For a lot of people, (and by a lot, I certainly mean the vast majority, on the 99%+ level), The past few years of AI, from Jarvis to ChatGPT, came out of nowhere. Heck, I’d probably say that for 90% of people they didn’t even use Jarvis and so just ChatGPT itself just suddenly appreared out of nowhere for them.
Very few people were much interested in the research science of AI. Most of you probably never heard of PaLM or LaMDA at all until very recently, despite both of those predating ChatGPT. That’s not a criticism, simply a statement of fact.
Heck, I saw articles and posts on social media from people who were ‘supposed’ to be smart SEOs, talking about the novel idea of maybe Bing using an AI in a search engine, completely ignoring that Google had already been using a dozen different forms of AI and Machine Leaning in their normal results for literal years before. It’s simply that those AI were not something you directly or openly interacted with. Those AI did things like process your query and attempt to get a better understanding of not just your meaning, but the intent behind that meaning.
There were AI that helped detect and remove spam, AI that had been installed into Google Maps so its directions would result in an average journey time reduction of 30%, and many, many others, but you couldn’t talk with any of them.
That’s a key difference, one of the absolute central points, and I want you to bear it in mind.
ChatGPT is very, very far from the first AI that could work out complex things. AI started to out play Chess Grandmasters long, long, long ago. Then it was Go players. An AI can attempt to model thousands of possible choices, and the predicted outcomes, in the blink of an eye to choose the one that has the statistically highest chance of being the best move. Whether that is a chess move, or mapping network connections, or route planning, or… predicting the next few words that are most likely to produce writing that matches the patterns it has been taught are ‘good’.
What it is not doing is ‘creating’. It is not ‘writing’ in the way we do, to think about the reader, to imagine ourselves communicating with them, and trying to think of how to keep them engaged and get our meaning across. Instead, it is looking at the patterns of every document it has been trained on and then predicting which words will most closely bring results that match one of those patterns. It is specifically, by design, trying to write like things that have already been written before, using facts and statements that have been used before, in the most successful (and thus widespread) patterns and styles.
All AI generated writing, at this stage in AI development, is inherently derivative. It is based on what already exists and was successful. It is formulaic, because it actually does detect and copy formula.
It is really important to understand this. AI can write great ‘filler’ content, the predictable blurb on the side of your box that is the same as the writing on the side of anyone else’s box, and that’s okay because nobody reads the box anyway. In addition, AI will create better filler text, faster, and cheaper, than using a human who knew nothing about the topic, and hadn’t one original thought or creative idea to their writing. But regardless of who wrote filler text, or why, or how fast, it is still filler text - it is still derivative, uncreative, formulaic, predictable, and ultimately, not something you’d use for an article, or where you want people to be impressed or persuaded by that writing.
Now, there are times where you can inject all the creativity and usefulness into the prompt, and all you need an AI to do is impose order and pattern on that information. That’s fine. If you are hooking up a couple of data sources into a prompt, data that has never been merged together in some cool and useful way, then AI can write the filler text and impose the structure at a scale that simply couldn’t be done by humans. That’s where AI generated content can shine.
But if there is something that could be written by humans, and is meant to be read by humans, and whether those readers are impressed and persuaded matters… That is the last place you should use AI.
So why are so many people using AI in the absolute worst way?
The answer is a cognitive bias called The Dunning-Kruger Effect. Those who are not good at writing, don’t have a great knowledge and experience of writing, and thus are not good at knowing good writing from bad. Their lack of ability in literature means that they literally can’t tell the difference between good and bad, and instead tend to judge writing by how many words it has.
When those people see that ChatGPT can write words, and they can’t tell that the words are unoriginal, derivative, formulaic, or even why those things matter, well, they don’t understand and figure words are words.
There’s a second part to that same flaw though. Those who use ChatGPT or any other AI because they are poor at writing, at using language, are missing the really, really huge ‘elephant in the room’. Large Language Model (LLM) based AIs use language. They are built entirely around a finer understanding of language than any prior generations of AI. Your sole control over those AI, your entire input into how well they will work, is all down to how well you can use that one thing it understands: Language.
If you want to be ahead in prompt engineering in the coming years, then the single most important skill you can learn, is better language skills. The ability to more clearly express your needs in a prompt.
AI is not a tool to help you avoid learning how to write. It is an entire new economy that is based entirely on how well you can use language to prompt a machine.