The Rise of the Weaponized AI Propaganda Machine
February 19, 2017 11:12 AM   Subscribe

There’s a new automated propaganda machine driving global politics. How it works and what it will mean for the future of democracy.
posted by a_curious_koala (40 comments total) 20 users marked this as a favorite
 
...and yet, were it not for the effed up electoral college system, Trump would have lost. So, I dunno.
posted by Klaxon Aoooogah at 11:18 AM on February 19, 2017 [7 favorites]




It's OCEAN shit, which is basically snake oil. They talk about the implementation as if it's a POMDP but it's prolly a buncha simple linear models. They were doing advertising like this in the 00's. The 1900's.

AI competence writes its own ticket too much to get in bed with these sorts of people. If Google will pay you $500k/yr to write multiarmed bandits with deep features why would you take a job for 80kquid/yr with these blowhards?
posted by hleehowon at 11:26 AM on February 19, 2017 [6 favorites]


(... and the real AI competence gets paid $millions/yr. Not single-digit millions, either, for a few)
posted by hleehowon at 11:28 AM on February 19, 2017 [2 favorites]


were it not for the effed up electoral college system

Good or bad, it's a well understood mechanism by all players involved, and they ran their campaigns to win the electoral college, not the popular vote. Clinton may have been robbed, but it wasn't the EC that did it.
posted by fatbird at 11:33 AM on February 19, 2017 [5 favorites]


POMDP

For a strange few moments I thought this was an odd use of the old coding/IT-support jargon "Phase-Of-Moon Dependent".
posted by Greg_Ace at 11:35 AM on February 19, 2017 [1 favorite]


Partially observed markov decision process

Actually means pretty much the same thing as "phase of moon dependent", kinda. It's stating that there's some markov process (with no permanent memory, maybe a little impermanent memory) that you're imagining the thing you're studying to be and then you're only observing a little bit of the whole of the markov process. You deal with the partial observation by adding more stochasticity

Shipped in production in a looooot of ad networks, for pricing and bidding
posted by hleehowon at 11:37 AM on February 19, 2017 [2 favorites]


That Buzzfeed article makes it sound like a typical start-up-y scam, like Theranos but more confusing to outsiders and so more plausible.

Not that we shouldn't all dial back on Facebook for a lot of reasons, but still.

What I never understood about Theranos was why they named it something that sounded so much like "Thanatos". It didn't sound healthy to me.
posted by Frowner at 11:41 AM on February 19, 2017 [7 favorites]


We can look at the results of the election and say, "under different rules, these results would have produced a different outcome." But under different rules, the candidates would have campaigned differently. Are we sure that if we fixed gerrymandering and got rid of the electoral college that these techniques would not work? Why?
posted by rustcrumb at 11:52 AM on February 19, 2017 [6 favorites]


Can I just mention that "Al" from the Weird Al post below and AI as in this post continue to be too similar in my browser display fonts and so the Al/AI confusion remains strong with me.
posted by hippybear at 11:59 AM on February 19, 2017 [8 favorites]


Either way, it's all about the Pentiums.
posted by Greg_Ace at 12:06 PM on February 19, 2017 [6 favorites]


I seem to have ninety-nine problems, but a beach/birch/female dog is not among them.
posted by hippybear at 12:08 PM on February 19, 2017 [1 favorite]


Weird AI is an app that will write automated parodies of your favorite songs. Enter a topic, choose a song, post to Facebook. Skynet's arrival will be soundtracked by uncanny valley Drake songs about the high school chemistry teacher. And accordions.
posted by wemayfreeze at 12:09 PM on February 19, 2017 [4 favorites]


The Revolution Will Not Be Al-ivized!
posted by hippybear at 12:10 PM on February 19, 2017 [1 favorite]


It's like the people who wrote this article have no idea how online advertising, and data collection/analysis on end users has worked for over 15 years.
posted by barkingpumpkin at 12:24 PM on February 19, 2017 [2 favorites]


On the one hand, I work in this field and these articles are more than a little hysterical. On the other hand, if anyone asks me what I think about these issues, I'm more and more outspoken in my opinion that the best thing you can do for yourself and for the world is to get the fuck off social media. I do feel like when my friends and family talk about Facebook and Twitter it's incredibly clear to me that most of them aren't engaged with these tools in a healthy or constructive way.
posted by potrzebie at 12:44 PM on February 19, 2017 [12 favorites]


it wasn't the EC that did it

I think it was. The fact that a system is well understood and candidates do their best within it does not mean it's fair. It could be bizarrely distorted and still be well understood. Most people's metric of fairness in an electoral system would be whether it delivers what the majority actually want. It seems relatively hard to argue that the popular vote isn't the best indicator there (you might argue that the USA is merely a federation, not a nation, so that the idea of 'what the majority of Americans want' is incoherent, though that seems implausible to me). The fact that the EC delivered the 'wrong' candidate by the standard of the popular vote seems to me enough to support the idea that Clinton was robbed, however well she understood in advance the means by which the robbery was achieved.
posted by Segundus at 12:58 PM on February 19, 2017 [5 favorites]


Yeah ... "Everyone understood that the rules of the only game in town were rigged before they started playing" does not mean that the game was fair.
posted by kyrademon at 1:20 PM on February 19, 2017 [9 favorites]


Modern propaganda, which arose out of the field of psychology and mass media technology, has been with us for over 100 years. It is a problem to be sure - we rely on media to make decisions about the world, and all outlets are biased to a particular elite perspective, and distorted by revenue needs. I agree that these articles are all hysterical, because all that is happening here is that PR gets more granular with time. If a PR firm could tailor messages to 10 demographics back in the 60s, now it's maybe all the way down to the individual level, although that is probably being too generous to these startups. The core concept of public manipulation hasn't changed.
posted by MillMan at 1:37 PM on February 19, 2017 [2 favorites]


The core concept of public manipulation hasn't changed.

Time to watch Adam Curtis' The Century Of The Self again.
posted by hippybear at 1:42 PM on February 19, 2017 [1 favorite]


It may sound like something out of a rejected Neal Stephenson manuscript, but I can assure you the "pepelisk" is very real, and for a very reasonable consulting fee I can instruct you how to defend against it.
posted by Pyry at 1:42 PM on February 19, 2017 [4 favorites]


What I never understood about Theranos was why they named it something that sounded so much like "Thanatos".

Where anus?
posted by Behemoth at 1:44 PM on February 19, 2017 [2 favorites]


"Everyone understood that the rules of the only game in town were rigged before they started playing"

This isn't it, though. "rigged" means that the output of the mechanism isn't what the understanding of the mechanism would produce. The EC did produce the outcome it should have, based on the vote counts and the state mechanics.

The fact that the EC delivered the 'wrong' candidate by the standard of the popular vote

Comparing the EC to the popular vote is like comparing the final score in a football game to the number of yards gained, and then arguing that the losers really won because they gained more, and shouldn't the winning team always gain more yards? Usually, but it's just a correlation, and this all assumes the actual popular vote is a clear representation of the voting public's preferred candidate. Any campaign worker will tell you that a thousand random factors impact turnout that have nothing to do with one's actual preferred candidate.
posted by fatbird at 1:49 PM on February 19, 2017 [2 favorites]


Can I just mention that "Al" from the Weird Al post below and AI as in this post continue to be too similar in my browser display fonts and so the Al/AI confusion remains strong with me.

Computers can't do Weird Al yet; so, it checks out
posted by thelonius at 1:49 PM on February 19, 2017


Even if this really didn't work in 2016, there's no denying that computing gets cheaper over time, and refinements to this technique will work in the near term future... again technology outpaces regulation.

This worries me.
posted by MikeWarot at 1:54 PM on February 19, 2017 [2 favorites]


This worries me.

I'm not worried, because the more I see of these sorts of "big data", "deep learning" AI efforts, the more I see a lot of noise producing little real benefit. Netflix famously paid a million dollars to the winner of a contest to see who could improve their recommendation algorithm that was already operating, according to them, at 90%+. The winner added a percentage point to that total and collected their money.

And yet, when I go on Netflix, everything that's being recommended to me is pretty obviously just because it's like some other show I watched; meanwhile, I get no surprise recommendations, some obvious ones are being missed (like "you watched all three seasons of this, and now there's a fourth season available"), and not many recommendations actually turn into a regular show for me. Their recommendations engine sucks, as far as I'm concerned. I have no idea where 90% came from. Same thing with Amazon, another supposed big data deep miner with boundless computing power. Yes, Amazon, I bought a pocket knife. That means I don't need another hundred pocket knives. You could probably show me fewer than average, just because you know I've already got one. Same thing with Facebook and Google Ads. If my life is an open book once these companies have full access to my digital life, as they do, why are they so bad at exploiting it?
posted by fatbird at 2:08 PM on February 19, 2017 [15 favorites]


Netflix can't work well because the choice of picking a movie is too granular, but if you look at the AI behind Youtube, you'd probably notice that you get lots of interesting things to watch, because its more fine grained, and less of an investment if you make a bad choice... I find that Youtube recommends lots of good full length features that have *ahem* mysteriously appeared there. Last night I ended up watching a PBS documentary about the building of the Panama canal, for example.

In summary, I believe AI is capable of doing the job of recommending things to watch, just not in the case of Netflix.
posted by MikeWarot at 2:24 PM on February 19, 2017 [2 favorites]


What I never understood about Theranos was why they named it something that sounded so much like "Thanatos". It didn't sound healthy to me.

An abbreviation of the really annoyingly dumb portmanteau "theranostic," a combination of therapeutic and diagnostic. Even more stupid, Theranos' purported main product had no therapeutic aspect whatsoever; it was just a classical diagnostic technology, except that it never actually worked.
posted by Existential Dread at 2:34 PM on February 19, 2017


if you look at the AI behind Youtube

Every time I watch a YouTube video about a video game, I get a bunch of recommendations for Pewdiepie and "FIVE HUNDRED REASONS SJWS ARE RUINING AMERICA". So, YouTube really isn't that much better, in my experience.
posted by tobascodagama at 2:51 PM on February 19, 2017 [7 favorites]


This isn't it, though. "rigged" means that the output of the mechanism isn't what the understanding of the mechanism would produce. The EC did produce the outcome it should have, based on the vote counts and the state mechanics.

The effect produced is contrary to democracy, not some arbitrary game or sport. Also, the contest isn't just between Trump and Clinton, so whatever expectations Clinton's team had cannot be taken as defining an aspect of the entire non-Trump contest. The claim of "rigged" stands.

Usually, but it's just a correlation, and this all assumes the actual popular vote is a clear representation of the voting public's preferred candidate. Any campaign worker will tell you that a thousand random factors impact turnout that have nothing to do with one's actual preferred candidate.

Now you're sneaking in "turnout," which is again a game aspect. Elections matter because they are where the rubber meets the road regarding opinion, they take something that may not have been measured, collapse the waveform, and measure it, and call that result the vote. That's not the best solution, but it's the best one we've got. It's inevitably game-like, but that doesn't mean those game aspects are good or fair. Take it from someone who thinks entirely too much about games. If the mechanisms of election do not reflect the popular will, then it is the election that is at fault.
posted by JHarris at 3:17 PM on February 19, 2017 [1 favorite]


Youtube has been a bit better for me than Netflix, but I still don't find it a very compelling example because for me it only surfaces content that's basically the result of a "more like this" query for certain variables. When I listen to one band's song, I get a a bunch of videos for the same band, and some other artists that are obviously similar, frequently because they're covering the same song.

I think all this talk about AI/Big Data/Deep Learning has a hidden premise: that there's a deep structure to be discovered and exploited by lots of data and computational power; that the reason I like both Longmire and Stranger Things is that I have an unfulfilled desire to live somewhere semi-rural, say, rather than both having dopey sheriffs. It's like this ongoing Freudian fallacy where everything is invested with much more meaning than it actually has, which causes a legitimate fear that others can command what we're not smart enough (or computationally endowed enough) to command in ourselves. And this is exactly what the propaganda machine in the link is trying to have you believe it can exploit so it can charge campaigns millions for their services.

My argument here is that Netflix, Facebook, Youtube, Google et al. are failing to demonstrating that they're doing anymore more intelligent than some fairly naive grouping, tagging, and "more of the same"; and I can't think of anyone with more data, resources, and motive to actually exploit this. The null hypothesis for deep learning should be that some non-AI algorithm can't produce the same output and success rate. I don't believe there's a deep structure there, just a lot of correlation chasing.
posted by fatbird at 3:17 PM on February 19, 2017 [6 favorites]


Fatbird - I buy your argument, thanks for relieving me of a bit of stress. You've changed my mind.

Wouldn't it be ironic if you were an AI? ;-)
posted by MikeWarot at 3:43 PM on February 19, 2017 [1 favorite]


I agree with fatbird. We would see more observable evidence of successful AI in this space if it existed, because the commercial incentives to deploy it are very strong.
posted by i_am_joe's_spleen at 4:06 PM on February 19, 2017


Wouldn't it be ironic if you were an AI? ;-)

How do you feel about you were an AI?

I'm sympathetic to the stress caused by the techno-hegemonist future we're supposed to believe is imminent. I also feel stress the other way, namely because of the implication there's no deep structure to exploit, which implies a very uncertain future. I've heard the motive for conspiracy theorists described as "it doesn't matter if the secret masters are evil, at least there are secret masters!"

the Electoral College is that it was intended to be a safety valve

It was also, like the Senate, a place where small states had outsized influence based on their population, which was an incentive for them to participate and a check, in theory, against larger, denser states that would dominate Congress. Wyoming's 3 EC votes give its voters 3.5x as much heft as a California resident's.
posted by fatbird at 5:12 PM on February 19, 2017 [1 favorite]


Mod note: Couple deleted. If you guys want to talk about the election or the electoral college, take it to the political thread catchall. Let's keep this more tightly focused on the AI Propaganda
posted by Eyebrows McGee (staff) at 5:36 PM on February 19, 2017 [1 favorite]


It's OCEAN shit, which is basically snake oil.

The five-factor model of personality, known as OCEAN or CANOE, was based on extensive literature reviews of psychological research about causal effects of different personality traits. It is way more empirically grounded that Meyers-Briggs typology that was dominant beforehand. On the other hand, I suspect that most of the peer-reviewed psychological literature is much too epistemologically modest to claim that they can predict or manipulate how somebody is going to vote. That's where the snake oil comes in.
posted by jonp72 at 5:51 PM on February 19, 2017 [5 favorites]


What I never understood about Theranos was why they named it something that sounded so much like "Thanatos". It didn't sound healthy to me.

What I didn't understand is why they chose to give their company a name that any moderately smartassed junior high school boy could make a "their anus" joke out of.
posted by jonp72 at 5:54 PM on February 19, 2017


>Time to watch Adam Curtis' The Century Of The Self again.

hippybear - maybe see Hypernormalisation if you haven't yet done so...
posted by Joseph Gurl at 5:54 PM on February 19, 2017 [2 favorites]


Fight the good fight, SakuraK. Whether we know it or not, you are our Champion, and with you our fates hang in the balance.
posted by Joseph Gurl at 12:24 AM on February 20, 2017


Nothing will happen with this until the first big propaganda botnet ends up inducing multiple hundreds of people to either suicide or murder. If we're lucky it'll be the former.
posted by benzenedream at 11:22 PM on February 20, 2017


« Older You can't go wrong with pizza, unless it's...   |   The Colors of Japanese Internment Newer »


This thread has been archived and is closed to new comments