Let's chat (no GPT required)
In which we poke the (in)famous ChatGPT and see what it can('t) do
Bleep bloop
ChatGPT is all the rage at the moment. The chatbot was launched by OpenAI in November 2022 and it’s built on top of the GPT-3 family of large language models.
Large language models are machine learning systems trained on massive datasets of text. The models then learn to predict which words or sentences are likely to co-occur. At this point, the amount of data developers have access to is staggering and machine learning algorithms get better all the time. The result is ChatGPT, which has been making headlines all over the place.
But.
Of course, there’s a but.
Chatbots based on large language models are designed to sound convincing, to cast their virtual eyes across immense swathes of text, and put together a reply to our query that their algorithms tell them is the most likely thing we’d like to hear/read. In broad lines, I agree with Sarah Myers West - managing director at the AI Now Institute - when, in this interview, she says:
ChatGPT is more effective than anything before it at producing text responses that closely mimic human writing. But even there, I’d understand it as the newest version of older technology.
What ChatGPT does is sound convincing, and it’s very good at this.
Here’s the problem: convincing does not equal correct, and convincing does not equal creative.
(Disclaimer: I write as part of the Dayjob™ and in my spare time, so I am naturally skeptical of tools that claim to take all the writing out of writers’ hands.)
Correct?
Let’s start with the worst downside of ChatGPT: it’s very convincing. Even when it’s spouting absolute nonsense.
There are plenty of examples of ChatGPT answers that are factually incorrect, logically inconsistent, and grossly biased.
Welcome to the age of mass-generated misinformation. As more and more people begin to use ChatGPT, it’s likely that more and more people will blindly accept its output. In the land of online content, time is money and quantity is king. Why fact-check? AI knows best, right? Except when it doesn’t.
Remember, what ChatGPT does is put strings of words together based on statistical patterns. It doesn’t care - can’t care - for correctness. For example, CNET, one of the largest tech news websites globally, recently had to pull a heap of ChatGPT-generated articles because they were riddled with errors. Whoopsie. Not much later, major plagiarism alarms went off. Double whoopsie.
The appeal for small businesses and content creators to hop onto the ChatGPT trend is strong, but please, please, fact-check whatever output you choose to work with. There is already far too much misinformation out there.
My writing is nowhere near perfect and I will make mistakes. However, when I make a claim, I look for evidence that supports or disproves it; I can consult and interpret primary research literature; I flag my biases when I’m aware of them; and I can adjust my claims to correct errors I’ve made. ChatGPT does none of those things.
Creative?
This one’s personal. I’ve been playing around with ChatGPT, trying different styles and lengths of text, for different audiences, on different topics. Almost all results are… bland. Stale. Sufficient, perhaps, for crappy clickbait, boring business updates, or empty-headed high school essays. (Although for essays, definitely fact-check, see above.)
ChatGPT text does not sparkle; it lacks both levity and layers, and remains woefully bereft of voice and emotion. (If you’ve read some of my writing, you’ll understand why I take that personally.)
Especially when it comes to fiction, ChatGPT-generated works are dead on arrival. One of the reasons for this, beyond ChatGPT’s lack of understanding about what it’s writing, is that good fiction (in my opinion, your mileage may vary) surprises me. It conjures up new metaphors, unexpected plot twists, deep, gut-wrenching emotional struggles, unseen worlds never previously imagined, and so on. In my experience, any ChatGPT attempts at fiction don’t really include these things. Much of what it generates in terms of fiction has a strong ‘Simpsons did it’ vibe. I’m not alone in being unimpressed at the bot’s fictional forays.
Along the same (song) lines, Nick Cave isn’t a fan of the chatbot’s attempts at lyricism ‘in the style of Nick Cave’. His words:
What ChatGPT is, in this instance, is replication as travesty.
Tool?
None of the above means that ChatGPT can’t be a valuable tool. But it’s just that, a tool.
It’s good at summarizing a given text, for example. Or write an abstract. It can give you computer code or help you refine descriptive paragraphs. It can help you put bullet points into text or vice versa. I haven’t tested this, but I’m sure it can give you a list of dad jokes. It might provide a spark of inspiration. Or, also valuable, it might show you what not to do in your writing.
It’s also a great educational tool. Have students generate essays. Then let them fact-check and edit. It’ll teach them how to search and vet information, as well as how to structure an argument.
And, because ChatGPTs job is to be a convincing chat partner, I wonder whether it can play a role in the management of loneliness.
I do not doubt that many business and content creators will jump/have already jumped on ChatGPT as a cheap and easy content generator. If you care about quality, however, about living and breathing prose, you’ll still need a human fact-checker and editor. (As I am finishing up this newsletter, I learn about a tool called GPTZero, which can give you a guess of how likely it is that a piece of text is generated by a large language model by - roughly - estimating how ‘predictable’ the text is.)
Two final, unpleasant notes to keep in mind:
ChatGPT needs labeled training data. That labeling was (is?) done by Kenyan workers who make less than $2/hour. Exploited labor is actually a big issue for most AI systems we are so eager to use, not to mention the massive environmental costs of training and maintaining these systems…
Right now, ChatGPT is free. But it is very costly for OpenAI to keep running. That’s not going to last. It might become a paying service. OpenAI might sell (your?) data. Microsoft might buy it and release a paying subscription. Yay capitalism, amirite?
Have you played around with ChatGPT? What did you think?