3 ms·

Framing it as "human psychological experiments" evokes famous psych experiments like the marshmallow test and the Stanford Prison Experiment. Tech companies run

by gradys 7y ago

Framing it as "human psychological experiments" evokes famous psych experiments like the marshmallow test and the Stanford Prison Experiment. Tech companies run these kinds of experiments occasionally (I believe Facebook is known to have tried to manipulate mood with a Newsfeed change), but I feel like this way of describing it masks the fact that 99% of large scale A/B tests are things like - We slightly changed the color of the submit button - The forward button was removed from the context menu on chat bubbles - An infrastructure change reduces the load time of the comments on articles by 10ms - The weekly ad relevance model update is being certified, yielding a 0.000001 increase in CTR for small segments of the market. On average, much more mundane than "human psychological experiments".

slavik81 7y ago
In academia, those sorts of tests would still need to be approved by the human experimentation ethics board. There are no exceptions for trivial tests.
- AnthonyMouse 7y ago
  > In academia, those sorts of tests would still need to be approved by the human experimentation ethics board. There are no exceptions for trivial tests. That's the point. It's like requiring a report to be filed whenever there is a "use of force" but then applying that rule using the Newtonian definition of force. Sat in your chair? File a report. Stand back up? File a report. Filed a report? File a report. Worse, this kind of thing can happen retroactively. If you discover that your numbers are different than expected, but you hadn't declared any experiment, comparing what changed before and after is the experiment. But you hadn't notified those users that you were doing an experiment because you hadn't expected to have any reason to, so now you can't even have the people with the before and after data communicate with the people who know what changes were made to the system in that time frame because comparing that information would constitute doing the experiment. It's like telling a car company they can't see their sales data when deciding which models to continue producing because it would constitute doing a psychological experiment on what kind of cars people like. (On the other hand, it sounds like the law would only apply to entities the size of Facebook, and screw those guys in general. But it really is kind of a silly rule.)
  - friendzis 7y ago
    > comparing what changed before and after is the experiment. IIUC, in order to do that comparison you still need to collect data. You may throw that data away and your experiment ends right there, you may do analysis on that data, but you said it yourself - it is an experiment.
    - AnthonyMouse 7y ago
      Right, so what are we trying to do here then? Having a notification that you're constantly participating in an open-ended experiment with a purpose to be determined at a future date seems worse than nothing. But if you require a more specific notification before the data is collected then the after the fact analysis doesn't just require user notification, it's inherently prohibited.
      - nojvek 7y ago
        Yeah i’d expect the notification as an opt-in? Do you want to be part of that experiment ?
        AnthonyMouse 7y ago
        The experiment where they change the color of the submit button? What should cause me to care about that? And what does opt-in even look like? No matter whether you want to "participate in the experiment" the submit button still needs to be some color for you, which is the only part of the "experiment" with any direct effect on you. The concern with psychological experiments isn't that they're collecting data. That's a different bailiwick. The major issue with psychological experiments is that they may have significant direct psychological consequences. If you show people only news stories about mass shootings and conflict it may cause them to become violent or suicidal -- which has nothing to do with whether you collect data on it or what you do with it afterwards. The experiment itself is the harm. Which means we would need some kind of principled and efficient way of distinguishing those kinds of "real" experiments from just measuring what happens when you make a subtle adjustment to a context menu.
- Mirioron 7y ago
  Yes, and it's dumb. It's a bureaucratic nightmare that most likely inhibits progress. Not only that, this is also being used as a cudgel to silence the people that did the grievance studies hoax. [0] [0] https://reason.com/blog/2019/01/07/peter-boghossian-portland-irb-hoax-griev https://reason.com/blog/2019/01/07/peter-boghossian-portland...
  - kristianc 7y ago
    Neither does it deliver a guarantee on the results you get out at the end - the Stanford experiment was faked, and academics have struggled to replicate the marshmallow test.
    - Mirioron 7y ago
      While I agree with your point, I do think that some kind of ethics oversight should happen over experiments like the ones you mentioned. I just think that it's absurd to expect the same from simple tests.
    - hopler 7y ago
      I thought marshmallow was replicated but clarified -- that it showed how children react to un/trusted adults, not uncover some genetic propensity to deferred rewards.

by gradys 7y ago

slavik81 7y ago

In academia, those sorts of tests would still need to be approved by the human experimentation ethics board. There are no exceptions for trivial tests.

AnthonyMouse 7y ago
> In academia, those sorts of tests would still need to be approved by the human experimentation ethics board. There are no exceptions for trivial tests. That's the point. It's like requiring a report to be filed whenever there is a "use of force" but then applying that rule using the Newtonian definition of force. Sat in your chair? File a report. Stand back up? File a report. Filed a report? File a report. Worse, this kind of thing can happen retroactively. If you discover that your numbers are different than expected, but you hadn't declared any experiment, comparing what changed before and after is the experiment. But you hadn't notified those users that you were doing an experiment because you hadn't expected to have any reason to, so now you can't even have the people with the before and after data communicate with the people who know what changes were made to the system in that time frame because comparing that information would constitute doing the experiment. It's like telling a car company they can't see their sales data when deciding which models to continue producing because it would constitute doing a psychological experiment on what kind of cars people like. (On the other hand, it sounds like the law would only apply to entities the size of Facebook, and screw those guys in general. But it really is kind of a silly rule.)
- friendzis 7y ago
  > comparing what changed before and after is the experiment. IIUC, in order to do that comparison you still need to collect data. You may throw that data away and your experiment ends right there, you may do analysis on that data, but you said it yourself - it is an experiment.
  - AnthonyMouse 7y ago
    Right, so what are we trying to do here then? Having a notification that you're constantly participating in an open-ended experiment with a purpose to be determined at a future date seems worse than nothing. But if you require a more specific notification before the data is collected then the after the fact analysis doesn't just require user notification, it's inherently prohibited.
    - nojvek 7y ago
      Yeah i’d expect the notification as an opt-in? Do you want to be part of that experiment ?
      - AnthonyMouse 7y ago
        The experiment where they change the color of the submit button? What should cause me to care about that? And what does opt-in even look like? No matter whether you want to "participate in the experiment" the submit button still needs to be some color for you, which is the only part of the "experiment" with any direct effect on you. The concern with psychological experiments isn't that they're collecting data. That's a different bailiwick. The major issue with psychological experiments is that they may have significant direct psychological consequences. If you show people only news stories about mass shootings and conflict it may cause them to become violent or suicidal -- which has nothing to do with whether you collect data on it or what you do with it afterwards. The experiment itself is the harm. Which means we would need some kind of principled and efficient way of distinguishing those kinds of "real" experiments from just measuring what happens when you make a subtle adjustment to a context menu.
Mirioron 7y ago
Yes, and it's dumb. It's a bureaucratic nightmare that most likely inhibits progress. Not only that, this is also being used as a cudgel to silence the people that did the grievance studies hoax. [0] [0] https://reason.com/blog/2019/01/07/peter-boghossian-portland-irb-hoax-griev https://reason.com/blog/2019/01/07/peter-boghossian-portland...
- kristianc 7y ago
  Neither does it deliver a guarantee on the results you get out at the end - the Stanford experiment was faked, and academics have struggled to replicate the marshmallow test.
  - Mirioron 7y ago
    While I agree with your point, I do think that some kind of ethics oversight should happen over experiments like the ones you mentioned. I just think that it's absurd to expect the same from simple tests.
  - hopler 7y ago
    I thought marshmallow was replicated but clarified -- that it showed how children react to un/trusted adults, not uncover some genetic propensity to deferred rewards.