. Biased AIs discriminate Biased algorithmic governance Many documented cases of harm eg: Dutch childcare benefits scandal ( ) Bias is a socio-technical problem a shortcoming that doesn’t annoy all Perpetuated in language models Represent less well South-East Asians [Chen et al 2024] G Varoquaux 16
18. An ever-rising footprint 2000 2010 2020 1012 1018 1024 Training FLOP 1 day compute of largest supercomputer Language Vision Other Unknown Games Multimodal Drawing Speech AI compute overtook largest supercomputers G Varoquaux 17
19. An ever-rising footprint 2010 2020 106 109 1012 1015 Inference FLOP FLOPS in $100 GPU Unsustainable inference – rebound canceling out hardware gains G Varoquaux 17
20. Concentration of power Big gets bigger More data, more compute Automation centralizes choices Couples to the other risks Mis-information Privacy Bias G Varoquaux 18
21. Need for multi-stakeholder AI Open source Openness enables innovation Helps mitigate bias Important notion: commons Challenges to open source Liabilities Can kill open source Lack of transparency What data went in the model? Open washing Source non-commercial open weights G Varoquaux 19
22. Need for international governance As internet, AI easily goes across boarders Training data span across juridictions creating more representative AIs Shared governance will help us build common objects As ICANN for Internet G Varoquaux 20
23. A World AI Organization Representatives of - states and inter-state organizations - research, general-interest structures, companies, & territories Tasked with - establishing standards, including of audit - establishing the state of knowledge on AI and its impact - advising on strategic orientations G Varoquaux 21
24. Our stance on AI Collective appropriation of AI will maximize its benefits The AI transformation is ongoing We need AI commons, openess G Varoquaux 22