Philosophy as conceptual engineering and responsibility
Before pivoting my career to data science and programming, while I was still studying philosophy, I encountered a persuasive idea called conceptual engineering. Conceptual engineering views philosophy as an enterprise that creates, edits, criticizes and develops concepts that are meant to be used for something specific.
A contemporary example, which also happens to relate to my current field, is that of responsibility and machine learning models. As humanity creates and implements ever more ML software artifacts that make predictions and decisions, responsibility for the real-world outcomes of these systems becomes much more diluted. The already classic example is that of a traffic accident where a self-driving car has been involved. The simple question is: who or what should be responsible for the accident? Who should we prosecute? Who should suffer consequences?
Now if we were to understand the work of philosophers as investigators who study the necessary and sufficient conditions for some concept to apply, we would view the issue of responsibility and the work related to it as a process of discovery. To answer the question of how responsibility should be allocated with self-driving car accidents, we need brave explorers willing to dive into the platonic world of abstractions and find the answers from there.
This view of philosophy is not entirely unfounded. Mathematics can be viewed as a process of discovery and the empirical sciences indeed do make discoveries – they certainly don’t engineer their results. So why not philosophy as well? If we were to ask what a just world is, can we not just investigate the concept of justice and through that investigation discover real answers to the problem?
I hold a very different view. I think that automated decision-making systems – abbreviated as ADMS by the cool kids – require us to develop a new concept of responsibility. The task is not to discover, but to invent and iteratively refine and improve concepts.
However, when we formulate the work of philosophy as that of engineering concepts, we need to define what constitutes good work. Here, I reveal my unashamed identity as a pragmatist. I believe that the concepts we cook up should serve us in practical terms and the merits of a concept should be measured by its positive real-world impact. I also think that the developed concepts should be maximally simple, be readily understandable for all stakeholders, provide guidance for decision-making and be appealing enough to gain popularity.
To apply this view to the issue of self-driving cars, we need our concept of responsibility first and foremost to impact the number of future accidents. For example, if we hold the car itself responsible and use a software update as a form of corrective punishment, then we are likely to fall short of our aims of making automatic traffic maximally safe. If we can trace the cause of the accident to a faulty software problem which was a result of haphazard programming, should we not also punish the programmer to discourage careless software delivery?
The view of philosophy as discovery can be supremely tempting. When considering responsibility, it is hard not to think about who really is responsible for the accidents instead of the most efficient targets of responsibility for making automated traffic safer. I personally feel the temptation to use counterfactual analysis to discover whose actions were most relevant to the accident. If a programmer had not been careless, if a manager had not been pushing for progress, if the investors had not been demanding a higher return on their investment, maybe the accident wouldn’t have happened. The problem is of course that all three can be and often are true, so the counterfactual approach just ends up assigning responsibility to many people. Do we send them all to prison? Do we give them a collective fine? Do we give up and say no one is really responsible?
The problem with the counterfactual approach is the assumption that someone is indeed responsible for the accident. But as I pointed out, this is the mistake that conceptual engineering aims to correct. I think we need a collective engineering effort to create new ideas of responsibility with the explicit aim that the concepts should allow us to strategically inflict punishment where it is most fruitful so that the ADMSs we implement become safer. This engineering effort can also lead to assigning punishment before any accident has happened – for example, I advocate for a whistleblower system, where management and investor pressure to hasten the delivery of these safety-critical systems can be reported and sanctioned properly.
The reason I feel more philosophers should start thinking of themselves as engineers is that I would like to see the field take a more pragmatic and engaged approach with the world we find ourselves in right now. Sure, feel free to study Aristotle and Plato, but ask yourself how you can use that study to create useful and informative concepts that solve real-world problems.
Grammar mistakes fixed with help from Claude Opus 5.5.