Has the AI ethics debate forgotten human flourishing?
It seems that everyone these days is doing or at least talking about AI ethics or one of the closely related areas of transparency, fairness or interpretability (as an aside: although the field of machine learning model interpretability has been around for a while it has experienced a resurgence of interest due to the introduction of the General Data Protection Regulation (GDPR) in EU and the right to explanation, which it is said to confer to data subjects). The examples brought up in these conversations center around the usage of machine learning algorithms in fields where the application of biased technology opens up all kinds of avenues for abuse: sentencing in courts, hiring job applicants, diagnosing illnesses and so forth. After all, these seem to be the areas where the blind trust in anything “artificially intelligent” seems to be unethical and perhaps even downright reckless.
Although conversations about AI ethics in these contexts is absolutely paramount, we seem to have forgotten another avenue where the application of ethics could lead us to machine learning technology that is designed to complement the human mind, not replace it. The Introduction to Data Ethics course from the Markkula Center for Applied Ethics at Santa Clara University defines ethics as the human quest to figure out how to live best, how to flourish. Keeping this definition in mind, it would seem pertinent that the discussion around AI ethics include not only ways to prevent harm from the misapplication of machine learning algorithms, but also considerations on how these systems could be used to promote human endeavours.
A lot of my professional work time these days is spent looking and analysing systems built with machine learning algorithms. These systems ingest large quantities of data and attempt to learn models or decision boundaries, which can then be leveraged for their predictive power to assign class labels or values to data points that model has never seen before. A classic example here is the rather innocuous Iris dataset. First collected by Robert Fisher, this dataset has been used by probably nearly all machine learning students and practitioners to generate models that classify flowers into one of the three Iris species based on measurements of the petal and sepal.
Many of the machine learning models that we develop in the tech industry are prescriptive: they spit out a decision for a new data point without providing much in the way of an explanation. Even obtaining the source code that trained the algorithm may not offer much to explain why a particular data point was assigned a particular class: most source code systems are thousand of lines long, can take months to fully grok even for a seasoned programmer and to even begin understanding the final model output from the algorithm would require the data that was used to train it. Because of their black-box nature, relying on these algorithms for decision makind requires an ounce of blind trust and unfortunately, does not usually end up teaching the human operator anything meaningful about the dataset being analysed. Blind trust is of course a less than desirable characteristic in fields where decisions (such as whether or not someone has an illness) can have life-altering impacts. Machine learning algorithms are, at the end of the day, expressed through a specialised type of software and like all software, their implementations too have bugs.
As part of the discourse on AI ethics, I’d thus like to focus on machine learning systems that don’t just focus on optimising some narrow performance metrics, but instead focus on finding ways to promote human flourishing in complementary ways to the human brain. Making sure that humans are not exposed to biased and opaque systems that are prone to error is one way. Another one is to focus on developing machine learning systems that, instead of focusing on just outputting a decision, focus on a collaborative approach. For example, it is well known that seeing patterns in multiple dimensions is quite an impossible task for humans, but not for ML algorithms. Once an algorithm has determined these patterns it can send feedback to the human user and await for further questions or instructions for analyses. Such systems are already being developed for example in the field of interpreting medical images (https://arxiv.org/pdf/1902.02960.pdf).
I’d like to conclude this barrage of thoughts by a machine learning paper that was first brought to my attention by Google Software Engineer and fellow Bryn Mawr alumna, Julia Ferraioli, “Machine Learning That Matters” by Kiri Wagstaff from the NASA Jet Propulsion Laboratory. Referring back to the Iris dataset mentioned before, Wagstaff asks “Legions of researchers have chased after the best iris or mushroom classifier. Yet this flurry of effort does not seem to have had any impact on the fields of botany or mycology. Do scientists in these disciplines even need such a classifier? Do they publish about this subject in their journals?”
There is a lot more to unpack in Wagstaff’s excellent article (including some areas where machine learning would be predicted to have an impact that turned out, in the years following the publication of the article, to go into the direction of sinister rather than impactful) and I am definitely going to return to it in later writing, but for now I am going to leave the discussion at this: what would machine learning systems that complement human abilities and facilitate human flourishing look like?