This page describes how I do and don’t use artificial intelligence for my science writing.
The main things you need to know:
- I DO NOT use generative AI to produce text for any of my freelance articles or books.
- I DO use AI-based transcription software to quickly create typed summaries of audio. This saves me hours of (otherwise unpaid) work and helps me to capture all important details from interviews. It also ensures that I get the most accurate quotes from my sources.
- I RARELY use generative AI as part of research or to come up with new ideas.
Here are some questions you might have:
Which AI-based audio transcription programmes do you use?
I use Otter to transcribe English-language audio and Amberscript for other languages (like Dutch)
How do you use the transcriptions generated by AI software?
I mainly use the text to help me recall the interview. I also highlight sections that I think could make good quotes, but I always listen back to the audio to make sure that the text is correct before I quote or reference anything in my article.
Do the transcription programmes use the interviews that you upload as training data?
Yes, that’s how the software learns and improves. Here are the data security and privacy policies for Otter and Amberscript that include information about encryption and my ability to remove content. Both companies are in jurisdictions with well-defined data privacy regulations (California and The Netherlands/EU respectively).
What is generative AI?
Generative AI is the type of AI that creates new text or images. It’s different from the type of traditional AI that, for example, finds patterns in microscope images.
Have you used generative AI for research?
Occasionally I have to search for difficult-to-find pieces of background information, and I sometimes use Perplexity to speed this up. I always double-check the information and source material it gives me.
An example of this type of search: Let’s say I’m writing a long feature about a new disease treatment. Somewhere in the introduction I might want to illustrate the impact of this disease by saying something like “affected X number of people in the last decade alone”. It reads like a quick fact, but it can take hours to dig up this sort of information in PubMed and Google, and then I still need to compare a few sources to see if they agree on the stats. Perplexity does the same Google and PubMed searches I would do, but much quicker, so I can get straight to the “see if sources agree” step. It provides links to the source material and prioritises academic information.
I’ve only used ChatGPT for research once, before I knew that more academically oriented platforms existed. After I spent a few days trying to identify trends in a field, I used ChatGPT to check whether it had picked up any other trends that I missed. (It hadn’t).
I’m very aware that generative AI programmes can hallucinate and return completely made-up information, so I always check all information and never copy/paste any text these sites generate.
Have you used generative AI for any type of content in the past?
In my own newsletter I used ChatGPT once, as a gimmick, in an issue that was all about generative AI. I disclosed it in the text that time.
When generative AI-created images were still new, and before I realised that they relied on stolen art, I did create a few images that I would have shared online at the time, but I have stopped doing this.
Will you use generative AI to create content in the future?
I don’t plan to, but if that ever changes I will update this page.
Has anything you’ve written been used to train AI algorithms?
Yes. Meta’s AI used LibGen, which includes several of the articles I’ve written for the front sections (the magazine part) of scientific journals.
I don’t know for sure if ChatGPT has used my writing, but I would assume that it has. I’ve asked ChatGPT what it knew about me, and in its early days it was hilariously wrong about that, but in a way that suggested that it had conflated “information about Eva” with “articles written by Eva”. So yes, I think my writing is in the training data. I don’t love it, but on the other hand I would prefer that it learns from a wide variety of good writing and for that it needs to grab well-edited articles.