Artificial Intelligence, Social Justice and Digital Civics

,Talk at DCitizens Seminar Series, 1st February 2024

https://alandix.com/academic/talks/Northumbria-2024/

It is hard to go a day without a new story about AI. Some are positive, such as potential treatments for disease, some negative such as the discriminatory bias in many automated algorithms. We are told that AI can create new jobs as well as replace old ones, but it is unclear whether AI will replace drudgery or just leave this to humans. We are at a time of unprecedented opportunity and threat and sadly the tendency for technological advance is rarely to improve social justice. It is therefore more important than ever that those in academia and the third sector do the things that will not naturally happen as a result of market economics. Part of this is about developing new applications that ameliorate the negative impacts of technology or achieve positive new things to improve the human condition. Part of this is about fundamental AI development, as the trend towards larger and larger AI has both environmental impact and reduces access to AI for all but the richest organisations and countries. If we want AI to enable a more just society, we have to work to achieve this.

1. Alan Dix https:// alandix.com / @ alanjohndix AI, Social Justice and challenges for digital civics

2. today I am not talking about … qualitative–quantitative reasoning deep digitality and digital thinking next generation UX tools long tail of small data physicality now digital light walking round Wales virtual crackers and slow time digital humanities and community heritage modeling dreams, regret and the emergence of self

3. 2 nd edition coming soon plus AI for HCI AI for Social Justice plugs

4. dealing with constant error from minimizing AI error to robust human-AI systems designing for appropriate intelligence AI + interaction => avoiding gotchas

5. synergistic AI design principles for both AI and UI change … epistemic interaction theory and principles building a roadmap … AVI Workshop 0.5 1.0 0.5 1.0 std. AI optimize output false positive rate (1–precision) true positive rate (recall) synergistic AI options + confidence for user

6. it takes two to tango a synergistic approach to human-machine decision making https://tango- horizon.eu /

7. impact of AI what AI does how AI shapes society

8. c.f. cars what cars do + ve transport, independence – ve accidents, speeding, drunk drivers how cars shape society suburban sprawl car poverty global warming

9. impact of AI what AI does how AI shapes society

10. ACCESS DOING IT RIGHT POSITIVE ACTION AVOIDING HARM POLICY PROCESS GOVERNANCE AI for Social Justice Clara Crivellaro

11. ACCESS DOING IT RIGHT POSITIVE ACTION AVOIDING HARM POLICY PROCESS GOVERNANCE

12. ACCESS DOING IT RIGHT POSITIVE ACTION AVOIDING HARM POLICY PROCESS GOVERNANCE AVOIDING HARM

13. Avoid harm Bans Inclusive data sets Auditing Base rates and correlates Neutral is not unbiased

14. ACCESS DOING IT RIGHT POSITIVE ACTION AVOIDING HARM POLICY PROCESS GOVERNANCE DOING IT RIGHT

15. UK A’level ‘mutant algorithm’ Entrenching inequality Tried to be fair Lack of data Lack of trust Unreasonable expectations of algorithms Mirror on existing system?

16. ACCESS DOING IT RIGHT POSITIVE ACTION AVOIDING HARM POLICY PROCESS GOVERNANCE POSITIVE ACTION

17. impact of AI what AI does how AI shapes society economy, education, environment politics, power democracy, digital poverty

18. AI … why now? 1990 2000 2010 2020 1980 1970 1950 1960 AI winters

19. AI … why now? 1990 2000 2010 2020 Early NN web search recommender systems MapReduce deep NN AlphaGo LLM ChatGPT Big Data Big Computation

20. AI and Computation deep NN were always known about … what changed? web (search, ecommerce) => big data => big data centres big data + big computation => DNN the new AI

21. new AI BIG data + BIG computation = BIG business

22. Datanomics 2000s – networks … 2 kinds internet global reach => transnational markets loss of national control (e.g. online gambling) human networks

23. network effects networks of connections within groups and between groups networks change value if your colleagues use Word it is more of value to you teachers parents children

24. feedback loops negative – stability, balance, golden mean smooth shapes, spherical water drop universal controller positive – instability, dynamism, extremes sharp edges, fractals, snowflake tipping points and hysteresis https:// commons.wikimedia.org /w/ index.php?curid =44338386

25. markets and monopolies market economics assumes open markets monopolies natural – single resource engineered => anti-trust laws (national) network effects => winner takes all => emergent monopoly https:// commons.wikimedia.org /w/ index.php?curid =127909261

26. digital breaks market economics

27. AI and peacock tails Jatin Sindhu, CC BY-SA 4.0, via Wikimedia Commons

28. why peacock tails? Darwinian explanation – runaway sexual selection economics often described in Darwinian terms – technological evolution – commercial survival of the fittest how does AI fit in?

29. the attention economy on the ‘free’ web … ultimately advertising pays web sites (e.g. social networks, search) best placement => best click-throughs advertisers want click-throughs => advertisers choose best placement winner takes all

30. add AI big data + best AI => best placement advertisers winner takes all => best AI takes most of the money small improvement in AI => lots more money + AI relatively small part of cost => runaway AI … well beyond ‘optimal’

31. financial and environmental cost

32. 2019 626,000 lbs CO 2 to train

33. 2020 warnings from within

34. 2023 trillion training tokens

35. who can afford it?

36. companies and countries?

37. Open Data “ I don’t know if any public sector has necessarily cracked the nut on attracting the right skills and capabilities ,” … “ The commercial sector has, because they’ve got the dollars to spend .” Ian Bartram, Gartner that was before big AI!

38. signs of hope

39. Google open LLaMa the model’s carbon footprint has been reduced to a seventh of GPT-3 the pretrained model will obviate the need to train the model … deploy the full model “using only 16 NVIDIA V100 GPUs.” ~ $400,000

40. LORA – open and (relatively) cheap reduce cost and carbon low dimension representation of inner layers

41. citizens and communities?

42. a growing divide?

43. signs of hope

44. generative enables niche education personalised tutors => choice of materials changes what we need to learn … not necessarily digital! minority languages accessibility no more alt tags? neurodiversity too https:// www.theguardian.com /world/2023/ sep /11/ sweden -says-back-to-basics-schooling-works-on-paper

45. summary AI and digital changes society & undermines market economics can entrench power and disenfranchise the poor signs of hope lower cost lower carbon foot print AI opportunities to redress inequalities

46. many challenges Alan Dix https:// alandix.com / what will you do?