5 things I learned about using ChatGPT in science communication

This month I attended an excellent online masterclass with Mohamed Elsonbaty Ramadan, a science communicator and AI enthusiast. His three-hour masterclass comes highly recommended for beginners in ChatGPT as well as those who have a bit of experience using the tool. It promised to turn me from a curious amateur into a confident composer of prompts, and it didn’t disappoint.

So, what did I learn?

1. There’s a lot of hidden power behind prompts

Before this course, my idea of a prompt was something like, “hello, can you please tell me who has published papers on X topic?” or “ChatGPT, can you please summarise this article for me?” (I am very polite to ChatGPT, just in case it attains consciousness unexpectedly while I’m talking to it).

But it turns out there’s waaaaaay more to prompting. For example, you can ask ChatGPT to populate a content format, ask the right questions to help you think through your ideas, and adopt different writing styles, all by wielding the magic of prompt patterns, which are a bit like prompt engineering but much more powerful.

Prompt patterns tap into the inherent properties of large language models and the internet-based training data ChatGPT is fed on. For example, saying “Acting as a science writer, tell me about X topic” is way more effective than just “tell me about X topic,” because it gets ChatGPT to tap into, and emulate, all the science communication content that exists online.

2. ChatGPT is extremely culturally biased

ChatGPT is trained on data from the internet, and most of this data is from countries that speak English. Unsurprisingly then, it’s predominantly western in its perspective.

Mohamed demonstrated this by asking the tool to complete popular nursery rhymes that are traditionally sung in a) English, b) French, and c) Arabic. The English one came with lyrics and a potted history of the rhyme. The French one came in French with an English translation. The Arabic one was just an English translation that wasn’t actually correct. A poor show.

3. It’s an intern with superpowers

Mohamed used the analogy of a super-powered intern throughout the course, and I found this really helpful. (Science communicators absolutely love an analogy, and we do come up with some good ones).

Like an intern, ChatGPT needs to be supervised, and the output it produces has to be checked before it can be used. You are responsible for everything it does, including its mistakes and those infamous hallucinations where it invents people, places, and things out of thin air.

However unlike a human intern, ChatGPT never gets tired, has data processing power out the wazoo, and can absorb more information in ten minutes than an intern could in ten days or more. Basically, we can tap into its power, but we also need to accommodate for its limitations.

4. ChatGPT is continually being updated

Whether you have the free version or the fancier ChatGPT4.0, the capabilities of the tool are being refined and updated all the time. So much so that Mohamed has to keep re-writing his masterclass slides to keep up with it.

5. The ethics part isn’t a nice-to-have

Before we do anything at all with ChatGPT, we need to have a handle on the ethical questions it raises. That’s stuff like privacy – everything we type or paste into ChatGPT goes through the OpenAI servers, which means it’s not kept confidential. It’s not going to be suitable for sensitive information or IP that’s protected by an NDA.

Then there’s the inherent bias, like the cultural slant I mentioned in point 2. Our creations, when assisted by ChatGPT, absorb its biases. So, we need to have a think about how we balance this out when we’re communicating to the public.

And of course, it’s very very power-hungry, chomping through electricity at a much greater rate than regular cloud computing. Environmentally speaking, it’s a concern, especially as the number of queries multiplies and the market for ChatGPT-like tools increases.

There aren’t any new dates scheduled at present, but you can register interest for Mohamed Elsonbaty Ramadan’s “Using ChatGPT to communicate science masterclass” to stay informed about future courses.