How we deploy AI will decide the advantage it confers

It could take some trial and error to work out how best to maximize AI’s potential, but it’s worth it

is a partner at Trilegal and also has a podcast by the name Ex Machina. His Twitter handle is @matthan

When ChatGPT was first made publicly available, I remember marvelling at the quality of its output and thinking that AI had finally improved to the point where it might be good enough to stand-in for me. Wouldn’t it be wonderful, I thought, if I could describe what I’d like my article to cover and have AI write it for me.

So I set about doing just that. I taught myself to get ChatGPT to generate sentences that I could fashion into paragraphs— which, in turn, could be woven into full length articles that, at least at first glance, looked like something I might have written. I even managed to use this process to generate an article that was accepted for publication. While the process was more laborious than I might have liked, it left me feeling that with just a bit more practice, I’d be able to get AI to write articles for me.

I am sorry to say that this was not how things turned out. Even after training it on every article I have ever written, I simply could not get ChatGPT to generate material that even remotely resembled what readers have come to expect from an Ex Machina column. Not only was the style stilted, the content was banal—utterly devoid of the sort of insights and nuance that I’d like to think characterize my writing. As for the one AI-generated article that did get published, almost everyone told me that by the time they got to the disclosure at the end, they were wondering why it felt strangely off-key.

At its present stage of development, AI can do little more than creatively regurgitate existing content. There is nothing objectively ‘new’ in the output of large language models considering the extent to which they remained constrained by what their training data-sets contain. And since they cannot offer a novel viewpoint, they struggle to even begin to approach the sort of writing I am looking for in either style or substance.

I had gone into this experiment hoping that AI would replace me—that it would allow me to dash out 950 words a week at a fraction of the effort I currently expend. I soon realized that, as useful as it was, getting AI to generate whole articles was impossible —not just because its ideas lacked nuance, but also because of a certain je ne sais quoi that, try as I might, I could not get it to create.

If I wanted AI to work for me, I needed to how it could be put to good use. And so I set about developing a brand new workflow—one that allowed me to leverage AI as much as possible while still ensuring that the final output continued to meet my expectations.

My core realization was that despite its impressive advancements, AI struggles to make the sorts of intuitive connections that humans do with ease—understanding the unintended consequences of new legislation, or extrapolating the impact of executive actions based on similar measures undertaken in unrelated domains. Humans excel at drawing analogies like this based on their knowledge and experience and using this to predict possible outcomes and risks. Since this is much of what my articles address, I have come to the reluctant conclusion that, at least for the time being, I will have to continue to ferret out new article ideas myself.

So I make note of ideas as they occur, logging them in a repository so that I always have a bank to choose from. Once I know what it is I am going to write on, I open ChatGPT and ask it to generate an outline for me. This is one of the things AI excels at and I find that not only do its suggestions organize my thoughts well, the structure that it suggests serves as the basis for a first draft with very little additional tinkering.

In almost every instance, some aspect of the outline calls for additional research —a deeper articulation of some concept or the identification of examples and use-cases that illustrate a point. For this, I turn to Perplexity, an AI-based research tool that not only finds me the answers I need, but does so with reference to source materials, allowing me to fact-check the material before I use it.

Armed with a ChatGPT outline and Perplexity research, I can finally sit down to write. This is when I turn to AudioPen, a wonderful little AI powered transcription service that takes the content you dictate, sprinkles some AI magic over it and generrethink ates clean text with none of the stutters and logical inconsistencies that inevitably insinuate themselves in regular dictation. I have found the final product to be polished and so unmistakably in my ‘voice’ that, more often than not, I have to spend no more than an hour editing it into an article that my editor would be willing to accept.

We spend so much time worrying that AI will replace us that we often fail to appreciate the extent to which it falls short of what humans are truly capable of. But even if AI cannot replace us, it can improve our productivity well beyond anything that is currently possible.

The reason I took you through my article-writing work-flow was to illustrate just how easy it is to integrate AI into our lives. Now that AI is widely accessible, we all need to figure out how best to leverage AI so that we can address the inefficiencies that hold us back.

What shape that will take, what it will actually do for us and how will surely differ for each of us, specifically suited to our individual ways of working. But one thing is clear: It is only those who learn to use AI most effectively who will have an advantage in a world where AI could become as ubiquitous as electricity.