Chat GPT feels like the dawn of AI and it’s both more mundane and yet more shocking than expected.
More mundane because it’s not doing anything unfamiliar: it’s not teaching us anything new, it’s just doing something very common that Humans can already do: write. More shocking because it’s showing us how producing prose that is grammatically correct and intelligent can be a lot more mechanical a process than we intuitively thought.
Nevertheless, it also feels like a trick. Spend time querying it and you will find that it repeatedly makes claims that are totally false. Sometimes it feels like a really intelligent, articulate student who has just realised they need to write a book review for their homework the next day, but haven’t read the book, so they decide to invent their review. So they have to write something as plausible and intelligent-sounding as possible yet without accurate knowledge or insight into the book itself.
If you don’t already understand a topic deeply, you could be tricked. If societal commentators have been worried in recent years about the spread of ‘fake news’ then this should give them many more sleepless nights!
Also, a lot of it feels like waffle. My (albeit limited) understanding of how it works is that having ingested much of Human textual output from the dawn of time to 2020, it looks for statistical patterns. It produces a few words and then searches for words that, statistically are usually likely to follow the previous words. It then throws in an occasional unlikely phrase, just to keep it from being too predictable. Yet this kind of automatic writing where one word triggers another is also the process that produces cliché in Human writing. So perhaps it’s not surprising that much of it reads like waffle. Yet this sort of process perhaps underpins how we talk and write most of the time, its just that we also have a guiding star of our own experiences, knowledge and goals.
If ChatGPT was a novelist it would be a ‘pantser’ – moving along rapidly by the seat of its pants, without the overall plan of its more sensible twin, the Planner. Pantsers produce successful novels, but I suspect few are truly as spontaneous as they seem. I reckon many are either the beneficiaries of structures that the writer expresses unconsciously or of the post-hoc reworking of the revision and editing process.
Does it pose a threat to Human writers?
The ability to simply churn out competent text is mostly not what we value in writers. We are seeking an interesting Human voice. This could be because they are an expert or because they have some interesting aspect to their life and personality that makes them interesting to read.
More specifically, we are seeking the expression of a Human voice that has been honed through editing, revision and judgment. The cliches I previously mentioned, bedevil much of our everyday speech are like weeds to be hacked away to reveal clearer and more elegant writing. This is the process of editing.
Simply writing a book, for example, is usually not enough to guarantee people will read it. What makes them want to read it is that they feel what the writer is saying is relevant to them. The book has been marketed to draw attention to it and enhance its relevance to readers. It’s not as though we are short of books per se. Even if all publishing ceased tomorrow there are many more than enough books already published to last the most voracious reader a lifetime.
Also, the very best writing is often about exploring the edges of Human intuition. It emerges out of the Human unconscious. Making ideas that we were once only dimly conscious of more concrete.
Nevertheless, there are many jobs writers do that aren’t quite so highfalutin. Large online retailers, for example, need product descriptions written for thousands of products. Some have already been using AI to assist in this process for a few years.
Equally, just giving an extended description of something – a sort of half-way point betweeen a google search and a Wikipedia article – seems like something ChatGPT is okay at (notwithstanding my earlier point about accuracy).
If we need to read something that will quickly impart some simple bit of news or describe a product or what is the general consensus on a particular subject, ChatGPT and it’s successors could be good enough. Particularly if the accuracy improves, or it has some way of admitting when it may be inaccurate.
Yet where ChatGPT and other systems like it might have their greatest impact is in more unexpected areas. Forms of writing that no Human has yet become really good at because it’s hard for our brains to do. For example, summarising lots of new or technical information quickly. Or, for example, imagine a nightly news show is due to air in minutes and a shocking event has just occurred, such as a natural disaster. an AI writer could rapidly summarise the salient responses from thousands of tweets from those on the ground. Something that no Human writer would have time to do. Or, if joined by AI that has greater analytical capabilities, it could ‘read’ hundreds or thousands of scientific papers and quickly synthesise and summarise their findings.
Another new area that it seems very useful for is summarising or simplifying complex text so that it’s possible for those with less knowledge or ability to understand it. One of my favourite sub-reddits is ‘explain like I’m five’, where experts take a complex topic (often something technical or scientific) and explain it in simple terms. AI could be great at that.
Such AI writing systems may also force Human writers to evolve. To become less formulaic. It seems, for example, unlikely that AI writing is going to make readers less interested in good Human writing. But it could very easily make them more interested in it.
Also, it’s worth considering that the real competition will not be Humans Vs AI writing but Humans using AI writing tools vs other Humans using AI writing tools. So the real question is how writers can use such systems to empower what they do.
What makes it hard to predict is that, unlike many things we encounter in life, things that rely on computer power and networks (like the Internet) can grow exponentially. A child improves their skills slowly and incrementally. An AI might double their skills overnight. This means it’s ultimately hard to predict where this is all heading.
It does mean that some of its current weaknesses are likely to be overcome. I would expect much of the inaccuracy will be overcome. Software that has a broader and deeper knowledge base or that at least warns you when it doesn’t know the answer, rather than generate a plausible-sounding lie.
The more salient questions are, I would suggest, how we will make best use of AI writing and how many of its current limitations are fundamentally part of it vs just temporary? It will be interesting to see how this new form of writing evolves.