In Defence of AI
During my PhD (2017–2024), I almost never considered using artificial intelligence. Looking back, there were plenty of opportunities where it might have helped.
I was building a corpus, and the idea of automating transcription certainly appealed to me. In practice, though, this never took hold because the corpus was a clinical dataset focusing on disorganised speech. Systems trained on healthy speech do not perform particularly well when transcribing thought disorder. There was another reason, too. Manual transcription gave me a level of control over my data that automated transcription simply would not allow. I was able to develop my own transcription convention, designed to capture details that I and other researchers could later analyse. This meant spending eight hours a day, for almost a year, holding a fairly complex transcription convention in mind while agonisingly working through more than twenty hours of fuzzy audio.
I also developed a linguistic experiment. The initial plan was to run it in a laboratory setting, with participants travelling to the university. The COVID pandemic required me to redesign the study so that it could be delivered remotely. I had already spent years designing the task and several months implementing it in SuperLab. Were it not for the release of SuperLab Remote at exactly the right time, I may have had to learn Python and rebuild the experiment from scratch. That could easily have added a year to the timeline and may even have threatened the completion of my PhD.
Most of my doctoral work involved wrestling with ill-defined concepts and a literature filled with gaps. There are only a handful of researchers working directly on thought disorder, and relatively little attention has been given to possible links between thought disorder and creativity, although there has been some renewed interest in recent years. An AI would not have been much help at that stage. Large language models generate responses based on existing patterns of knowledge, and what we knew about creativity and thought disorder prior to my thesis was very little.
So I spent a great deal of time with books. Books on thought disorder, creativity, linguistics, psychology, research design, and corpus building. I found myself tracing what had been written about thought disorder over the last century, identifying where interest had faded, and considering what broader schizophrenia research might contribute to the puzzle. Much of the work consisted of following loose threads and seeing whether they connected.
I am glad I took that route because completing a PhD without involving LLMs forced me to develop skills I would not otherwise have acquired. If I had asked an AI to design an experimental task, it would almost certainly have produced something competent. It would also have been heavily shaped by the conventions that dominate psycholinguistics, experimental linguistics, and experimental psychology. The more interesting outcome of doing things the hard way was that I continually encountered methodological gaps. Nothing that had been done before quite met the needs of my project. Nothing directly addressed the specific problems I was trying to solve.
I would not have developed a novel method for norming linguistic stimuli in a clinical population by consulting with AI. Nor would I have thought to take a reviewer's criticism that “this reads like a tutorial for students” and rewrite the paper as an actual tutorial in order to get it published. That remains one of my favourite publication experiences. Rather than defending the paper against the criticism, I accepted the premise and leaned into it. AI does not generally suggest that sort of lateral move. It weighs, averages, predicts, and often reassures. Useful qualities, but not quite the same thing.
After my PhD, I began experimenting with AI. The result of doing so is this website, my YouTube channel and podcasts, the Python implementation of my experimental task, and three of my four books.
AI did not script my videos or generate the ideas behind my books, but it allowed me to take material I had already created and repurpose it into new formats. This is the workflow that has worked best for me since integrating AI into my projects. Rather than asking AI to generate ideas, I ask it to transform my existing work.
Androids DO Dream is one example. The book began as a collection of ideas spread across my thesis, a memoir of psychosis, years of video essays, and podcast discussions. I gave the model those materials and asked it to produce a short mass-market book in my voice that adhered closely to the themes of my existing work. The result sounded surprisingly like me, not because AI can read minds, but because it had access to a substantial amount of personalised source material and a very clear prompt.
The Python implementation of my experiment is another example. AI did not design the study. It did not spend two years standing at a whiteboard, chain-smoking with a migraine, trying to make linguistic stimuli, theoretical concepts, research design best practices, and statistical rules fit together. It was given my methodology paper and PhD thesis and instructed to build a software replica of what participants experienced during the study. It succeeded because it was working from detailed, high-fidelity source material. The difficult intellectual work had already been done.
How Language Holds provides another example. This book was produced by an LLM using my own work as source material. It was given my thesis, memoir, and corpus and asked to identify patterns in the interview transcripts that reflected preserved ability, expression, and adaptation rather than simply signs of disorder. The instructions were explicit: be transparent about the method, acknowledge the limitations, and avoid presenting the analysis as rigorous academic research. The result is not a scholarly monograph. It is a way of showing readers another side of the data, one that they can engage with and reflect upon. Had I not transcribed those interviews manually, many of the patterns highlighted in the book would likely never have been captured in the first place.
This is the use of AI that I find most compelling. Giving AI real human ingenuity, creativity, and effort and asking it to automate tasks for which the original creator no longer has the time, energy, or inclination. That feels fundamentally different from asking a model to generate a novel, a painting, or an entire creative project from scratch. While newer models are increasingly capable of such things, I still believe there is more satisfaction in creating something yourself and then asking AI to do something interesting with it.
What strikes me about my own use of AI is that every successful project has followed the same pattern. The model performed well whenever it was given a large body of high-quality human work to transform. It was far less relevant during the earlier stages where the actual conceptual breakthroughs were being made. The corpus had to be built. The experiment had to be designed. The ideas had to be developed. The books had to begin somewhere.
Every successful use of AI in my own work has followed the same principle: start with something that required human effort, expertise, or creativity, and then use the model to transform, extend, or repurpose it. My thesis became software. My videos became books. My corpus became a different way of thinking about language and psychosis. None of these projects began with AI, but all of them were made easier by it. To me, that feels less like cheating than a natural extension of creativity itself.
