4 ms·

I really wish someone would give an example of an ethically-biased algorithm in common use. So many times I discuss this topic with colleagues and come to reali

I really wish someone would give an example of an ethically-biased algorithm in common use. So many times I discuss this topic with colleagues and come to realize that their complaints are not valid. The algorithm would perform "unbiased" when given different data. To me this indicates a complaint about disparity of the data, but in a more sinister, impossible to reason about way since it is not logically sound (i.e. the counter example with alternate data producing unbiased results exists.)

deadbunny 7y ago
While I can't provide any examples; given we have stories of other biases in algorithms coming out fairly regularly why would ethical bisses be any different?
- zarathustreal 7y ago
  I've never personally seen an account of an algorithm with any ethical bias. I would love to see one in all seriousness, the implications of such an algorithm would be massive. What I have seen is objectively unbiased algorithms (such as neural networks) given input data with objectively true labels (such as racial, gender, etc,) producing objectively unbiased (albeit alarming) results. Now I would love to sit here and argue about equality, equity, and the like but I'll still be waiting for the example of a biased algorithm.
sudosteph 7y ago
Even if there exists a set of data that would produce a more ethical result, would it be fair to say that a potentially biased algorithm cannot be considered ethical until there is a corresponding algorithm that can either filter out biased data or generate unbiased input data? Otherwise it seems a bit pedantic to differentiate between data bias and algorithm bias when both need to be so closely coupled together to produce any meaningful output at all.
- zarathustreal 7y ago
  >would it be fair to say that a potentially biased algorithm cannot be considered ethical until there is a corresponding algorithm that can either filter out biased data or generate unbiased input data Absolutely not, if the sample is large enough to be considered representative how can that data be ethically-biased? Raw data has no ethical implication whatsoever, human emotion must be considered when assessing ethical value. The existence of a second representative sample which produces a different ethical implication is proof that the algorithm would not be ethically biased. >it seems a bit pedantic to differentiate between data bias and algorithm bias when both need to be so closely coupled together to produce any meaningful output at all On the contrary, it seems disingenuous to argue that an algorithm, a set of computational steps, is ethically-biased when a counter example (in the form of a data set) exists. This is akin to blaming the process of eyesight for an observation of a ethically-repugnant phenomenon. It is not the process of seeing that causes observation of ethically negative phenomena, one can just as easily observe ethically positive phenomena through the process of eyesight. We do not need to have invented glasses that filter out ethically negative phenomena to prove that the use of ones eyes is a ethically-unbiased process, the fact that it's possible to observe both ethically positive and ethically negative phenomena is enough proof. Bias could be introduced by seeking negative phenomena and avoiding positive phenomena, but that doesn't imply anything about eyesight itself - it remains an unbiased process. If we are going to come to any meaningful conclusion about how to change the results of unbiased algorithms (insert photons into pupil, decode electric signals into conceptual models, etc,) we need to stay as logically consistent with our conclusions as possible. Observations should inform conclusions, not the other way around.

deadbunny 7y ago

While I can't provide any examples; given we have stories of other biases in algorithms coming out fairly regularly why would ethical bisses be any different?

zarathustreal 7y ago
I've never personally seen an account of an algorithm with any ethical bias. I would love to see one in all seriousness, the implications of such an algorithm would be massive. What I have seen is objectively unbiased algorithms (such as neural networks) given input data with objectively true labels (such as racial, gender, etc,) producing objectively unbiased (albeit alarming) results. Now I would love to sit here and argue about equality, equity, and the like but I'll still be waiting for the example of a biased algorithm.

sudosteph 7y ago

Even if there exists a set of data that would produce a more ethical result, would it be fair to say that a potentially biased algorithm cannot be considered ethical until there is a corresponding algorithm that can either filter out biased data or generate unbiased input data? Otherwise it seems a bit pedantic to differentiate between data bias and algorithm bias when both need to be so closely coupled together to produce any meaningful output at all.

zarathustreal 7y ago
>would it be fair to say that a potentially biased algorithm cannot be considered ethical until there is a corresponding algorithm that can either filter out biased data or generate unbiased input data Absolutely not, if the sample is large enough to be considered representative how can that data be ethically-biased? Raw data has no ethical implication whatsoever, human emotion must be considered when assessing ethical value. The existence of a second representative sample which produces a different ethical implication is proof that the algorithm would not be ethically biased. >it seems a bit pedantic to differentiate between data bias and algorithm bias when both need to be so closely coupled together to produce any meaningful output at all On the contrary, it seems disingenuous to argue that an algorithm, a set of computational steps, is ethically-biased when a counter example (in the form of a data set) exists. This is akin to blaming the process of eyesight for an observation of a ethically-repugnant phenomenon. It is not the process of seeing that causes observation of ethically negative phenomena, one can just as easily observe ethically positive phenomena through the process of eyesight. We do not need to have invented glasses that filter out ethically negative phenomena to prove that the use of ones eyes is a ethically-unbiased process, the fact that it's possible to observe both ethically positive and ethically negative phenomena is enough proof. Bias could be introduced by seeking negative phenomena and avoiding positive phenomena, but that doesn't imply anything about eyesight itself - it remains an unbiased process. If we are going to come to any meaningful conclusion about how to change the results of unbiased algorithms (insert photons into pupil, decode electric signals into conceptual models, etc,) we need to stay as logically consistent with our conclusions as possible. Observations should inform conclusions, not the other way around.