Thread · 6 tweets · 10 Apr 2020

I have been thinking along similar lines as this thread (h/t @djleufer & @mariafarrell) but with somewhat different views. I think that ML has its own somewhat novel position in the "Seeing Like A State" framework. 🧵

ML doesn't *create* the legibility. It can only operate if there is already data, so the legibility has to already exist. It does however make the data more actionable such that you can mould the world to match the utility function that trained the ML.

This matches the James C. Scott framework since he discusses the State changing the world back. You start wanting to organise the world's knowledge, you end up organising the world to match your knowledge model. Building from a utility function is also very high modernist.

One distinguishing aspect of ML is that it makes the State more illegible to its subjects. This strengthens the power imbalance beyond what Scott discusses and I think is novel.

Another novel aspect is that ML enables scale with fewer natural checks and balances (from physical realisability) than anything in Seeing Like A State. And scale without checks & balances is inherently bad.

That said I reach similar conclusions to @emmibevensee that ML is not inherently bad but rather requires adapted governance. In fact, ML can be used *with* mētis. In fact there's an upcoming @timesopen post on a sociotechnical system associating ML with people for good results!