Dec. 10, 2020, 9:07 a.m.
Ethical Technologist Issue 12: From _The Great Instauration_ to Jeffrey Eppstein: The inevitability of bias in AI.
The Ethical Technologist
Howdy again!
This week Chip and I delve deep into the world of bias and artificial intelligence, and how those who profit from this technology work to shape public discourse.
One of my first academic talks, so long ago when I was an undergraduate torn between AI work and philosophy, was a critique of Francis Bacon’s Instauratio magna. Bacon was one of the earliest proponents of this newfangled thing called “method”, and he saw method (or what we might today recognize as “algorithm”) as foundational to eliminating human bias from scientific judgments. His work formed the basis of what we now are taught as the “scientific method”, but he was especially keen on issues of classification—something that biologists had recently started doing. Classification was seen as extracting objective, abstract truths about the universe from individual observations, and the scientists of the era believed the practice gave them unique and unbiased insight into God’s creation.
I was therefore struck during my studies by the first chapter of my undergraduate machine learning textbook, which emphasized that classification cannot proceed without bias. To decide when two things are different, or are the same, you need to first decide which features are important, how to measure those features, and what differences in measurement constitute “difference” vs merely “variation”. There can be no classification without bias. Jorge Luis Borges touches on this same point, some forty years before that moment, in his short story “Fuñes the Memorious”, about a boy who retains perfect memory of everything he encounters, and thus loses the ability to discern abstractions: He cannot create or understand classifications, because all he sees are an infinitude of differences.
Bacon was wrong. And yet, what a stranglehold Bacon’s thesis seems to have still over AI researchers, indeed so much of Western thought.
In this issue: How to build unjust bias into an AI model, the politics of machine translation, the standard head, the fiction of “fair” AI, and the facial recognition feint.
If you have a recommendation for a theme, or an article you’ve seen recently that you think I ought to share, please do let me know. You can reply to this email, or hit me up on Twitter
As always, you can find back issues of The Ethical Technologist in the archives. And if you found this issue thought-provoking and informative, please share with your friends and colleagues!
[ 🙅♂️ How to Build Bias Into an AI Model ]
Training bias in machine learning is a well-known phenomenon, yet even in the tech industry too many mistake a trained model for objective truth. This paper explores the problem in the context of one of the biggest recent fads in natural language processing: word embeddings.
The fact that it is commonplace to use pre-trained models from tech giants like Google, rather than collecting data and training one’s own model, only compounds the problem. Those who are providing models that could potentially be used by millions have an added responsibility to reduce bias, and to be transparent about what biases do exist.
[CH]
[ 🌅 The Politics of Machine Translation ]
Speaking of language models. If you haven’t heard, the co-lead of Google’s Ethical AI team, Timnit Gebru, was recently fired. Many have speculated the firing was triggered by her unpopular opinions about the ethics of Google’s AI programs. Here’s a look at the paper that may have led to her firing. It seems that bias isn’t the only problem with these large language models: there are environmental and social costs as well.
> “We are working at a scale where the people building the things can’t actually get their arms around the data,” she said. “And because the upsides are so obvious, it’s particularly important to step back and ask ourselves, what are the possible downsides? … How do we get the benefits of this while mitigating the risk?”
There’s also a reminder that no amount of data can take the place of a human in the loop:
> In 2017, Facebook mistranslated a Palestinian man’s post, which said “good morning” in Arabic, as “attack them” in Hebrew, leading to his arrest.
[CH]
[ 👱🏻♂️ The Standard Head ]
A recent art exhibition at the Pace Gallery dug into the history of unjust bias in automatic classification. Of course, all classification demands bias; there are an infinite number of ways to group things together, but we as humans must select the relevant metrics, and the relevant exemplars, to train a classification system. Trevor Paglen’s art displays the subtle and invisible ways that unjust selections and unjust outcomes creep into the mainstream.
> Norms, classifications and categories always have a politics to them … What kinds of judgment are built into technical systems? Why are they made that way, who are they benefiting, and at whose expense?
[DGW]
[ 🎓 The Fiction of “Fair” AI ]
As tech companies come under increasing scrutiny about how they use and deploy artificial intelligence, corporations are rolling out “ethical” principles to make AI “fair” and remove bias. TODO wrote last year about how companies are increasingly leaning on academia, and in particular on Joichi Ito’s ethical AI program at the MIT Media Lab (now defunct because of his connections to Epstein), to control the narrative around ethics and AI.
[DGW]
[ 📹 The Facial Recognition Feint ]
Fast-forward six months. Microsoft has a staggering array of mass surveillance programs for sale to police forces and government entities, and is partnering with the NYPD and other departments to build highly controversial and unjust predictive policing systems. Yet Microsoft has largely escaped scrutiny by focusing negative press coverage on facial recognition—a technology that, while deeply problematic, is but the tip of the iceberg. The public commitments to not sell facial recognition technologies, Michael Kwet argues, are carefully designed to distract from just how involved Microsoft really is in crafting surveillance states for profit.
[DGW]
[ ❤️ Sponsor ]
Does your business need a developer strategy? You’ve heard of developer relations, but what is that? Do you even need it? Katsudon.tech can help you navigate the complex world of developer relations. Every business is different, and we can help you evaluate your developer goals, determine the best way to get started, and ensure you feel confident that you are spending your money effectively. To find out more, and contact us to set up an initial consultation, visit our website: https://katsudon.tech/.
So much for Issue 12! Thanks for subscribing and reading, and thanks to the many folks who have continued to share content with us! If you enjoyed this issue, share with your friends! If you have an article that should be featured in an upcoming issue of The Ethical Technologist, let me know by either replying to this email (what, a newsletter you can reply to!?), or pinging me on Twitter.
Until next time, yours, Chip Hollingsworth & Don Goodman-Wilson