by KM Wehrstein, Associate Editor and AI Consultant
About a year and a half ago, someone I know asked ChatGPT to name five reincarnation researchers. In the list it spat out in seconds, Ian Stevenson was first, of course, followed by Jim B Tucker, James G Matlock and another name I knew though I can’t recall, plus a Jane somebody, of whom I had never heard. As this news circulated in the community, a person or two asked me if I knew Jane, and I had to say no. She seemed oddly absent when I Googled her. Also oddly, a citation of hers for a paper published in the Journal of Scientific Exploration that ChatGPT had provided was missing some of the necessary information one finds in citations. Growing suspicious, I went on the JSE site to look it up, and found that it didn’t exist.
The fact that I ran into this unforgettable hallucination of an entire reincarnation researcher just as I was entering the world of AI has stood me in good stead ever since, as it etched on my mind the fact that AI can make things up out of the whole cloth, and prevented me from having any notion that it was infallible. The love-hate relationship the world has with large language models (LLMs – this being the proper term for ChatGPT, Claude, Gemini, Grok and the others that have become so familiar) divides people into two categories: those who think LLMs are infallible, necessitating those little cigarette-package-esque warnings on their sites, ‘<X LLM> can make mistakes; check important information’; and those who think that everything they produce is crap and you always end up doing more work anyway. The truth, as is usually the case with such dichotomies, is somewhere in between. It is also, with effort and know-how, controllable.
Since July of 2025 I have worked fairly extensively with LLMs on a variety of endeavours, textual, mathematical and visual. Some of the key truths I have learned are:
- It’s all in the prompt. The lesson of Jane is not ‘never use AI’, it’s ‘never ask an LLM to find X number of anything; ask it to find all that it can.’ They’re like the ensorcelled broom in the Sorcerer’s Apprentice: ‘Master told me to fetch water so that’s what I’ll do until I get a command to stop’, even until the apprentice is riding waves. AI work on the Published Reincarnation Cases Database (PRCDb) project especially has taught me that with good enough prompting, hallucination and LLM absent-mindedness can largely be tamed.
- LLMs are only as good as the people running them. Yes, the internet is awash with AI slop and academic journal editors are being swamped with it – because LLMs’ astonishing capacity to crank out content at the speed of light is being used by sloppy-minded humans. No creative/intellectual product of quality can be produced by someone who cannot discern or envision quality. Sturgeon’s Law is being proven on a previously unimaginable scale these days.
- They love to throw a lot of info at you. As an experiment (and every prompt is an experiment, one reason I so enjoy using LLMs), I’ve tried to train ChatGPT to write in my own writing styles. Its biggest challenge is conciseness.
- They will help solve their own problems. They’re the first to ask.
AI came to the Psi Encyclopedia (PE) after General Editor James (Jim) Matlock and I had a Messenger conversation about how it might help with editing. Having decided to add to each PE article an introductory summary of about 45 words and three highlights of less than 20 words each, Jim was anticipating spending months writing 725+ of these. I suggested that we have ChatGPT (5.4, Extended Thinking) do it instead. We tested and tweaked the prompt for about 45 minutes, I resolved that Chat would complete the entire job in under 24 hours, and it succeeded.
It was time to set AI policy for the PE, so between my recommendations and Jim’s modifications, we adopted the following:
- All articles must be researched and written by humans. Large-language models (LLMs) may only be used to write connective material such as introductory summaries, highlights, category and subcategory descriptions and similar.
- All writing by LLMs must be edited by a qualified human editor prior to posting.
- AI-assisted editing must be for PE style (British spellings, British-style punctuation, etc.) only, and may not alter the writer’s personal style or the article’s content or meanings.
- A formal statement declaring and describing our use of LLMs will be posted on the ‘About’ page.
In the ensuing months, Jim has asked me if ChatGPT can do the following tasks. I have answered in every case ‘yes’, and we’ve then had it either complete or begin to:
- Edit articles to PE style (Chat got into trouble for editing with too heavy a hand until I put the fear of Jim into it), including the always-tricky lists of citations
- Search the web for article illustrations
- Write 8–15 tags for each article (useful in searches)
- Fill in important information it can glean from the site into our master spreadsheet (our tool for organising the PE workflow)
- Work on editing and ordering the Links Library
- Work on compiling and editing comprehensive PDF bibliographies for downloading from links in articles about our most prolific researchers
- And a few other more minor tasks.
We’ve also planned for the future:
- Creating an AI-powered ‘Psi Chatbot’ for the PE to interact with users, answering questions and referring to articles
- Creating a keyword system for better searching and to encourage LLMs generally to access the PE for their education
- Testing Claude and possibly other AIs for performance comparison with ChatGPT.
So why are we employing a technology which in some quarters is considered a travesty against human reason, sense and sentience?
I can explain it simply by a personal policy I adhere to with AI generally, and that the PRCDb team has adopted also: use it for everything it can do so as to free humans up to do that which only humans can do: have original ideas, draw directly from human experience, meditate for answers, contact other humans, exercise agency, etc. By having the more mechanical tasks performed by LLMs, Jim, the other associate editors and I are freeing up our time and SPR’s money for the human-specialised tasks. We haven’t calculated the savings exactly, but a rough calculation on writing introductory summaries and highlights for each article, based on Jim’s estimate of a half-hour to write one set and the 15 hours it took ChatGPT to write 720 sets, shows a 96% savings. This massive efficiency has enabled us to either complete or get a good start on PE-improvement projects that otherwise would have been on a years-long to-do list. It has also opened up possibilities we would not have considered before, the Psi Chatbot being a prime example.
We hope our use of AI has made and will continue making your use of the PE more productive and enjoyable. Excelsior!