4 min and 16 sec to read, 1066 words

One of the areas I want to explore more in the coming weeks is automated science over all — the idea that we could automate science, the entire process and perhaps the institution¹ This is one of the interesting questions — what do we mean when we speak about automated science? Do we mean automating the process of science or the institution of science? Are these really different, and if so how? If we believe that it is the institution, then one question quickly becomes the conditions for and ability of us as citizens, individuals, humans to participate in science as a practice. Process, practice, institution — science can be described as all of these, and when discussing the automation of science we need to figure out which it is we are automating – or if it is all of them. is both exciting and somewhat worrying.

These are some preliminary thoughts – written down mostly to map how I view the field before digging in more deeply into the literature and thinking that has already happened.² I find it helpful to write out your own questions and ideas before digging into the literature, to have a naive mapping of the subject and how it seems shaped to me at this point in my work. This means that anyone reading this might be underwhelmed with ideas I find interesting – or frustrated that I am going over old ground, but methodologically this helps me enough that I am willing to risk that even in sharing these sketches.

1. Let’s imagine a sciencomat. It is a machine that produces hypotheses, tests them, orders them into theories and explores the world all on its own. How would we design the user interface for such a machine? What would it need to be able to do for us to rely on it and trust its findings? Say we design it with a voice interface, much as a generative AI model today: we can ask it questions and get answers. Now imagine that this automat stands in a lab, and when we come to the lab in the morning it has a small red lamp that flashes if it has made a scientific discovery – and when it has discovered something we can then ask it what it has discovered. What kind of discoveries could such a machine make? What is the problem space it could operate in?
2. Say we come in one morning and the machine blinks red. We ask it what it has discovered and it says « I have discovered a new fundamental force in the universe – and I call it xalitation. It is what produces gravitation in low temperature settings such as ours.» — how would we then proceed to validate this finding? And what if it answered « I have discovered a new drug that cures 30 different forms of cancer? » or if it said « I have a new theory about the fall about the Roman Empire»?
3. One question that comes up quickly when we look at automated science is if it could ever really create the kind of paradigm shift that we saw from Newtonian to Einsteinian physics. And if it did if we would recognize it and find it legitimate.
4. You could imagine a machine that finds different interesting mechanisms and functions in reality and shows how they can be exploited to reach fascinating results – but does nothing to explain those functions or mechanisms. A machine that produces engineering without science — and isn’t that what evolution is?
5. Could evolution be automated, or would we be stuck in the conceptual frameworks we need to create the machine with which we would hope automate it?
6. There is something here about observation. A machine could be made to observe according to different categories and templates – and then report on those observations. Automated observation seems to be the least controversial of all the steps in the scientific process. We could have a radio-telescope listen instead of us for structured signals that would indicate that there could be extraterrestrial intelligences — and if it recognized such a signal it could alert us. Here the machine is essentially engaged in a search in a pre-defined space of possible signals — something that could be easily automated.
7. What about hypothesis-generation? A machine could, trivially, produce many random hypotheses. It is easy enough to imagine a machine that just randomly fills ut a pattern like « substance X causes disease X » – and then proceeds to test the resulting hypothesis space. In principle you would then only be dependent on the data available. In testing « liquorice causes lung cancer » we are left powerless if there are no data sets that document patients in the two dimensions we are interested in, liquorice consumption and the prevalence of lung cancer. 
8. This suggests another kind of automated observation: large scale data collection on all kinds of possible variables in a hypothesis space: if we want to isolate the causes of cancer we should observe all the possible antecedents — but that isn’t possible. Not least because we could look at combinations of different substances and environments.
9. Causality, then, quickly becomes a problem for any automated science. This is not surprising – since we know already that it is challenging for machine learning overall.³ As evidenced by the work Judea Pearl and others in e.g. Pearl, J., and Mackenzie, D. (2018). The book of why: the new science of cause and effect. NY, United States: Basic Books. But why do we believe causality is so important? Knowing that X causes Y makes it easier for us to understand the world – but what if we were offered a non-causal piece of knowledge that could cure a disease? Is this possible? There could be an ethics of causality somewhere here.

There is much more to come back to in all of this. I am interested in also combining this first sweep with thinking about Stanislaw Lem’s ideas in Summa Technologiae on automated science – not least the necessity he sees in automation. There seems to be a set of interesting ethical questions here that could be worth exploring: is there an ethical imperative somewhere? Is it ethically defensible not to automate science? Assuming that automation is the only way to deal with the complexity or « megabyte bomb » that Lem speaks?

Much more to dig into here.

Footnotes and references

• 1
  This is one of the interesting questions — what do we mean when we speak about automated science? Do we mean automating the process of science or the institution of science? Are these really different, and if so how? If we believe that it is the institution, then one question quickly becomes the conditions for and ability of us as citizens, individuals, humans to participate in science as a practice. Process, practice, institution — science can be described as all of these, and when discussing the automation of science we need to figure out which it is we are automating – or if it is all of them.
• 2
  I find it helpful to write out your own questions and ideas before digging into the literature, to have a naive mapping of the subject and how it seems shaped to me at this point in my work. This means that anyone reading this might be underwhelmed with ideas I find interesting – or frustrated that I am going over old ground, but methodologically this helps me enough that I am willing to risk that even in sharing these sketches.
• 3
  As evidenced by the work Judea Pearl and others in e.g. Pearl, J., and Mackenzie, D. (2018). The book of why: the new science of cause and effect. NY, United States: Basic Books.