Reading Time: 4 minutes

Ethan Mollick, whose One Useful Thing posts are always readable, suggested 15 immediate uses for GenAI. I disagreed with some of them, but I thought the exercise was worthwhile, so, here are my suggestions for research but also for general use. Most importantly, I’ve tried to steer clear of the “by all means use genAI if you are confident you can evaluate the answer”.  Instead, what these suggestions have in common is that they are ring-fenced, limited to collating, or focusing, or amplifying, or simply providing alternatives, to a prompt. This means you don’t always have to know lots about the subject to benefit from the AI results. And let’s be open about this: why would you use Google if you already knew the answer?

  1. Summarization – of short documents, up to (say) article level. Mollick links to a fascinating paper that got genAI tools to summarize whole novels, and then check the results by humans who had recently read the relevant novel. My immediate comment on this is that even literature students (of which I was one) have great difficult summarizing a novel they read just a few days ago, let alone one a year ago. But in any case, the paper revealed a poor ability by the AI to summarize correctly, over a large scale (they say around 121K tokens per novel, or 85,000 words). The main problem was specific events or claims, such as claiming the two characters had a brief affair, when the narrative states clearly they were married. Does this kind of thing affect summaries of scholarly articles? I think not.
  2. Collating several individual viewpoints. A great example of this is to look for recipes. Google Search will give you plenty of recipes, by named authors, and you can then review them and create a mix of different ideas to meet your needs. GenAI, in contrasts, will give you one recipe, without attribution, and does the work of merging the different approaches into one. The result is immediately usable, although perhaps less valuable, if you want to see what the individual permutations might be.
  3. The initial literature review. By this I mean the process many academics need to carry out, in order to keep up with the literature in their domain. Every few weeks then need to reduce, say, 100 papers, with titles that look relevant, to the 10 or so that they will start to read in detail. This sifting process is already being done by a number of commercial tools, following a simple methodology: provide the system with a few training articles, then see the results, and fine-tune them by selecting from the results the hits you think are most promising: in other words, prompt engineering without using the term.
  4. Expressing something with a different tone or point of view. Asking the AI to repeat the answer more simply, or in a more complex way, or from a different angle, can be very effective.
  5. Providing a supplement to human thinking. When I have to create a form, or an application, I write something, then run it past the AI to see if there is anything I’ve missed. Crucially, I’m not asking the machine to be comprehensive – I’m asking it to supplement what I have already thought.
  6. “Just one more viewpoint”. Mollick co-authored a paper on this. which is essentially supplementing an individual’s brainstorming. However many ideas you might have, the AI will hve more. the AI can provide hundreds that do not meaningfully repeat.¹
  7. Definitions of technical terms. This proves surprisingly problematic with Wikipedia, not because the definitions are not correct, but because Wikipedia’s level of complexity can vary widely for different terms in the same domain, and the same thing can happen with Google Search, for similar reasons. I find genAI provides a more consistent level of technical difficulty and a well-structured explanation.
  8. Completing something – Mollick describes it as “work that will keep you moving forward”. Again, the key is that you have provided a detailed framework within which the system can achieve its results.
  9. Asking natural-language questions while you are reading, to help your understanding – this is more or less happening already with Google’s AI Overview, which frequently just lifts the most appropriate source from Google Search hits. All undergraduates would benefit from this.
  10. Chronologies. Ask genAI for the ten most important dates in the history of France, for example, the result is surprisingly good – and considerably more useful (in my opinion) than reading the Wikipedia entry for “France: History”. Is it 100% reliable? I don’t have any doubt that the key dates and events are all genuine. Whether I can extend this excellent result to, say, the ten most important dates in Nigerian history (about which I know nothing) is another matter – but this is no different to using Google, or Wikipedia, for that matter. Based on the results I have seen for countries I am familiar with, I would certainly use it.

I should point out that the answers for this post, using Copilot, varied when I carried out a similar request on a different PC. But the difference could be minimised by providing a more precise prompt.

I’d love to hear from others if they have found other uses for genAI that they are confident to put into practice immediately, and trust the results.

Views: 260