Is Generative AI coming for your job, or is it a bubble?

I’ve spent a fair amount of time over the past few months trying out different practical applications of Generative AI as well as trying to separate reality from hype. It’s been an enlightening learning experience with many “Oh, wow” moments, but also a lot of “Is that it?” results. On balance, I’m a lot less worried about mass unemployment and Skynet becoming sentient than I was at the start of the year.

It’s less than a year since the hype cycle around Generative AI went into overdrive as ChatGPT and other models started to gain widespread use. Predictions about future impact have ranged from dystopian to utopian. Being old enough to have been through the dotcom bubble, rise of the mobile internet and smartphones, the emergence of social media and crypto, the noise around GenAI is sounding very familiar.

AI is a broad concept that encompasses Generative AI - systems like ChatGPT that are able to generate new content based on natural language input. These systems are underpinned by Large Language Models (LLMs) that have been trained on staggering amounts of data, predominantly scraped from the Internet. The more training data you feed into an LLM with, the better it will perform. Because textual data is so abundantly available on the Internet, text-to-text GPTs have rapidly developed impressive capabilities.

The scraped text is divided into chunks that are given a context. When a user provides a prompt, the GPT uses neural networks to string together chunks with relevant context to create an answer. It does this by selecting the most likely next chunk in the chain based on probability derived from the learning data. That is obviously an absurdly simplified description of a very complex process, but the point is that the answer is a probabilistic concatenation of chunks rather than the result of a thinking process.

If you have spent any time generating text content, you’ve probably been impressed by the initial output based on a fairly simple prompt. A 1,500 word blog article in a split second looks close to magic. The issue with GPT output tends to be that it sounds generic. I think that is an inescapable result of how LLMs work. They give you the most likely output based on the data they were trained on, so what you end up with is an average of relevant scraped data. It’s OK, but it’s not new in any real sense.

With the expanding capabilities of text-to-text GPTs, what does the future hold for jobs in creative writing? I think pure content creation can be handled by a GPT, for example if you need to generate SEO-optimised text for a specific purpose. But if you are looking for content with a novel point or view or a tone-of-voice that is brand specific, you will still need a copywriter. Sure, you can improve your prompts or tweak the GPT in any number of ways, but you will still get better creative content from a human. I expect that a lot of writers will be using a GPT as a co-pilot, but I don’t see GPTs replacing copy writers.

Another group of professionals that have been warned of obsolescence are software developers. GPTs that generate code will soon replace developers according to the hype. Reality is a long way away from that. Broad GPTs like ChatGPT or Gemini generate code of questionable quality that requires extensive review. The reason is likely that the volume of quality code available to train LLMs isn’t sufficient to deliver consistent quality output. If you need to prompt the GPT and verify the output, you still have a developer with a job, only using different tools.

Is there the same irrational exuberance regarding AI today that we have seen in other bubbles? Absolutely, and several of the emerging vendors will be the boo.com, pets.com and MySpace of the AI bubble. That doesn’t mean that AI won’t have tremendous long-term impact on society, I’m convinced it will, but it will likely be in ways we haven’t even conceived of yet. I refuse to believe that the best we can achieve using AI are better ways to churn out blog articles and TikTok videos. That would show a depressing lack of intelligence, artificial or otherwise.