The magic of "AI", like so much magic, comes from being able to make a request with your words and have it fulfilled.
To do that, our genie has to turn words into meaning. So, at the core of most AI is an LLM, or Large Language Model. Their most famous ambassador is GPT, the Generative Pretrained Transformer. They're big, complex maps of the patterns in a language, and how they group together into things we might call "ideas", but which to the model are just clusters of data points that it's blindly seen together often.
So for starters, don't call it AI I have no time for people who think GPT is a bad name. It does exactly what it says on the tin. It's generative: It's set up to output something that didn't exist before. It's pretrained: It's already absorbed most of the internet, with all that that entails. It's a transformer: It excels at transforming text, from one format to another.
Having some understanding of the origins of the technology really helps. Transformers originated as translators. They matched clusters of words in one language to clusters of words in another language, and figured out the patterns for moving between them. In a lot of ways they still are - they're just translating within a language, or from language to picture, or from language to code, rather than between human languages. From full length to summary, from boring sentence to joke, from list of themes to narrative and back again. From description to image and back, from English to JavaScript and back - it turns out a lot of useful tasks are, basically, translation.
Don't trust any information you didn't give it. These models excel at generating patterns, and if you don't tell it what to populate that pattern with, it'll make up something that looks right that seems to fit. They can't, as I have seen some people claim, "show you their sources"*. What they can do is generate something that looks like a source list. Likewise, you can't correct them. You can just ask it to generate something that looks like it's been corrected. It's fantastic at structure, but has no idea about meaning, let alone truth.
* Although between me starting this draft a month ago and finishing it, (along with more-or-less live access to the internet) citing sources does seem to be on the cards, and I think that's probably a good thing.
Be aware of your biases around language. As humans, we're so strongly primed to believe anything that uses language is smarter than it is. A conversational chatbot was by far the most compelling and intuitive, but also perhaps the most misleading, way to introduce a large number of people to this tech. It's not a coincidence that nobody's really claiming the image generators are sentient. Haunted, maybe, but not sentient.
An AI powered future looks less like consciousness, or conversation, and more like autocomplete. We'll have autocomplete on everything - but it's still going to need humans in the loop to decide what actually needs doing. It isn't agentive, and never will be - not this tech, anyway. We've found a way to turn words and pictures into maths and back again, and found so many applications for that, and it is going to be transformative - pun absolutely intended. But as far as "consciousness" goes, this is a dead end - GPT is no more "conscious" than a thesaurus with particularly good index.
This is going to do for words, and art, what spreadsheets did for numbers. Nobody's claiming Excel is sentient because it can see patterns and autofill cells and make a guess at what graph would best fit this data series. And, similarly, Excel didn't put (say) accountants or analysts or engineers out of business. It made them more efficient, and increased their throughput, and made them more capable, and maybe increased expectations - for better or worse - around what it's reasonable for them to accomplish.
Think of the the things it can't do, because most people work in the physical world still. Just like spreadsheets haven't changed those jobs, neither will AI. When you're claiming that this changes everything and affects everyone, be more precise in who and what you're talking about. Yeah, they can restructure an essay like nobody's business - but they're an extremely long way off changing hospital sheets or stopping kids from eating too much paste.
Think of the things it can do as well. It's going to keep getting easier to generate industrial scale filler. Google is already useless, clogged up with fake SEO cruft, and this will only make it more so. I think we're already seeing some realisation of this take hold - it's fascinating to watch people's realisation to the internet go from fairly openly trusting everything, to curating their communities and sources much more tightly - or at least trying to hack things so they know there's a real human involved somewhere. Even if it is, god forbid, on Reddit. It's also fascinating (and a little ironic) watching the the trainers scramble for uncontaminated, known-human training data.
Think of the things that you do, too. I don't see a future where I substantially rely on AI for my writing, because for me, the process of writing is the important part. That's why I'm writing this. Thinking about how you explain, and express, and structure an idea is how I learn things, and the output is almost irrelevant. One of the valuable parts of being of a science communicator, I think, is in having someone who knows how it feels to understand the thing you're talking about, and who can guide others through the same process - and the end result is irrelevant to that, whether it's an article, a show, an activity, whatever. It's the art of knowing where you need to put the stepping stones for people, how to space them so it's always just at the edge of what they can accomplish to keep things challenging and exciting, which bits you can skip and which bits are foundational, and how to adjust all of that for the people you're trying to bring with you on the journey. The article, the image, the video, is the starting point for that journey, not the end.
Someone still has to be responsible. I don't see a future where scientists don't need someone to do their comms with them, because writing the press release, or the tweet, or the human readable summary, is the least important part of the job. Someone still has to take responsibility for getting the summary written, for getting it up on the website, for getting it out there to people, for determining if it's important and what the story is. And I think responsible is the key word - even if you automate that whole pipeline and ask scientists to drop a raw abstract in one end and auto post the results to some kind of feed, that's still an investment of time and it's still an investment of responsibility. That applies everywhere. Generative models will never have intent, and will never be able to take responsibility, any more than a hammer or a dictionary or any other tool can.
I'm sure this will all change. It's changed between me drafting this post and reading these articles a few months ago, and me posting it now. Right now, I think these are sensible guidelines. I think these are useful things to remember. I think generative models will be useful. I think they'll become normal. I think they'll become another tool in the toolbox, and probably nothing more.
But I could be wrong, and I guess we'll have to wait and see.