2026-05-28: If LLMs can write abstracts, what's our job? The Uncanny Valley and Gell-Mann Amnesia Effect in the ACM Digital Library
If LLMs can write abstracts, what's our job?
The Uncanny Valley and Gell-Mann Amnesia Effect in the ACM Digital Library
Michael L. Nelson
2026-05-28
I serve on the ACM Digital Libraries Board, and we are navigating a number of changes to the ACM's Digital Library, which as a professional society and memory organization, is arguably the ACM's primary asset. A recent article (March, 2026) by Jack Davidson and Wayne Graves provides a status update of the ACM's move to open access, which includes establishing a "basic" and "premium" service level. Although there are some questions regarding the long-term implications of moving to open access, I, and presumably all authors, welcome the ACM's bold strategy for ensuring that our content reaches the widest possible audience.
Jack's and Wayne's article also addressed the DL's recent experimentation with AI/LLM enrichment of articles, specifically landing pages. And unfortunately, the experimentation got off on the wrong foot. Just before the holidays in 2025, the landing page for articles in the DL added AI-generated summaries as a sort of alternate or rival abstract. To make matters worse, these summaries were shown by default, and users had to select a tab to show the original, author-supplied abstracts. The figure below is an example taken from Dr. Casey Fielder (CU Boulder), whose social media post about the summaries being shown by default instead of the abstracts gained a lot of traction.
AI-generated summary shown by default (2025-12-16) for https://doi.org/10.1145/3706598.3713322
Fortunately, the expected behavior of showing the authors' abstract by default returned very quickly, and the AI-generated summary is now clearly marked as such, including the date that the summary was generated:
Author-generated abstract is now shown by default https://doi.org/10.1145/3706598.3713322
The AI-generated summary is now clearly marked as such, and includes the date the summary was generated https://doi.org/10.1145/3706598.3713322
First, let me be clear: showing the AI-generated summary by default instead of the authors' abstract was a terrible idea and was uniformly rebuked. The DL board was not informed that this was going to happen, and I can't recall anyone on the DL board even suggesting it; perhaps it was just an oversight by an ACM staff member or engineer at Atypon. I don't recall exactly when the expected default behavior was restored, but it was soon after the author community complained.
My original suggestion at the DL board meetings (echoed by Dr. Fiesler) was to provide wiki-style editing on the AI-generated summaries, possibly limited to logged-in authors (a possible premium feature?). One can make a good argument for either opt-out or opt-in, but neither option adequately addresses the problem of the sizable back catalog of unreachable authors (JACM began in 1954).
But what I find interesting is the level of author backlash against AI-generated summaries, at least as I observed on social media. This is all anecdotal, and I realize people don't post about things for which they are neutral or have even mildly positive feelings about because, let's face it: carping is a lot more fun. But Dr. Fiesler and the others in the thread are all reasonable people and aren't just trolling. I think there's something more fundamental happening. I think our collective reaction (revulsion?) to AI-generated summaries can be explained by adapting two phenomena: the Uncanny Valley, and the Gell-Mann Amnesia Effect.
The Uncanny Valley is an hypothesis that posits that our emotional response to depictions of humans (expressions, speech, movement, etc.) initially rises as the likeness becomes more human-like, and then takes a sharp dive as the likeness becomes nearly human-like but not quite. Basically, most cartoon characters, anthropized animals, etc. are "cute", but the more realistic animated humans in movies like "Polar Express" (2004) are just creepy.
The Uncanny Valley (Source: Wikipedia)
I propose that something similar happens with text. Most authors have no problem with AI tools enriching the work, for example: language translation, extracting citations, repairing/rewriting hyperlinks, suggesting related works, suggesting/assigning keywords and ACM CCS values, and any number of other services and derived content. But generating a summary that rivals the abstract? Yuck. No thanks. An error in citation parsing or CCS assignment? Meh, who cares, either ignore it or fix it, but no one takes to social media to complain. A subtle but detectable (if only by the author) error in a summary? That's glaring and viscerally wrong. And even if we can find no substantive errors, knowing the text is AI-generated, we will find fault with phrasing, the structure, and various minutiae (cf. humans' negative attitudes to replicants in Blade Runner). Extracting keywords is what computers do. Writing abstracts is what we do. If LLMs can write abstracts, what's our job?
Those assessments inevitably derive from us reviewing AI-generated summaries of our own work. Presumably, no one knows the material better than us, so the best anyone / anything else can do is be "as good as", certainly not "better". We're writing for our peers, and we share a nuanced, high-bandwidth vocabulary that outsiders just can't appreciate. On the other hand, if we have to read articles outside of our area of expertise, we often wonder why are the authors so obtuse? Why can't "those people" just write plainly?
This is the essence of the Gell-Mann Amnesia Effect, which was coined by Michael Crichton to describe the phenomena that the more you know about a topic, the more likely you are to see the flaws in a third party analysis, but at the same time not being as critical when that same third party summarizes a topic on which you are not an expert. Anyone who has been interviewed by the media has experienced this: the reporters inevitably butcher your hour-long exposition, provided in painstaking detail, covering all the nuances, edge cases, historical review, and possible future directions – all reduced to a minute or less of decontextualized soundbites. But that news outlet suddenly becomes a trusted and valuable source when they cover a topic outside of your expertise.
I suspect the Gell-Mann Amnesia Effect applies to AI-generated summaries as well: they are an abomination when applied to my work, but a useful de-jargoning tool for exploring unfamiliar or even adjacent sub-fields. This even presupposes that there should be multiple AI-generated summaries, aimed at different audiences (e.g., lay person, High School, undergraduate, researcher). In fact, the rival abstract in Dr. Fiesler's example might be the least useful summary, precisely because it does rival the author's abstract. But writing for audiences other than our own is a different skill set: writing for my fellow researchers at JCDL, Hypertext, Web Science, etc. is what I do, but writing for high schoolers is not what I do. Casting my work into something appropriate for high schoolers would be a good use of LLMs, and simplifications (if not outright errors) are to be expected.
In summary, I think it's natural to feel revulsion when the LLMs are used to rival our work: it falls into the textual uncanny valley, in a way that other generative works, such as translation, do not (at least not currently). But at the same time and based on the Gell-Mann Amnesia Effect, our harshest judgement of AI-generated summaries is reserved for areas in which we are an expert, and our assessment of AI-generated summaries improves as we apply them to areas further from our own.
With that in mind, it would make sense for the ACM DL to enable wiki-style editing on summaries, move away from the model of a single summary that rivals the author's abstract in length and complexity, and introduce multiple summaries, tailored to audience and intended purpose.
–Michael
2026-05-29 Update: I was chatting with Martin Klein, and he informed me that bioRxiv introduced in late 2023 on-demand summaries at variable reading levels. bioRxiv is far from my field, so I'm not completely clear on its status as a production service or just a prototype. For example, this recently published preprint doesn't show the option for AI-generated summaries:
Clicking on the "Automated Services" for the recently published https://www.biorxiv.org/content/10.1101/2025.05.23.655690v1 …
…shows "There are no automated services for this paper."
However, I was able to find this preprint from a year ago that does have that option available:
The "Automated Services" option is active for https://www.biorxiv.org/content/10.1101/2025.05.23.655690v1
When clicked, the default AI-generated summary is for the "General" audience:
The "General" AI-generated summary for https://www.biorxiv.org/content/10.1101/2025.05.23.655690v1
The "Expert" AI-generated summary for https://www.biorxiv.org/content/10.1101/2025.05.23.655690v1
Are these good summaries? I guess so – although I'm not sure what else to evaluate them against. I don't know the first thing about proteomics, so the "General" summary is certainly the most accessible to me. The "Expert" summary is more detailed than the "General" summary, but still more accessible to me than the authors' abstract. That's not a surprise because 1) I haven't studied biology or chemistry since High School, some 40 (!) years ago, so Schär et al. aren't writing for me, and 2) the summaries are both about half the length of the authors' abstract. I saved all three into separate files:
% wc -w bio-*txt | grep -v total
219 bio-abs.txt
107 bio-expert.txt
88 bio-general.txt
Two hundred words is a good target for abstracts. I'm guessing the prompts for the AI-generated summaries had a target of about 100 words, so by design even the "Expert" summary will not rival the authors' abstract (though metadata and wiki-style editing would be nice). The "Automated Services" tab has at the bottom a link to "Explore Further on ScienceCast":
The target of the "Explore Further on ScienceCast" link https://sciencecast.org/casts/jpdm4k710oet
I don't have an account (yet) on ScienceCast, so that's the end of my exploration for now. But there's clearly a bigger AI↔paper ecosystem to explore, for both me personally and the ACM DL.
–Michael
*Apologies for including Dilbert, but the options for Gell-Mann Amnesia Effect cartoons are limited.
Comments
Post a Comment