Your AI Summary is (probably) Not What You Think
A few posts coming through my feeds recent circled this idea of using AI for summaries. In the L&D space I’ve seen this put forth as a potential/recommended use case for GenAI. First up was an article shared by Stephen Downes about someone who used a bunch of Adobe products to convert speech in a conference session to text and then save it to PDF (because Adobe). I share Downes’ side eye about that bit. As for the workflow, it’s fine if you already have (overpriced) Adobe licenses, but I’m pretty sure you’d have a simpler workflow using MS Word’s built-in speech-to-text or something similar at a fraction of the price (time, money, and effort). And you’d get it immediately in a more useful medium. Downes added,
It made me think back to attending conferences in person, but always having an eye on Twitter (RIP) for what the buzz at a conference was. I’m not sure you’d get the same effect from summaries of sessions that weren’t filtered through the people in attendance. Some of the best conference tweets were snark, and I’m not sure you’d get that in summaries. I’m ok with waiting generally for people to get on their blogs, although that doesn’t cover every session. Anyhoo, onto the next piece, also talking about summaries. Clint Lalonde recently recapped his experience with an AI tool called Recall. I love the breakdown and analysis provided, but what stood out to me were these bits,
The auto-generated questions was kind of interesting, I use Nolej for similar sorts of workflows. But this bit about the summary not being great. I’ve noticed myself that summaries can be hit and miss. I’ve been using Grammarly for a long time and using Grammarly GO since it became available in Canada. One of the suggested prompts they provide is “shorten it”, which I make extensive use of. That command in Grammarly is a ruthless editor, going well beyond what it needs to depending on how you use it. But the pattern of how it shortens text I’ve found to be pretty similar to what I find with “summarize it” style prompts. Whether you put the text to be summarized in the chat box with Copilot, or load in a PDF, or link to a web article, the summary has this quality about it that is similar to “shorten it”. That is qualitatively different from what we tend to mean when we ask for a summary of something.
To that, I ran across a post from Ethan Mollick about an Australian study that creates a typology of errors GenAI makes in the summarization of books. He does point out that the researchers only just published this, meaning that the experiment was done a while ago, and that they didn’t use frontier models. So there are some particularities that would affect the results. He also links to studies that find GenAI does an adequate job of summarizing books.
Ok, away from blogs, and social media, I also ran into this topic of AI summaries and problems on the podcast, Mystery AI Hype Theatre 3000. In the linked episode they mentioned an article When ChatGPT summarises, it actually does nothing of the kind. I encourage you to check out this article, because they give an excellent breakdown of their experience using ChatGPT to summarize documents, in potentially high profile environments, and the implications of that. Their detailed account brought me back around to that similarity between “shorten it” and “summarize it” and I see those similarities even more now. Maybe confirmation bias.
What have your experiences with GenAI summaries been?