Global Health prompts to an Artificial Intelligence language model
We have interacted with Bloom, an open source AI tool trained on a gigantic corpus of text, in search of insights into Global Health and our role as a non-profit organization
Samuel Volk & Miquel Duran-Frigola
Fundraising for a charitable project requires tranquility, self-awareness and diligence. When you are trying to find your voice in Global Health while still struggling to understand your own principles, the available funding channels can be distracting, if not incompatible, with your ethical stance. You need to be very strict about the funding you pursue. You start with a concession, then comes another one, and before you know it nothing matters anymore. This became apparent to us during the past summer. For some reason, Ersilia gained traction amongst the “for-good” branches of large corporations. We also met more frequently with philanthropists and family funds, perhaps over-strategizing and putting ourselves at risk of becoming a type of organization that we simply are not. All of this triggered internal discussion, exchange of emails and reading sessions on the topic of Global Health. The goal was to come up with a clear, robust protocol to decide which money (or let’s call it “opportunity”) to accept, and which to kindly decline.
We failed to create this protocol. We did not even get close to a draft, actually, because as soon as words of the kind “colonialism” and “capitalism” entered the discussion (they do often), the spirals began and all became absurd. Words like these have this mysterious property: you put them in a sentence and the sentence is suddenly ruined, devoid of meaning. How are we going to question, let alone “kindly decline”, opportunities that stem from colonialism if we don’t know how to use this word? Finding good advice from colleagues, especially good disruptive advice, is difficult since we are all bought into the system in one way or another. Said more mildly, we all belong to the same episteme and it is hard to escape from it.
In July, an Artificial Intelligence (AI) model called Bloom was released. Bloom stands for ‘BigScience Large Open-science Open-access Multilingual Language Model’. It may also refer to the Leopold Bloom of Joyce’s Ulysses, or to a blooming flower, which would be equally evocative. Bloom was trained on an enormous slice of the internet containing 46 natural languages in 1.6TB of text. You can try it here for free. Other large language models exist. This one is remarkable because it is not an industrial product but the result of a fully open, collaborative effort between scholars.
The fascinating fact is that these large language models can be thought of as informal knowledge systems. Our culture is primarily stored in language, and Bloom is an intelligence that has only ever experienced language. As powerful as they are controversial, large language models have received a great amount of attention recently for their ability to produce increasingly realistic text responses to a prompt, for the inordinate energy cost of their training, for their pernicious embedded biases, and for a host of additional ethical concerns surrounding their use and development. Whether we like it or not, Bloom is a representative of the knowledge system and cultural assumptions of the resourced world, which is the world of the creators of the text content that constitutes Bloom’s training data. Bloom may be a good way to navigate, and maybe even demystify, our current episteme.
The most common usage for Bloom is the autocompletion mode. You write some language in a prompt console, Bloom reads and interprets it, and then it helps you write the rest. For example, when we typed “Ersilia is an invisible city, and it is a nomad city”, Bloom added: “because only its core is real”. Only its core is real — that was interesting. Then Bloom continued: “Ersilia is a city of copies, and it is a city of research”. A city of copies… that was interesting, too, and to-the-point, and not very assertive, perhaps, but true after all. This is what the Ersilia Model Hub is about: a collection of research tools developed by others and publicized by us.
The next inevitable prompt to Bloom was to gather a definition for Global South, because, admittedly, even if we work in and for the Global South, it is not entirely clear to us what the term means. To the prompt “Global South can be defined as” Bloom responded “the countries of the world that are not part of the Global North.” Then it went on to enumerate some of these countries in the Global North, which was not what we asked about. Only after a few lines it went back to the Global South, that is “the developing countries of the world, which are the rest of the world”. That was a terrible definition, by all means. A 90º rotation of the orientalist idea that there is an “us” and there is a “them”, that the East is everything that is not the West, and that the East needs to be invented for the West to make sense and exist. Bloom wrote “the Global South is the part of the world that is exploited by the Global North”, and then it stopped.
We may not have learned anything new here, but then again, Bloom is not a reasoning intelligence. It is more of an average language generator, an average discourse generator, if you will. Asked about the difference between Global North and Global South, Bloom clarified that “the Global South is not a place, but a condition”. A condition… is this how we all think about the Global South, on average? “A condition of being poor”, Bloom clarified in a strange rhetorical shift, “a condition of being oppressed, a condition of being exploited, a condition of being colonized, a condition of being marginalized, a condition of being excluded, a condition of being dispossessed, a condition of being discriminated against, a condition of being underdeveloped, a condition of being underprivileged, a condition of being underpaid…” The “unders” went on and on, and the repetitions began, in an infinite enumeration that wasn’t going anywhere.
To be completely honest, the prompting session with Bloom started to become frustrating, with so much emphasis on the North-South dichotomy. “What is the similarity between the Global North and the Global South?”, we asked. “The similarity is that they are both dominated by the same economic system, which is capitalism”, but then, for some reason, Bloom was blocked and fell back on the differences and the infinite enumeration. “The difference is that the Global South is the place where the Global North dumps its waste”, and “exploits its people”, and “exploits its women”, and “exploits its men”, and “exploits its elders”, and “exploits its culture”, and “exploits its history”, and “exploits its future”, et cetera, et cetera.
All of this, and much more, is true but incomplete — we might expect these kinds of messages only from a language model that is mostly trained on text written in the resourced world. Asked to define “exotism”, Bloom responded “cultural imperialism”, asked to define “otherness”, it responded “cultural racism”. Asked to define “progress”, it responded “an increase of the number of people who are able to live a life of dignity” and then it proactively elaborated on dignity (freedom, not money; people, not things; and so on). Asked about “neglected tropical diseases”, Bloom was able to itemize each and every one of them. Chagas, dengue, sleeping sickness, the whole list. So the tool is impressive, but at the same time uninteresting, the same way most of us are uninteresting when we talk about the world.
It was time to ask Bloom the question that overflies Ersilia’s day-to-day. “Will AI solve inequalities in Global Health?” It took some extra prompting to get an informative answer to this question, but we finally found a worthy bit: “there is nothing like absolute equality; all resources are subjective.” This response speaks to one of our main preoccupations: we can’t control whether the resources we are creating will ultimately be valued. Bloom then wrote a complex sentence, grammatically: “If you charge for your resource, then it needs to be something that could not, in any practical sense, be done without paying you for it”.
This is a strange concept of value, one that certainly did not clarify our fundraising dilemmas. It is like offering good conversation over a cup of coffee but refusing to talk to anyone without them paying first. Ersilia cannot be accessed through payment (we are and will always be free and open source), but this doesn’t make us valuable, necessarily. Aren’t there values that are not so transactional? Bloom doesn’t equivocate: “when we drink coffee with someone, we are making a transaction. In this case, we are trading the cost of the coffee with the social exchange of the information. If the coffee costs more than the value of the information to me, I might not want to engage in the exchange. Otherwise, I would consider the exchange worth the cost of the transaction to me.” At this point, we’ve fallen back onto the same concept several times: an essential resource is essential only insofar as we hold it, a value is only a value relative to us. It is somewhat futile to think in ownership terms if you pursue open science, especially if most of what you do is just recycle work from others and make it available in the public domain.
After a lot of prompting, it was clear that Bloom was not catching our concept of “value”. We didn’t derive much new insight into the AI side of things, either. It was not clear, at this point, whether AI will solve inequalities in Global Health. “At least”, we told Bloom, “help us understand the ethics of AI.” Are we causing any harm by using AI in Global Health research? “The potential uses of AI in healthcare highlight issues of bioethical exclusion”, Bloom argued. “Differential statuses are produced and reinforced through various mechanisms. The advent of AI may intersect with existing hierarchies that differentially privilege individuals and institutions based on their access to and control of technology and biomedicine.” We agree with much of what Bloom said here. As the promises of AI start to yield for biomedicine, issues of access to these developments are enormous, and access is indeed determined by pre-existing hierarchies. Nonetheless, it’s hard to shake the feeling that Bloom just dropped us into an uncanny valley of ethical thought. After the second and third reads, Bloom’s comments feel more like an offering of the internet’s discursive tokens rather than a sharp critique. But, don’t we all do the same? We exchange the same comments over and over, over and over, until they’ve become detached from the reality they once described.
Resolving our lingering questions about Global Health with Bloom proved difficult, not because Bloom could only parrot commonplace ideas, but because Bloom’s shortcomings are too similar to our own. Trained on the data we already have, the words we’ve already written, AI may never innovate for us better than it shows us a mirror. But mirrors are still interesting, especially when you are looking at a reality that is difficult to look at. Global Health is a difficult reality to look at, with questionable funding and power dynamics. Bloom didn’t provide us with many answers, but it gave the necessary language to fuel a conversation that, in the end, we need to have with ourselves.
