“Part of explaining what an AI system is doing and knowing why it is doing it turns on knowing what (if anything) the system is representing when reaching an output.”
That’s a small excerpt from the entry on “representation” at the new Philosophical Glossary of AI, a growing collection of explanations of key concepts often used in discussions of artificial intelligence (AI), edited by Alex Grzankowski (Institute of Philosophy and Birkbeck College, University of London).
What topics would you like to see covered in it?
The entries, each several paragraphs longs, aim to elucidate both the significance of the concept in regards to AI as well as the philosophical dimension of the concept (disputes over its meaning, implications of different conceptions of the concept, questions related to it, etc.).
The glossary aims to be a resource for the various people working on and thinking about AI, providing a common understanding of key terms (and the issues surrounding them) for AI researchers, lawyers, policymakers, businesspeople, philosophers, journalists, and the public.
Entries are written by academics and practitioners and reviewed by the glossary’s editorial team, and sometimes by outside experts. The entries are also “tested on different audience groups such as policymakers, lawyers, and journalists at stakeholder events.”
Dr. Grzankowski says he hopes readers will provide feedback on the glossary and suggest additional possible subjects for it. He adds that the team is also looking for good candidates to write new entries. You can comment with your suggestions here or get in touch with him directly by email.
The glossary is an initiative of The London AI and Humanity Project.
It’s pretty worrying that searches for ‘attention’, ‘transformer’, ‘convolutional’, ‘diffusion’, ‘recurrent’, and so on — the most broad-bucket categorizations of machine-learning architectures — don’t turn up literally anything. I hope the editors plan to remedy this soon. Philosophical thinking about AI (so-called) has been limited by a lack of awareness and willingness to engage with how various kinds of systems actually work under the hood.
Hi Turing,
I’m the managing editor of the Glossary.
Thanks for the feedback! You’re right, those terms are missing. We do plan to expand the glossary with entries on how AI systems work, including the terms you mentioned.
Finding authors with expertise in both AI and philosophy is key, and we’re working on that. If you or anyone you know might be interested in writing an entry, please reach out!
Benjamin Henke
I’m not a philosopher( professionally ) but it seems to me that this subject is of such immense importance that it would be better done by a fairly prestigious academic institution. The Stanford Encyclopedia of Philosophy (SEP) is an excellent source for people like me with little formal education but nonetheless a strong interest in philosophy. I don’t know what stature the SEP has among academic philosophers but it must be used at the very least by graduate students in philosophy. The reputation of an institution is on the line when they put their name on a project. Wouldn’t that also encourage experts to write articles for that effort. The impact of AI is gargantuan (no mention of pantagrual(sp?). There is so much talk by people who are clearly clueless. It seems like it would be a tremendous community service.
I like the idea of having both search systems. We can step out of this either/or , zero sum patterns of thinking
I would like to see terms like interpretability, clamping, and alignment. I would also like to see the inclusion of terms from the broader conversation, such as instrumental convergence, the orthogonality thesis, and recursive self-improvement. It’d be nice to have some philosophical clarity on those rather controversial ideas
Thank you! We’re interested in including these concepts as well (and will soon post entries on a few of them). Stay tuned!
If you have expertise in these areas and would like to contribute, please feel free to reach out!
This is a great idea, and it’s very sleek and user-friendly. However, I have two concerns about its quality as a resource.
First, I checked out the entry on creativity, the topic on which I have the most expertise (entry here: https://www.aiglossary.co.uk/index/ai-creativity). It begins with the pronouncement that “creativity is the disposition to produce outputs that are original and valuable.” This is certainly not a consensus definition in the philosophical literature. There is a lot of disagreement among philosophers working on creativity about what its necessary and sufficient conditions are. I’d encourage anyone using this glossary to take it with a grain of salt as an accurate representation of how philosophers understand these concepts.
But perhaps more concerning is the fact that these entries do not cite any philosophical sources or direct readers to academic articles, despite clearly drawing upon ideas in the literature. What justifies the lack of references?
Hi Lindsay,
I’m the managing editor of the AI Glossary. Thanks for your feedback!
Regarding the first issue, I think that’s right. It’s important to point out that the AI Glossary doesn’t aspire to articulate either the consensus view about a philosophical topic nor speak to significant controversies in the literature. Of course, we do want the claims to be relatively neutral, but often you can’t offer much clarity about the relative topics while remaining neutral. The aim of the AI Glossary is to provide clarity to debates in AI from a philosophical perspective, not to offer overviews of the philosophical literatures on the relevant terms (and there are much better sources for this kind of thing). A related constraint is that entries are presently limited to three pages of a Google doc. We think that’s a good call, given our interest in having the entries be usable by the general public, but I’ll admit it’s often frustrating work for us philosophers.
Regarding the second issue, this was a tough one as well. We don’t consider the entries scholarly contributions of that kind. They’re more akin to brief newspaper articles (or even opinion pieces), which more rarely cite sources. We considered whether each entry should include a ‘relevant readings’ list, but concluded that the literature on these topics moves so quickly that this list would quickly be useless, and so would not serve the central goals.
That said, it’s early days, and we don’t know of a clear model on which to base these decisions since this particular task is somewhat unique. (If anyone has clear analogues in other domains, we’d be very interested in hearing about them as well.)
Thanks again,
Ben
Thanks for your response, Ben!
I think, given what you’ve said about the goal here, it might be worth rethinking what is written about the process and approach on the “About” page (https://www.aiglossary.co.uk/about). The first criterion listed for glossary entries is: “Accuracy: Entries should reflect the breadth and key content of relevant academic literature in philosophy and empirical sciences at time of writing.”
If this doesn’t mean that the entries aspire to accurately represent what philosophers who specialize in the relevant areas think about the issues, then what does it mean?
The glossy would benefit from expanding the entry on AI chat bots and meaning to include some discussion of bullshit, or from a separate entry on the concept. The authors of the article “ChatGPT is Bullshit” would be excellent contributors to the glossary.
Thanks Aaron!
Ethics of Care would be an important term to be able to search.
I have some reservations about this project. It seems to me that its scope is somewhat unclear and a little confused. The “About” page on the website states that the glossary aims to contribute to the establishment of a common set of working definitions for terms like “intelligent”, “AGI” and “bias”. In my view, this is extremely premature. Having worked in the intersection of cognitive science, philosophy and AI for twenty years, I don’t think we have settled the disputes between theoretical frameworks which would be required to offer even working definitions of terms like “intelligent” (at least, not without implicitly privileging a particular theoretical framework).
The entry on “representation” is a good example. Within cognitive science, the “representation wars” are notorious, with some prominent theorists claiming that the notion of representation is actively harmful for understanding cognition in natural or artificial systems. Rodney Brooks’ essay “Intelligence Without Representation” is an early example of anti-representationalism in behaviour-based robotics. Anti-representationalists such as Inman Harvey have long challenged representationalists to provide explicit and unambiguous definitions of “representation”, which have not (in my opinion) been convincingly forthcoming.
How is the glossary going to help establish a common set of working definitions, if the theoreticians who work in these areas disagree on those definitions (or in some cases whether the concept is worth using at all)? For what it’s worth, the current entry does not even acknowledge how contested the notion of representation is.
So I think the supposition that the glossary can provide clarifying definitions (even working definitions) is misguided. At best, it might be able to offer an indication of how people in different contexts might be thinking when they use particular terms. Even that seems like quite a considerable challenge, especially without citations and references for the interested reader to pursue further.
While an entry on explainability and interpretability is forthcoming, the index should also include entries on justifiability (https://link.springer.com/article/10.1007/s10676-024-09754-w) and contestability (https://www.sciencedirect.com/science/article/pii/S0933365720301330) among others.
Thank you! Great ideas!