9 ms·

I talked to the Data scientists at Cambridge Analytica before and after the scandal, including in-depth on the technical details. Couple of things that they tol

by bertil 7y ago

I talked to the Data scientists at Cambridge Analytica before and after the scandal, including in-depth on the technical details. Couple of things that they told me that change a bit the interpretation you can have of the scandal: - There was a moat between the operations/sales team (Alex Nix, Brittany Kaiser, etc.), the people you see on most documentaries and the technical team (Alex Taylor, etc.): Sales would say anything to convince buyers (campaigns) and had no idea how the technology works. You can take any sentences uttered by them and quote it to the data scientists, their reaction would be… bow, shake their head, “Yeah, not, that’s… not it.” Probably the worst case of over-selling I’ve heard of. - The company kept Kogan’s data in archives (breaching their agreement with Facebook) but overall, data scientists considered it useless because it was ageing, not helpful for models. They never used it in production. - Instead (and you can see that at the 1:00:00 mark in _The Big Hack_) they used election data (participation, name, demographic, address), which magazine people are subscribed to (name, address and the title) and credit information, notably what car they are paying back (name, address, make and model of the car). How are magazines and car helpful? Well, in their words: “guess who the driver of a Ford-150, reader of _Guns & Ammos_ vote for?” Both are a well-documented source of political insight, so much that there’s an open-source model for parsing Google Street View, classifying the car parked and making precinct-level predictions. - Kogan data could not be used on Facebook (it was blacklisted) but they didn’t know that.

johnfactorial 7y ago
> There was a moat between the operations/sales team... and the technical team. It's one thing to claim two teams _don't_ work together regularly; it's another to say there's a moat between them, indicating it was difficult or impossible for the teams to interact due to an external factor designed into the organization's structure. What does this moat mean practically, and why is this claimed division pertinent to the points the documentary makes?
- bertil 7y ago
  Intellectual moat: the sales team refused to acknowledge that over-selling was going to hurt them in the long run; the technical team regularly asked them to re-phrase copy on sales material that was not technically true, the sales would say yes and not do it, or change the written statement but make the false claim, verbatim, in front of cameras. Numbers in their presentations didn’t add up, they confused information and inferences, interpolation, models and individual data, etc.
ryanmarsh 7y ago
there’s an open-source model for parsing Google Street View, classifying the car parked and making precinct-level predictions Mind blown. The lengths campaigns will go to for the win. Can you share any more stories like that?
- Mikeb85 7y ago
  > The lengths campaigns will go to for the win. In general, if there's enough money to use for a specific goal, humans will go to great lengths to win. Look at finance companies building microwave towers to lower trade latency, hiring the top minds for salaries in the millions, building in-house supercomputers, etc... Obama famously used technology in his campaign, why people are surprised that the other side is using it is beyond me. Also, while individuals can make unique decisions, in general humans are very predictable, which is why fields like economics and sociology are even able to exist.
  - bertil 7y ago
    > Obama famously used technology in his campaign, why people are surprised that the other side is using it is beyond me. I suspect a lot of the difference in judgement is less about the tools and more about the messaging that it allowed: What CA did, quite effectively under the brand SCL Election, was dog-whistling and demotivation campaign, like “Do So” in Trinidad presented in the Big Hack. If you can tell which one of your voters are racist, you can be a lot more blunt about what you mean by “revisit immigration policies”. People suspect that targetting allowed Trump to talk as if no one else was in the room — although, it’s not like he was very respectful in public. One the other hand, people are likely to assume that Obama’s analytics, even if they were more detailed, where about pointing potential voters are the relevant part of his program: you are a teacher? Let’s put the promises about increasing the budget for education in bold. Veteran? Talk about red-tape at the VA. Everybody can see those, but people might not care as much. This is still targeting, probably more effective than cloak-and-dagger messaging, but it doesn’t come off as problematic. Of course, a lot of the tools are similar: scraping Facebook Graph should be frowned upon in both cases, it wasn’t in 2008, 2012; it was in 2016, although it’s hard to see if that’s more awareness, and from whom, or partisanship. I do think that the distinction is key, and why Facebook’s main visible effort was to make all political ads publicly visible: it avoided to let the argument about secret offensive ads fester. That might have some unintended consequences, though.
amrrs 7y ago
Michael Lewis' Podcast `Against the rules` did a good job in bringing a different perspective of the entire thing Kogan did with data was not significant yet he was made a scapegoat by journalists https://atrpodcast.com/episodes/the-alex-kogan-experience-s1!d20f3 https://atrpodcast.com/episodes/the-alex-kogan-experience-s1...
- holografix 7y ago
  That is indeed a very interesting perspective
- bertil 7y ago
  Ignoring the unrelated 30-min digression about grammar (sic), I’m not tempted to excuse him: he shared personal-level data, including names, with a company that was openly Orwellian. They didn’t hide what they were doing. People started paying attention when they went from sowing division and crafting racist campaign in South-East Asia and Africa to Western countries, but they did not hide their methods before or after. I guess the podcast is making a good case that he’s awkward and naive but that shouldn’t prevent him from seeing the consequence of his sharing his research data. The biased story seems to ignore that Facebook did ask CA to delete the data as soon as they knew about the sale. I can’t imagine that Kogan didn’t hear about that. On that, I wouldn’t be able to say anything more relevant than the last sentence of the podcast on “creating your own reality”. Kogan’s avoiding public opinion and his self-excuses are blocking the entire research on social networks. Now, Facebook PR expects that any similar academic collaboration will end badly — and I’m not surprised that those are stalled.
bart_spoon 7y ago
> Sales would say anything to convince buyers (campaigns) and had no idea how the technology works. You can take any sentences uttered by them and quote it to the data scientists, their reaction would be… bow, shake their head, “Yeah, not, that’s… not it.” This is frankly par for the course in data science. I work as a data scientist/statistician for a relatively newly established (~5 years) analytics team at a large company, and the first order of business was weaning our teams off of the snake oil that grifter consulting companies were making millions selling them. The things I've seen referred to as "machine learning" and "AI" are absolutely bonkers. The reality is that most of the analytics services being sold B2B is total garbage, and yet it thrives because most customers aren't aware of why its bad. To be fair to these consulting companies, my impression is at least some of the time they aren't knowingly selling BS, and that often times they just have very poor analysts/data scientists who have a very surface level knowledge of what they are doing, but not enough depth to realize why they can't do/aren't doing what they think they are. And the actual sales people know literally nothing but buzzwords. These days we've been able to largely get the company away from vendors where possible and to rely on internal analytics teams, but it is still an issue when our team will come back to business partners with an answer someone doesn't want to hear, and they have external vendors essentially promising they can solve literally any problem with AI, and magically always produce the safe answer the customer wants to hear (but they never are willing to share code, the low level details of their process, or really anything that can make their analyses reproducible).
- ryanmarsh 7y ago
  Relevant quote from TFA: Dave Karpf, who has written two books on the use of data in modern campaigns, calls Cambridge Analytica “the Theranos of political data.” Eight GOP political consultants told Ad Age’s Kate Kaye that the company was “all hat and no cattle.” Maybe it's because I'm from Texas but "all hat and no cattle" is about as damning an insult as they come. It's the equivalent of "full of shit".
  - pacaro 7y ago
    Or as they would say in Cambridge, "all mouth and no trousers"
1v1id 7y ago
> there’s an open-source model for parsing Google Street View, classifying the car parked and making precinct-level predictions. I tried searching for this but wasn't able to come up with anything. Does anybody know what this is referring to? Edit: Found this in another comment https://news.stanford.edu/2017/11/28/neighborhoods-cars-indicate-political-leanings/ https://news.stanford.edu/2017/11/28/neighborhoods-cars-indi...
mzs 7y ago
As an outsider to me the worst was targeting people that were paranoid (also not in some technical psychological sense, again as an outsider) while violating agreements and exploiting that for elections.
- bertil 7y ago
  That would be the worst if it happened. I don’t think that Cambridge Analytica had the ability to do that.
  - mzs 7y ago
    I appreciate your POV, thanks. I do think psychographic profiling was pretty roundly proven by reporting and testimony at this point though. You might be interested in this piece from today though, it takes a bigger view outside CA as well: "It was somewhat surprising that Facebook succeeded in hogging the limelight in a field so rich with villains. The bigger picture…was the way in which the Cambridge Analytica story opened a window onto a new constellation of international billionaires, corrupt politicians, and war profiteers who were apparently amassing enormous power. That story isn’t only about technology, data, and psychographic profiling; it’s also, at root, a story about the consequences of entrenched economic inequality, the privatization of essential public assets and government functions, including even national security, and the challenge to conventional foreign policy posed by the bargains being struck between international kleptocrats…" https://www.nybooks.com/daily/2019/08/27/the-oligarch-threat/ https://www.nybooks.com/daily/2019/08/27/the-oligarch-threat...

by bertil 7y ago

johnfactorial 7y ago

> There was a moat between the operations/sales team... and the technical team. It's one thing to claim two teams _don't_ work together regularly; it's another to say there's a moat between them, indicating it was difficult or impossible for the teams to interact due to an external factor designed into the organization's structure. What does this moat mean practically, and why is this claimed division pertinent to the points the documentary makes?

bertil 7y ago
Intellectual moat: the sales team refused to acknowledge that over-selling was going to hurt them in the long run; the technical team regularly asked them to re-phrase copy on sales material that was not technically true, the sales would say yes and not do it, or change the written statement but make the false claim, verbatim, in front of cameras. Numbers in their presentations didn’t add up, they confused information and inferences, interpolation, models and individual data, etc.

ryanmarsh 7y ago

there’s an open-source model for parsing Google Street View, classifying the car parked and making precinct-level predictions Mind blown. The lengths campaigns will go to for the win. Can you share any more stories like that?

Mikeb85 7y ago
> The lengths campaigns will go to for the win. In general, if there's enough money to use for a specific goal, humans will go to great lengths to win. Look at finance companies building microwave towers to lower trade latency, hiring the top minds for salaries in the millions, building in-house supercomputers, etc... Obama famously used technology in his campaign, why people are surprised that the other side is using it is beyond me. Also, while individuals can make unique decisions, in general humans are very predictable, which is why fields like economics and sociology are even able to exist.
- bertil 7y ago
  > Obama famously used technology in his campaign, why people are surprised that the other side is using it is beyond me. I suspect a lot of the difference in judgement is less about the tools and more about the messaging that it allowed: What CA did, quite effectively under the brand SCL Election, was dog-whistling and demotivation campaign, like “Do So” in Trinidad presented in the Big Hack. If you can tell which one of your voters are racist, you can be a lot more blunt about what you mean by “revisit immigration policies”. People suspect that targetting allowed Trump to talk as if no one else was in the room — although, it’s not like he was very respectful in public. One the other hand, people are likely to assume that Obama’s analytics, even if they were more detailed, where about pointing potential voters are the relevant part of his program: you are a teacher? Let’s put the promises about increasing the budget for education in bold. Veteran? Talk about red-tape at the VA. Everybody can see those, but people might not care as much. This is still targeting, probably more effective than cloak-and-dagger messaging, but it doesn’t come off as problematic. Of course, a lot of the tools are similar: scraping Facebook Graph should be frowned upon in both cases, it wasn’t in 2008, 2012; it was in 2016, although it’s hard to see if that’s more awareness, and from whom, or partisanship. I do think that the distinction is key, and why Facebook’s main visible effort was to make all political ads publicly visible: it avoided to let the argument about secret offensive ads fester. That might have some unintended consequences, though.

amrrs 7y ago

Michael Lewis' Podcast `Against the rules` did a good job in bringing a different perspective of the entire thing Kogan did with data was not significant yet he was made a scapegoat by journalists https://atrpodcast.com/episodes/the-alex-kogan-experience-s1!d20f3 https://atrpodcast.com/episodes/the-alex-kogan-experience-s1...

holografix 7y ago
That is indeed a very interesting perspective
bertil 7y ago
Ignoring the unrelated 30-min digression about grammar (sic), I’m not tempted to excuse him: he shared personal-level data, including names, with a company that was openly Orwellian. They didn’t hide what they were doing. People started paying attention when they went from sowing division and crafting racist campaign in South-East Asia and Africa to Western countries, but they did not hide their methods before or after. I guess the podcast is making a good case that he’s awkward and naive but that shouldn’t prevent him from seeing the consequence of his sharing his research data. The biased story seems to ignore that Facebook did ask CA to delete the data as soon as they knew about the sale. I can’t imagine that Kogan didn’t hear about that. On that, I wouldn’t be able to say anything more relevant than the last sentence of the podcast on “creating your own reality”. Kogan’s avoiding public opinion and his self-excuses are blocking the entire research on social networks. Now, Facebook PR expects that any similar academic collaboration will end badly — and I’m not surprised that those are stalled.

bart_spoon 7y ago

> Sales would say anything to convince buyers (campaigns) and had no idea how the technology works. You can take any sentences uttered by them and quote it to the data scientists, their reaction would be… bow, shake their head, “Yeah, not, that’s… not it.” This is frankly par for the course in data science. I work as a data scientist/statistician for a relatively newly established (~5 years) analytics team at a large company, and the first order of business was weaning our teams off of the snake oil that grifter consulting companies were making millions selling them. The things I've seen referred to as "machine learning" and "AI" are absolutely bonkers. The reality is that most of the analytics services being sold B2B is total garbage, and yet it thrives because most customers aren't aware of why its bad. To be fair to these consulting companies, my impression is at least some of the time they aren't knowingly selling BS, and that often times they just have very poor analysts/data scientists who have a very surface level knowledge of what they are doing, but not enough depth to realize why they can't do/aren't doing what they think they are. And the actual sales people know literally nothing but buzzwords. These days we've been able to largely get the company away from vendors where possible and to rely on internal analytics teams, but it is still an issue when our team will come back to business partners with an answer someone doesn't want to hear, and they have external vendors essentially promising they can solve literally any problem with AI, and magically always produce the safe answer the customer wants to hear (but they never are willing to share code, the low level details of their process, or really anything that can make their analyses reproducible).

ryanmarsh 7y ago
Relevant quote from TFA: Dave Karpf, who has written two books on the use of data in modern campaigns, calls Cambridge Analytica “the Theranos of political data.” Eight GOP political consultants told Ad Age’s Kate Kaye that the company was “all hat and no cattle.” Maybe it's because I'm from Texas but "all hat and no cattle" is about as damning an insult as they come. It's the equivalent of "full of shit".
- pacaro 7y ago
  Or as they would say in Cambridge, "all mouth and no trousers"

1v1id 7y ago

> there’s an open-source model for parsing Google Street View, classifying the car parked and making precinct-level predictions. I tried searching for this but wasn't able to come up with anything. Does anybody know what this is referring to? Edit: Found this in another comment https://news.stanford.edu/2017/11/28/neighborhoods-cars-indicate-political-leanings/ https://news.stanford.edu/2017/11/28/neighborhoods-cars-indi...

mzs 7y ago

As an outsider to me the worst was targeting people that were paranoid (also not in some technical psychological sense, again as an outsider) while violating agreements and exploiting that for elections.

bertil 7y ago
That would be the worst if it happened. I don’t think that Cambridge Analytica had the ability to do that.
- mzs 7y ago
  I appreciate your POV, thanks. I do think psychographic profiling was pretty roundly proven by reporting and testimony at this point though. You might be interested in this piece from today though, it takes a bigger view outside CA as well: "It was somewhat surprising that Facebook succeeded in hogging the limelight in a field so rich with villains. The bigger picture…was the way in which the Cambridge Analytica story opened a window onto a new constellation of international billionaires, corrupt politicians, and war profiteers who were apparently amassing enormous power. That story isn’t only about technology, data, and psychographic profiling; it’s also, at root, a story about the consequences of entrenched economic inequality, the privatization of essential public assets and government functions, including even national security, and the challenge to conventional foreign policy posed by the bargains being struck between international kleptocrats…" https://www.nybooks.com/daily/2019/08/27/the-oligarch-threat/ https://www.nybooks.com/daily/2019/08/27/the-oligarch-threat...