Dear EAs,

After the interview with Tristan Harris Rob asked EAs to do their own research on the topic of aligning recommender systems / short term AI risks. A few weeks ago I already posed the short question on donating against these risks. After that I listened to the 2.5 hour podcast on the topic and read multiple articles on this topic on this forum (https://forum.effectivealtruism.org/posts/E4gfMSqmznDwMrv9q/are-social-media-algorithms-an-existential-risk, https://forum.effectivealtruism.org/posts/xzjQvqDYahigHcwgQ/aligning-recommender-systems-as-cause-area, https://forum.effectivealtruism.org/posts/ptrY5McTdQfDy8o23/short-term-ai-alignment-as-a-priority-cause)

I would love to collaborate with some others on collecting more in-depth arguments including some back- of-the-envelope calculations / Fermi estimates on the possible scale of the problem. I've spent some hours creating a structure including some next steps. In particular we should focus on the first argument since I feel the mental health one is very shaky after recent research.

I didn’t start reading the mentioned papers (except for the abstracts). We can define multiple work streams and divide them between possible collaborators. My expertise is mostly in back- of-the-envelope calculations / Fermi estimates given my background as management consutant. I am especially looking for people who like to / are good at assessing the value of scientific papers. 

Please drop a message below or send an e-mail to jan-willem@effectiefaltruisme.nl if you want to participate

Arguments in favour of aligning recommender systems as cause area

Social media causes political polarization 

(http://eprints.lse.ac.uk/87402/1/Social-Media-Political-Polarization-and-Political-Disinformation-Literature-Review.pdf)

• Polarization causes less international collaboration
  • Proved by paper?
    • Brexit: What are the odds that social media was decisive in the outcome?
      • https://journals.sagepub.com/doi/pdf/10.1177/0894439317734157
      • https://comprop.oii.ox.ac.uk/wp-content/uploads/sites/93/2017/12/Russia-and-Brexit-v27.pdf
      • https://www.theguardian.com/commentisfree/2020/jul/21/russian-meddling-brexit-referendum-tories-russia-report-government
    • Trump: What are the odds that social media was decisive in the outcome?
    • Less appetite for EU
  • This increases risks for other (X-)risks:
    • Extreme Climate change scenarios
      • Estimate Trump’s impact on climate (What is the chance that an event like this causes certain amplifiers of climate change to push us towards extreme scenarios?)
        • Direct
        • Indirect through other countries doing less
      • Show that more instead of less international collaboration is needed if you want to decrease the probability of extreme scenarios
    • Nuclear war (increasing Sino-American tensions)
    • AI safety risks from misuse (through Sino-American tensions)
    • (Engineered) Pandemics
      • Can we already calculate extra deaths caused through misinformation?
      • Increased chances of biowarfare
    • Lower economic growth because of trade barriers / protectionist measures
      • Calculate extra economic growth from EU to calculate what it will cost is the EU falls apart?
        • Convert to possible additional QALY’s?

Next steps here:

• Find papers for all the relevant claims
• Look at Stefan_Schuberts counterarguments below (https://forum.effectivealtruism.org/posts/E4gfMSqmznDwMrv9q/are-social-media-algorithms-an-existential-risk)
• Synthesize findings from papers that prove political polarisation
• Show papers that prove that political polarisation led to less international cooperation
• Look for cases (Trump / Brexit) where social media are to be blamed. Calculate the chance that social media actually flipped the election there
• Make back- of-the-envelope calculations / Fermi estimates for all relevant negative consequences

Social media causes declining mental health

• Doesn’t seem to be the case after doing a short review
  • https://journals.sagepub.com/doi/10.1177/2167702618812727)
  • https://lydiadenworth.com/articles/social-media-has-not-destroyed-a-generation/
  • https://www.sciencedirect.com/science/article/abs/pii/S0747563219303723

This one looks interesting as well: https://docs.google.com/document/d/1w-HOfseF2wF9YIpXwUUtP65-olnkPyWcgF5BiAtBEy0/edit#

Next steps here:

• Countercheck these paper
• If we find prove that it causes mental health to decline: calculate lost QALY’s (see https://forum.effectivealtruism.org/posts/xzjQvqDYahigHcwgQ/aligning-recommender-systems-as-cause-area#Scale for first estimate)

Aligning recommender systems is “training practice” for larger AI alignment problems

See https://forum.effectivealtruism.org/posts/xzjQvqDYahigHcwgQ/aligning-recommender-systems-as-cause-area#Connection_with_AGI_Alignment

Next step here:

• Should we expand this argument?

The problem is solvable (but we need more capacity for research)

See e.g. https://www.turing.ac.uk/sites/default/files/2020-10/epistemic-security-report_final.pdf

Next step here:

• Collect additional papers on solutions
• What kind of research is interesting and would be worthwhile investing in?