Friends with no previous interest in AI ethics have been asking me about it recently, so I want to share several underlying concepts about AI & power that are important to understand. 🧵 1/ AI and Power: The Ethical Challenges of Automation, Centralization, & Scale

Based on 20-min talk:

Blog post version: rachel.fast.ai/posts/2023-05-…
2/

Viruses: The Silent Triggers of Autoimmune & Neurodegenerative Diseases (how a simple cold can lead to life-changing disease) 1/

My new post: rachel.fast.ai/posts/2023-03-… Rheumatoid arthritis, Crohn's disease, Multiple sclerosis, Type 1 Diabetes, Lupus, Hashimoto's, & Psoriasis impact a range of body systems, but all are autoimmune diseases.

Developing a lifelong autoimmune disease is often first triggered by an infection. 2/

Even common viruses can have long-reaching, surprising, & devastating consequences. Fortunately, there are simple steps we can take to reduce transmission. 1/

My new post: rachel.fast.ai/posts/2023-03-… The idea that a common childhood virus can quietly hang out in your nervous system, reactivate decades later to cause shingles, and then months AFTER shingles blisters clear up cause blood clots & strokes is mind-boggling to me 2/

theconversation.com/chickenpox-and…

After 12 years working as a data scientist & AI researcher, I have gone back to school for a Masters in Immunology. When I become fascinated by a topic, I want to learn as much as I can. 1/

rachel.fast.ai/posts/2023-02-… My ultimate goal is to apply my machine learning & data ethics skills to immunology, but I want to make sure I fully understand the underlying domain & relevant context first. (With ML, it’s important to not just be a hammer searching for a nail). 2/

I made a mastodon account a few weeks ago: @math_rachel@mastodon.social (in case twitter falls apart)

I haven't used it yet and have mostly been taking a social media break prior to this. 1/ I used to post a lot on twitter (about machine learning, working in tech, algorithmic harms, & later covid). I loved this place and I learned a lot from others here. Thank you to those I met & learned from. (I plan to stick around, but who knows what will happen.) 2/

This is a thorough, incredibly well-researched explanation of the Qld government's unscientific & unethical pursuit of a "wall of immunity", which was doomed to failure before it had even begun. by @ColinKinner

medium.com/@ColinKinner/l… 1/ Please watch the embedded video of Chief Health Office Dr. Gerrard. Note that when he repeatedly says that it is inevitable & necessary for us to all catch covid, he is speaking as someone with significant political power. 2/

I hate the fatalism & false logic of “everyone’s going to get covid many times, so just give up; it doesn’t matter how often (or how close together infections are); no point in trying to keep hospitals, schools, or pharmacies safe.” 1/ First, this becomes a self-fulfilling prophecy. If you tell people there’s no point in avoiding infection (or worse, that they are building a “wall of immunity”, masks are opposite of smiles, etc) and abandon protective policies, that is going to discourage cautious behavior. 2/

My daughter & I had fun together decorating our elastomeric respirators with paint & gemstones. Mine is a GVS Elipse P100 and hers is a @flo_mask. Many thanks to the amazing @nickelpin for her inspiration! 1/ Why wear an elastomeric respirator? They work and are comfortable! A Texas hospital used Elastomeric Respirators to eliminate Health Care Worker infections, and save $40,000 per year. 2/

kens5.com/article/news/h…

It is urgent that we implement community-wide mask wearing in public indoor settings to protect Australians from infections, reinfections, & chronic disability caused by Long-COVID. Please sign the petition 🇦🇺. More in🧵 1/

chng.it/NLK2cCKZpW Having spent year 1 of pandemic in USA & then moving to Aus, my perspective is that understanding of airborne spread is even worse in 🇦🇺 than 🇺🇸. Because covid was so successfully managed via border closures, many people never had to learn about masks, airborne spread, etc 2/

A public health academic saying non-fit-tested N95s are "useless", and that those w/out public health credentials can’t disagree/engage.

This sort of elitism & gate-keeping causes so much harm.

(screenshot: red is PH academic, blue is other person) 1/ First, there is previous evidence that non-fit-tested N95s can still be effective 2/

fast.ai/2022/07/04/upd…

Why masks are still important:
- vaccines offer limited protection from long covid
- unconstrained transmission → more variants evolve, faster
- keep essential, public spaces accessible to high-risk ppl, babies, & immunocompromised
- constant reinfections → mass absenteeism 1/ Being vaccinated & boosted reduces an individual's chance of hospitalization or death, but does little to stop infection (& even initially mild infections can lead to a host of complications later on) nor transmission at this point 2/

https://twitter.com/Bob_Wachter/status/1543780610920947712

Why leaders can’t be honest, direct, & clear about long covid, risk of reinfections, waning immunity, airborne spread, & importance of N95s 🧵 1. Extreme short-term thinking: some people would *choose* to go out less if they knew the risks, which is bad for economy in the short-term (of course, constant sickness, absenteeism, & disabling part of the workforce is also bad for the economy, in both short- & long-term)

I really enjoyed our @DiagnosingML discussion yesterday on algorithmic bias in healthcare AI. I will link to some of the research & ideas I mentioned in this 🧵 1/

https://twitter.com/DiagnosingML/status/1539861823200976897

@yvessj_aquino shared results from interviews w/ 70 different stakeholders in healthcare AI, including software devs, doctors, & startup founders, to explore different conceptions of algorithmic bias. Here is a previous lightning talk he gave on this 2/ fast.ai/2022/03/14/ADS…

The video from our 3rd @AustralianData Ethics Workshop is up. Here are a few highlights 🧵
australiandatascience.net/data-ethics-wo… The current status quo on AI governance often involves harmful silos.
System owners wrongly assume bias metrics have removed ethical risk.
Model design & validation decisions uninformed by system impact.

@lmccalman1 @GradientInst

Recent writing:
- The problem with metrics
cell.com/patterns/fullt…
- Qualitative research is crucial to AI fast.ai/2022/06/01/qua…
- AI harms are collective, not just individual fast.ai/2022/05/17/soc…
- No such thing as not a "math person"
fast.ai/2022/03/15/mat… Reliance on metrics is a fundamental challenge for AI, with David Uminsky @DSI_UChicago @Patterns_CP

https://twitter.com/math_rachel/status/1526040262782636032

6 months ago, Australia largely went all in on a vaccines-only strategy, with > 90% of adults double-vaxxed. Since then:
6,000 deaths
7 million cases
likely 600k - 1 million cases of LongCovid

In response, we are going to double down on vaccines-only... 1/

https://twitter.com/AlboMP/status/1532258784684277760

Even back when masks were being worn (almost nobody does now), I never heard mainstream messages about benefits of N95s (instead of loose-fitting or cloth masks) nor explaining airborne spread. 2/

https://twitter.com/AnnastaciaMP/status/1496000157246451719?t=hDYS4Z9oRge1ad4MHEKt1A&s=19

Qualitative research is crucial in AI to understand:
- more than what is captured by short-term metrics
- what is missed by large-scale studies (which can elide details & overlook outliers)
- circumstances in which data is produced 1/

fast.ai/2022/06/01/qua… @louisa_bartolo Following the thread of any seemingly quantitative issue in AI ethics (such as determining if software to rate loan applicants is racially biased or evaluating YouTube's recommendation system) quickly leads to a host of qualitative questions. 2/

Q: "We take risks every day. Isn't driving in a car riskier than covid if you're vaccinated?"

A: No.

I recently heard this from someone who sees herself as cautious. My longer answer:🧵1/ Between 2017-2019, average of 37,000 Americans died per *year* in car accidents. 80,000 VACCINATED Americans have died of covid in first 5 *months* of 2022. And USA could have 9 MILLION new Long Covid cases this fall alone. (links for numbers below) 2/

In the west, our ideas of harm are largely anchored to an individual impacted by a particular action at a discrete moment in time. Yet harms caused by algorithmic systems are often collective & communal

new by @louisa_bartolo @AdmsCentre & me 1/
fast.ai/2022/05/17/soc… Before passage of environmental laws, many existing legal frameworks were not well-suited to address environmental harms, which are accumulative, build over time, and not necessarily linked to a single discrete incident. Parallels w/ AI. @NathalieSmuha 2/ policyreview.info/articles/analy…

• Any metric is just a proxy for what you really care about
• Metrics can, and will, be gamed
• Metrics tend to overemphasize short-term concerns
• Many online metrics are gathered in addictive environs
• AI is very effective at optimizing metrics
1/

cell.com/patterns/fullt… Algorithms that heavily emphasize student test scores to rate & fire teachers, has led to widespread cheating scandals, as well as beloved teachers being fired (eg if student scores drop when current school DOESN'T cheat, teacher is penalized)
washingtonpost.com/local/educatio… 2

I took a twitter break for a while, b/c I’d been repeating myself for 2+ years about masks, post-viral disability, etc. My family will be avoiding unmasked, in-person situations for a long time, and our lives are drastically different than 2019. 1/ I've had brain surgery 2x and am high risk. However, I think the idea of a high-risk vs. low-risk dichotomy can be misleading, as many healthy people are greatly underestimating the risk & potential severity of long covid 2/

https://twitter.com/ahandvanish/status/1504971105269915648