[1] https://blog.photofeeler.com/okcupid-is-wrong-about-smiling-...
They were heavily involved in researching what factors more reliably led to not just better matches, but better relationships -- when you disabled your account, they'd ask if it was because you'd met someone through OkC and ask you to pick who, if you were willing to share.
I don't think there was anything fucked up about it, as long as it was all anonymized and at scale. Trying to understand what messaging strategies worked better or worse could be a major part of figuring out how to improve matches.
Like, one obvious factor could be to match people who send lots of long messages with lots of questions with each other, while a separate set matches people who's messaging style is one sentence at a time. I'm not saying that would necessarily work well, but it's not crazy to research if NLP analysis of messages can produce additional potential compatibility signals.
The whole point of OkC back then was to try to develop as many data-based signals as possible to improve matches.
34. In response to this request, Humor Rainbow gave the Data Recipient access to nearly three million OkCupid user photos. Humor Rainbow’s President and Chief Technology Officer were directly involved in facilitating the data transfer.
35. In addition to user photos, Humor Rainbow shared other personal data with the Data Recipient, including each user’s demographic and location information.A company running NLP on its own user data doesn't have anything to do with third parties.
I'm confused why "as long as" carries so much weight here considering the article that started this discussion. You seem to trust that they stopped their privacy fuckups with third parties. I don't know where your trust comes from.
My privacy problem exists with third parties AND internally within OkCupid. At some point, running NLP over messages necessitates that an analyst or ops person actually has to look at the underlying data at some point.
The 3rd party stuff just exudes something like a privacy "code smell" -- "orthogonality" doesn't really matter to my point because 1. these things don't exist in a vacuum and 2. stench has a way of permeating through layers, regardless of how orthogonal underwear is.
Like, do you trust that OKCupid staff didn't download user messages to their laptops?
Christian Rudder's OKTrends blog (and Sam Yagan's presentation at their acquisition celebration) even spelled out the reason why: some women on OKC (or, more rarely, men) would acquire the "Replies rarely"/red color on their profile, for almost never (<10%) replying to initial messages, which was generally considered to be undesirable behavior, even in the negative (there is value in a negative message: "Thanks but I'm not interested due to age/location/other factor". And also OKC could then measure whether users' stated preferences mismatched preferences inferred from which set of users they message e.g. people who say they're looking for 30-55 for LTR but tend to message people 21-35 for short-term). And before anyone points out that younger more attractive female profiles would get more initial messages than males (up to 200:1 more), OKC used to allow you to set filters on the other user's age/distance/other criteria, so you could automatically filter those out. Also, factor in the usual caveats that many users on dating sites tend to lie about their age/weight/height/location/status/etc.
Anyway, to avoid getting labeled the dread "Replies rarely", some (mostly female) users got in the habit of sending one-liner responses that were ambiguous/non-committal/cryptic/negging. And then not responding further (but without unmatching, which only took a single click). This was making their profile look less undesirable but generating pointless message traffic and reducing the overall utility of the platform at actually attempting to match people (for compatibility, not just initial attraction). Hence, OKC tried to actually measure initial exchanges to figure out which ones led to genuine back-and-forth conversations of 3+ messages (which is an ok proxy for inferring a match, certainly a better proxy than just counting initial messages/likes/votes on photos). Yagan jokingly referred to this as "Every Monday morning, we ask 'How many three-ways did we set up over the weekend?'"). (PS, Rudder and Yagan both stressed that users' names/identities/ identifying characteristics were kept out of the analysis.)
After the 2011 IAC acquisition, most of this platform quality control (and looking for constructive insights) went out the window pretty quickly and the three cofounders moved to OKCupid Labs. But it was good for the brief while it lasted. By 2013 a chainsaw had been taken to most of OKC's unique features, esp. for free users.
As such it wasn't "reading other people's messages". That's according to the OKC founders description of what they did, pre-acquisition. Since they were reasonably upfront about what they did, and since that functionality worked even for free users and they didn't aggressively push premium or gate the features, I believe they were being truthful. Furthermore, after the 2011 acquisition and when they stopped being active on OKC itself, there was a palpable degradation in site quality. (And post-2014, IAC went on to sell personal information about users' substance use etc. to insurers, to which users had never given informed consent).
So I think your date explained things badly and you picked up the wrong end of the stick. It's trivially easy to write a script that strips usernames and identifying information. And it's not too hard to distinguish "Hey baby" or "DTF?" from more meaningful messages. For the platform to do that with an intent to improving matches was strongly positive, not negative.
> For the platform to do that with an intent to improving matches was strongly positive, not negative.
Do you realize how much work that OkCupid has put into their product to make matching intentionally worse (through analysis!)? Take off your rose-tinted glasses!!
Edit: aaaand there's a new article about OkCupid giving facial recognition data to 3rd parties.
"the odds are good, but the goods are odd" may apply here
I think the "only thing" that would make me cherry-pickable from their data is that I used an autoclicker to give everyone a 5 star... I have mixed feelings about doing that, but I got a couple (surprisingly nice) dates out of it that never went anywhere.