Algorithmic and Technological Transparency
Today I had a talk on OpenFest about algorithmic and technological transparency. It was a somewhat abstract and high-level talk but I hoped to make it more meaningful than if some random guy just spoke about techno-dystopias.
I don’t really like these kinds of talks where someone who has no idea what a “training set” is, how dirty the data is and how to write five lines in Python talks about an AI revolution. But I think being a technical person let’s me put some details into the high-level stuff and make it less useless.
The problem I’m trying to address is opaque systems – we don’t know how recommendation systems work, why are seeing certain videos or ads, how decisions are taken, what happens with our data and how secure the systems we use are, including our cars and future self-driving cars.
Algorithms have real impact on society – recommendation engines make conspiracy theories mainstream, motivate fascists, create echo chambers. Algorithms decide whether we get loans, welfare, a conviction, or even if we get hit by a car (as in the classical trolley problem).
How do we make the systems more transparent, though? There are many approaches, some overlapping. Textual description of how things operate, tracking the actions of back-office admins and making them provable to third parties, implementing “Why am I seeing this?”, tracking each step in machine learning algorithms, publishing datasets, publishing source code (as I’ve previously discussed open-sourcing some aspects of self-driving cars).
We don’t have to reveal trade secrets and lose competitive advantage in order to be transparent. Transparency doesn’t have to come at the expense of profit. In some cases it can even improve the perception of a company.
I think we should start with defining best practices then move to industry codes and standards, and if that doesn’t work, carefully think of focused regulation (but have in mind politicians are mostly clueless about technology and may do more damage)
The slides from the talk are below (the talk itself was not in English)
It can be seen as our responsibility as engineers to steer our products in that direction. Not all of them require or would benefit from such transparency (rarely anyone cares about how an accounting system works), and it would normally be the product people that can decide what should become more transparent, but we also have a stake and should be morally accountable for what we build.
I hope we start thinking and working towards more transparent systems. And of course, each system has its specifics and ways to become more transparent, but I think it should be a general mode of thinking – how do we make our systems not only better and more secure, but also more understandable so that we don’t become the “masterminds” behind black-boxes that decide lives.
Today I had a talk on OpenFest about algorithmic and technological transparency. It was a somewhat abstract and high-level talk but I hoped to make it more meaningful than if some random guy just spoke about techno-dystopias.
I don’t really like these kinds of talks where someone who has no idea what a “training set” is, how dirty the data is and how to write five lines in Python talks about an AI revolution. But I think being a technical person let’s me put some details into the high-level stuff and make it less useless.
The problem I’m trying to address is opaque systems – we don’t know how recommendation systems work, why are seeing certain videos or ads, how decisions are taken, what happens with our data and how secure the systems we use are, including our cars and future self-driving cars.
Algorithms have real impact on society – recommendation engines make conspiracy theories mainstream, motivate fascists, create echo chambers. Algorithms decide whether we get loans, welfare, a conviction, or even if we get hit by a car (as in the classical trolley problem).
How do we make the systems more transparent, though? There are many approaches, some overlapping. Textual description of how things operate, tracking the actions of back-office admins and making them provable to third parties, implementing “Why am I seeing this?”, tracking each step in machine learning algorithms, publishing datasets, publishing source code (as I’ve previously discussed open-sourcing some aspects of self-driving cars).
We don’t have to reveal trade secrets and lose competitive advantage in order to be transparent. Transparency doesn’t have to come at the expense of profit. In some cases it can even improve the perception of a company.
I think we should start with defining best practices then move to industry codes and standards, and if that doesn’t work, carefully think of focused regulation (but have in mind politicians are mostly clueless about technology and may do more damage)
The slides from the talk are below (the talk itself was not in English)
It can be seen as our responsibility as engineers to steer our products in that direction. Not all of them require or would benefit from such transparency (rarely anyone cares about how an accounting system works), and it would normally be the product people that can decide what should become more transparent, but we also have a stake and should be morally accountable for what we build.
I hope we start thinking and working towards more transparent systems. And of course, each system has its specifics and ways to become more transparent, but I think it should be a general mode of thinking – how do we make our systems not only better and more secure, but also more understandable so that we don’t become the “masterminds” behind black-boxes that decide lives.