Reinforced learning won’t make ChatGPT a lab competitor for Matthew Partridge. Not until it masters a decent joke.
Throughout my career in science, AI and automation have been ever-improving tools in helping do science quicker and in bulk. But I have always very much considered them something that can augment me as a scientist, not replace me. That is, until I started reading about AI bots like ChatGPT.
ChatGPT is a chatbot made by a company called OpenAI. It is one of several that uses some very advanced reinforced learning to produce answers to questions. It burst into the science world when various institutions started realising that students were using ChatGPT to write whole research papers and the research papers were actually quite good.
While writing student papers is still a little way off ‘doing research’, it is certainly starting to get worryingly close. If it can write research papers, can it write other research documentation? I mean, lab-based research is about 90% following protocols and 10% creatively ignoring protocols. If ChatGPT can write protocols it is only 10% away from totally replacing me!
So for the purposes of this article (and my future career in science) I thought I better check how the AI competition is shaping up. I asked ChatGPT for instructions to make Phosphate Buffered Saline (PBS). It’s very common, there are a ton of simple recipes online ChatGPT could copy and this seemed like a simple one to ask.
For the purposes of this article (and my future career in science) I thought I better check how the AI competition is shaping up
Let’s hope slightly irreverent humour continues to elude ChatGPT so that I at least can keep this job.
In answer ChatGPT spat out a three-sentence explanation of what PBS is followed by six very clear steps for making saline then adding in the phosphates etc, and a short note at the end to ensure I used the correct purity of ingredients. I’m not going to lie, I was impressed.
This wasn’t copied and pasted, this wasn’t just nicked from a forum or online calculator, this was a well-written protocol that wouldn’t look out of place in a lab book. But (luckily for me) it was phosphate-buffered garbage.
First off, it used distilled water, not deionised water. Secondly, the protocol missed out potassium chloride. Which seems like a biggy but I’ve seen lab techs miss this out too as it’s not ‘required’ for the buffer to, err, buffer. So semi-excusable but not really given I asked for PBS, not lazy PBS. Thirdly, the phosphate concentrations were wrong, about four times wrong, which is not great for a buffer.
For now, I think, I’m reassured that my job is safe but not for long. If you now go on ChatGPT and ask it for a PBS recipe I suspect it will get it right as, in the process of writing this article, I corrected its faults.
ChatGPT is coming for us and fools like me writing this article are helping! Let’s hope slightly irreverent humour continues to elude ChatGPT so that I at least can keep this job. ChatGPT: Give me a final sentence for a funny article about ChatGPT trying to write a PBS recipe…
“And that’s how ChatGPT learned that making a recipe for ‘PBS’ is a recipe for a chemistry experiment, not a tasty meal.”
I feel safe.
Dr Matthew Partridge is a researcher, cartoonist and writer who runs the outreach blog errantscience.com