I'm confused about the difference between "ethics" and "bias" when those concepts are discussed in the context of Machine Learning (ML). In my understanding, ethical issue in ML is pretty much exactly the same thing as "bias": say, the model discriminates people of color and this is the same as to say that the model is biased. In short, "ethics is always a bias, but it is not necessarily true that a bias is always an ethical issue". Is this true?
The term bias is, to my knowledge, not related to ethics in the context of ML. Instead, it usually refers either to the bias–variance tradeoff or to a learnable parameter of a model, e.g. bias in a neural network. (Note that in statistics the term is commonly used to refer to biased estimators which is related to but more general than its use with regards to the bias-variance-tradeoff.)
In contrast, when making a connection to ethics (aka fairness) you most likely use the term in a more general way or how it commonly used in science. (But it is important to note that this is not what bias refers to in ML.)
However, even when applying the general scientific definition of bias its relation to ethics/fairness in ML is limited:
Let's assume you apply a model to classify images of chairs and tables. If your dataset contains 99.999% chairs the naïve classifier which always predicts chairs would perform very well in terms of accuracy. (side note: let's ignore the fact that "very good" is task-specific and accuracy might not be the best metric here) This model would be very biased towards chairs in the general meaning of the term. However, we would not consider this an issue of ethics or fairness (unless you're a big fan of tables).
Now let's assume that you have a model applied in a self-driving car task. One could think of a situation where the model needs to decide to either run over a group of two people or a group of three people. The decision the model has to be made is an ethical ML problem. (probably it's actually an AI issue, i.e. involves other sub-fields of AI too since self-driving cars usually apply techniques from multiple AI-fields and not just ML) Clearly, this is an ethical problem independent of any bias, i.e. you can have ethical considerations in ML without necessarily involving bias.
As the above two examples show you can have ethical issues in ML without bias (in the general sense) and vice versa.
I'd say the reverse is true: "bias is always an ethical issue, but it is not necessarily true that an ethical issue is always a bias".
"bias is always an ethical issue" because like you said, a model may discriminate against people of color when it comes to hiring, for example. Even if the model is correct (this is very unlikely and if it happened it is caused by confounding variables obviously), it is unethical to decide not to hire someone because of their skin color. More generally, we should not leave out a group of people entirely.
"an ethical issue may not be a bias" because there are many ethical issues related to Machine Learning that do not involve bias. For example:
- Government Surveillance: the government uses cameras incorporated with facial recognition to track activities and location of certain people of interest.
- Deep fake videos: a video of a person in which their face or body has been digitally altered so that they appear to be someone else, typically used maliciously or to spread false information. (example: https://youtu.be/bE1KWpoX9Hk)
There is a confusion between "bias" in the mathematical/statistical sense (that is the difference between 2 values) and "bias" in the psychological/social sense which means prejudice. In fact the two can be related in all sorts of ways, not in a simple one-to-one sense.
For example a biased model (in the mathematical sense) may counteract the effect of social prejudice (eg because the data are imbalanced due to social bias, so another bias can equalize this), while another unbiased model may support social bias/prejudice effects. The other way is also possible. There is no simple one-to-one mapping.
This researcher focuses on Ethics, transparency and accountability of AI. It can be a start for a more rigorous approach to discrimination in machine learning.
See also: What is Inductive bias?
In statistics bias is often used to refer to a statistics inability to measure the true population value of the statistic. This is usually shown analytically by computing the expected value of the statistic. If it is equal to the true population parameter, it is 'unbiased'. This is sometimes hard to do, so you can sometimes use simulation to look at the resultants distribution to see if it looks biased or not. Small biases will usually not matter.
This also translates to machine learning. If your sampling method is bad, your statistics (means, correlations, models) will always be bad because your sample may not reflect the population. If you choose different models, most likely it will favor certain groups over others. Sometimes this is by optimization, but often it can just occur randomly. If you omit an important variable from a model, it may be biased towards one group, and if you then include the important group it may eliminate the bias to the group, or it may increase the bias toward the other group. You never really know how bias will work. You just need to be aware if it. Bias in statistics is not inherently right or wrong. But ethics is a different matter. That usually includes a statement of fairness and equity.