New APPS: Art, Politics, Philosophy, Science

Google fires another AI ethicist

This time Margaret Mitchell, one of the other authors on the fabulous "Stochastic Parrots" paper (that's my post on it.  The paper is here) on natural language processing.  This was obviously coming, since they'd suspended her email account weeks ago.  In case you haven't read the paper (and you really should!), it's worth mentioning that the paper does not mention Google, and the goal is to point to strategies to make AI better, i.e., less likely to entrench racism, sexism and so forth while destroying the planet in climate catastrophe.  It turns out that those are related, because bigger and bigger datasets are both carbon intensive and likely to pick up on and magnify hegemonic and toxic speech.  Attention to efficiency and curation moves the needle on both.

No commitment to even looking like they care about ethical AI.

Reading List: Bender and Gebru, “Stochastic Parrots”

By Gordon Hull

Not long ago, Google summarily dumped Timnit Gebru, one of its lead AI researchers and one of the few Black women working in AI.  Her coauthor Emily Bender has now posted the paper (to be presented this spring) that apparently caused all the trouble.  It should be required reading for anybody who cares about the details of how AI and data systems can perpetuate racism, or who cares more generally about the social implications of brute-force approaches to AI.  Bender and Gebru take up a common approach to natural-language processing (NLP), which involves an AI system learning how to anticipate what speech is likely to follow a given unit of speech.  If, for example, I say “Hello, how are,” the system learns by studying a dataset of existing phrases and text snippets that the next word is likely to be “you,” but almost certainly will not be “ice cream.”  How good the computer gets at this game is going to be substantially determined by the quantity and quality of its training data, i.e., the text that it examines.

Bender and Gebru outline the social costs of one approach to this problem, which is basically brute force.  As processing power increases, it’s possible to train computers with larger and larger datasets, and the use of larger datasets reliably improves system performance.  But should we be doing that?  Bender and Gebru detail several kinds of problems.  The first is environmental justice: all that processing power uses a lot of energy.  Although some of it may come from carbon-neutral sources, the net climate cost is significant.  Worse, the NLP systems being produced don’t benefit the people that are going to suffer the most from climate change.  As they memorably put it:

“Is it fair or just to ask, for example, that the residents of the Maldives (likely to be underwater by 2100) or the 800,000 people in Sudan affected by drastic floods, pay the environmental price of training and deploying ever larger English LMs, when similar large-scale models aren’t being produced for Dhivehi or Sudanese Arabic?”