Why We Need Ethics for AI: Aethics
We need to think specifically about the implications of classification, machine learning and artificial intelligence on decision making processes.
In 2017 Kate Crawford presented at NIPS, the Neural Information Processing Systems Conference. She convincingly showed that analytical models trained on existing data exhibit undesired behaviours, like discriminatory behaviour.
As an example she mentions the famous ProBublica news article. The article examined police algorithms identifying blacks as being more likely to (re)commit a crime. In the article ProBublica gives examples of inequitable scores. The black woman Brisha Borden was given a higher recidivism score than the white Vernon Prater —who was already a recidivist and who would go on to be convicted of new crimes.
Many more examples of undesired behaviours in analytical models are available; the famous ‘Google-Gorilla’ case is another one, or women being denied loans after breast cancer.
In her talk Kate Crawford dives deep into the notion of bias. What do we mean when we say that a model or algorithm is biased? She differentiates between allocative and representative harm and does a masterful job showing the nuances in biases in machine learning — if you haven’t seen the talk, I suggest you do so.
She finishes her talk remarking that bias might be a side effect of classification in general. Classification is a dimensional complexity reduction technique and these undesired behaviours are an artefact of the classification. And that we are currently in the largest experiment of classification in the history of mankind.
But what do we actually mean when we say models ‘got it wrong’ or that the models exhibits ‘undesired behaviours.’ What if a certain subpopulation is indeed over or underrepresented in the data. Should we go to lengths to adjust our predictions in the algorithms to suit some sort of equality? This not a technical question, it is an ethical normative question.
It would be too easy to simply ascribe all model failures to either the underlying data, or a failure of the operators tune the model ‘correctly.’ The question goes deeper — what do we mean when we talk about undesired behaviours, and what does discrimination by algorithm mean?
In order to answer these questions we need to have knowledge about right or wrong, i.e. ethical, and that we have specific notions on what to do, i.e. normative.
Consider for instance that the underlying data correctly establishes that some groups are associated with a higher risk. The case in The Netherlands where women who had previously been diagnosed with breast cancer got higher premiums, or were denied mortgages and loans. We intuitively feel this is wrong, even if the underlying data is (probably) correct in assuming that they are a higher risk to the loan issuer. It would be a mistake to blame the AI however— which is simply learning from all the historic cases. Note that, once the ‘problems’ of undesired behaviour are identified, technically suppressing certain model behaviour is simple.
The normative questions is harder: should we make the loan issuer carry the risk of giving out mortgage to these women? If your reaction is ‘yes,’ then I’m with you, but ask yourself what the consequences of such thinking are? Who gets to determine who carries the burden? The women? The loan issuer? Or governments? People are denied certain services all the time — no doubt some reasons being dubious from an ethical perspective.
The ‘discrimination’ in this case is an ethical normative boundary drawn by a specific group at a specific time. What we need to think about is the ethics and implications of large scale adoption of artificial intelligence in the form of aethics.
The same issues comes up in businesses where data scientist are asked to create a model that predict some metric: sales, churn, markup. But depending on your goal, many models and model-variations are possible.
It is much easier to hide behind crime statistics than to engage in a debate on what should be done. Analytical models simply give the ‘as-is’ and ‘to-be’ states with no ethical correction in between. Data Science does not show us ‘what is good’ or ‘what should be,’ it does not help us with what is right or what we should do.
Even worse, we know that science — and the scientific approach — have severe flaws. Considering all issues in modelling merely technical issues would be a mistake. Flaws in the structure of our scientific inquiry might result in wrongfully published papers; that is, papers that have claimed to have found the answer to a problem which actually turned out bogus.
Consider for a moment bias in research. The bias is towards solving problems for which scientist believes there is a true positive relationship, it means that less research is done towards those where we think no such relationship exists.
Bear with me: assume an arbitrary statistical power of 80%. In a group of 1000 experiments in which, let’s say, 100 have a true positive relationship, we should correctly identify 80 of the 100 as having a true positive relationship.
However, we will also identify 20% of the remaining 900 as false positives; totalling 180. If all positive findings are published we have a ratio of 80 (true positives) to 180 (false positives). Only 30% of the published research is actually ‘true.’
All of this should put the scientists at the edge of their seat when using technical findings to base technical decisions on. The cure, as Sam Harris likes to say, for bad science isn’t more science. We cannot hide behind ‘more science’ to figure out how to deal with ethical questions in AI.
The only way we can overcome these technical issues is to not us the same kind of thinking that we used when we created them. We need to think specifically about the ethics related to the application of AI in the public domain where we apply AI to make automated decisions. We need to have non-technical thinking to overcome this technical issue.
Therefore we need aethics—ethics for AI.
