SciSpace: a quick story

I must confess that, sometimes, I Google myself. I typically know what I’ll find: I have enough of an online presence that most of the results are related to me (and one other Adrien Foucart who competed in Judo fifteen years ago), and it’s typically a mix of my blogs, scientific papers, social media, and from time to time a post about things I’ve written or done. Finding those is the main reason that I do this exercise every once in a while. Today, however, I got a surprising result, from a website that I had never heard about: SciSpace.

SciSpace is yet another GPT-powered chatbot, aimed at scientists who want to outsource their research to a machine. You ask a question, it answers with a summary built from scientific papers, with citations to those papers so that you can read them if you want to do some work somewhere in the process. I don’t think it’s a good idea: doing those kind of summaries is how you actually gain the understanding of your field, and you’ll necessarily miss a lot of the nuance of what’s happening in the field if you just get the AI-generated “summary”. So even if it worked perfectly as advertised, I wouldn’t recommend using it. But the reason I’m writing this is that it fails pretty spectacularly at its job.

It seems that SciSpace allows you to browse questions, presumably asked by other users. Google indexed a question where, surprisingly, I appeared in the answers. I say surprisingly because the question is not quite in my field: “What are the specific cultural criticisms associated with the implementation of Panopticon in various societies?”

The beginning of the answer seems to be on-topic, although since it’s a topic I know nothing about, it could all be bullshit for all I know. But it’s around the end that I suddenly appear, with this tangent:

Lastly, Adrien Foucart and colleagues critique the Panoptic Quality metric in digital pathology, illustrating the challenges of applying panoptic principles to complex, nuanced fields [10].

This is a reference to my Scientific Reports paper “Panoptic quality should be avoided as a metric for assessing cell nuclei segmentation and classification in digital pathology”, which is probably the paper I like the most out of my publications, but which has absolutely nothing to do with the Panopticon, outside of the fact that there is “Panoptic” in the title.

And this, I think, is where GPT went wrong. Because GPT predicts tokens, which encode parts of words, “Panoptic” and “Panopticon” likely share one or several tokens in common. This alone shouldn’t be enough to trip the model, but it also happens that one of the names most associated with discussion of the Panopticon is Michel Foucault, and “Foucault” also shares at least one token with “Foucart”. This taken together probably sent SciSpace’s GPT into the wrong direction. Because, as we should all know by now, GPT has no fucking idea what it’s talking about. It’s all just a statistically likely string of tokens, with no understanding whatsoever of what’s going on.

So Foucault and the Panopticon end up mixed with Foucart and Panoptic Quality, even though these come from completely different domains. I should also note that for some reason my paper is mistakenly cited as coming from “Dental science reports” instead of “Scientific Reports”. No idea what happened there, but another clear reason that this tool is absolutely useless.

From the CEO’s LinkedIn profile, the platform is “used by more than a million researchers worldwide, including Nobel Laureates”. I highly doubt, however, that Nobel prize worthy science will be done with it anytime soon.