34c3: Dude, you broke the Future!

Ich hole immer noch Vorträge vom 34c3 nach – hier eine YT-Playlist – und habe mir gestern endlich Charles Stross' Keynote angesehen, hier sein Transkript.

Stross redet in seinem Vortrag über Manipulationsmöglichkeiten durch AI, die Erzeugung von gefälschtem audiovisuellem Material mit einem Klick, das nächste Level der Fake News und Unknown Unknowns, die alle Vorhersagen über den Haufen schmeißen und die anscheinend immer unkalkulierbarer werden. Vor allem redet er aber über juristische Personen (also Corporations) als künstliche Intelligenzen, ein noch relativ junger Gedanke, den ich zuerst von Ted Chiang (Arrival aka The Story of your life) in seinem Essay für Buzzfeed gelesen hatte und der ein paar Wochen zuvor von Jeremy Lent formuliert wurde. Ich überlege die ganze Zeit, ob „Corporation as AI“ ein Gleichnis oder eine Metapher ist – aber tatsächlich denke ich, eine juristische Person ist exakt das: Eine künstliche Intelligenz. Guter Gedanke!

Wenn ihr euch auch nur einen Talk des Kongresses anschaut, lasst es den von Charlie Stross sein.

If I predict that in 2027 LTE cellular phones will be everywhere, 5G will be available for high bandwidth applications, and fallback to satellite data service will be available at a price, you won't laugh at me. It's not like I'm predicting that airliners will fly slower and Nazis will take over the United States, is it? And therein lies the problem: it's the 1% of unknown unknowns that throws off all calculations. As it happens, airliners today are slower than they were in the 1970s, and don't get me started about Nazis. Nobody in 2007 was expecting a Nazi revival in 2017, right? (Only this time round Germans get to be the good guys.)

My recipe for fiction set ten years in the future used to be 90% already-here, 9% not-here-yet but predictable, and 1% who-ordered-that. But unfortunately the ratios have changed. I think we're now down to maybe 80% already-here—climate change takes a huge toll on infrastructure—then 15% not-here-yet but predictable, and a whopping 5% of utterly unpredictable deep craziness. […]

Someone out there is working on it: a geolocation-aware social media scraping deep learning application, that uses a gamified, competitive interface to reward its "players" for joining in acts of mob violence against whoever the app developer hates. Probably it has an inoccuous-seeming but highly addictive training mode to get the users accustomed to working in teams and obeying the app's instructions—think Ingress or Pokemon Go. Then, at some pre-planned zero hour, it switches mode and starts rewarding players for violence—players who have been primed to think of their targets as vermin, by a steady drip-feed of micro-targeted dehumanizing propaganda delivered over a period of months.

And the worst bit of this picture?

Is that the app developer isn't a nation-state trying to disrupt its enemies, or an extremist political group trying to murder gays, jews, or muslims; it's just a paperclip maximizer doing what it does—and you are the paper.