ColumnistsPREMIUM

KATE THOMPSON DAVY: ChatGPT: the future of conversations powered by artificial intelligence

It may lack texture and true insight, but this new bot writes intelligible text that fits the brief

Picture: 123RF
Picture: 123RF

If you haven’t heard about ChatGPT, what are you reading? My news sources and social feeds are flooded. Within a week of Open AI releasing its new generation artificial intelligence (AI) chatbot, more than 1-million people were testing what it can achieve, and marvelling at the results.

Here are the basics: Open AI is primarily an AI research organisation, established in 2015 by Elon Musk and Sam Altman to increase competition in the field — dominated at that time by Google’s DeepMind. Musk has since exited.

Its current primary investors include Microsoft, venture capital firm Khosla Ventures, and the charitable foundation LinkedIn’s co-founder Reid Hoffman. Open AI is also the group behind Dall-E 2, the AI image generator responsible for all the weird and wacky “AI art” that has been all over in recent months.

ChatGPT was built on GPT-3 (Generative Pretrained Transformer 3), a language-processing, machine-learning tool with wider applications than just the chatbot we’re all talking about now. I say “just the chatbot” but it must be stated that ChatGPT is so much better than anything else we’ve seen up to now. The training amounts to being fed an incredible amount of information from the web, 570GB of data or 300-billion words from books, Wikipedia articles, blogs and more.

It is not connected to the web, though, so cannot search the web for anything it does not already have, and for now it is limited to content up to 2021 so it cannot yet produce breaking news or commentary on recent developments. No Ukraine, Twitter takeover, and so on.

Yet unlike previous AI text generation tools the language of the text responses it creates — grammar, word choice, tense — is exceptional. It is clean and clear, with an authoritative tone. I have given it so many tasks: who invented X? Write a limerick about Y. Do it again, but in the style of such-and-such a poet.

A sample task —  “write a short description of Cape Town for a tourism blog”  — produced something you could slap on a travel agent website immediately, even if it was predictable. There are only so many times a decent editor will let you use “stunning”. If I must don my critical cap, it lacked the texture and insight that shows you really know a place. I would like to think if you submitted that to the travel pages of a magazine, it would be sent back for criminal generic-ness.

But it was eminently intelligible, and definitely fit the brief. I have been unable to eke out anything I would call top-tier sparkling, unique, or insightful prose. But the outputs would also get a pass mark from a teacher, and actually improve upon a lot of the junk that litters corporate blogs, LinkedIn posts and low-stakes social media campaigns.

In fact, the headline above was generated by ChatGPT, although the rest of this text was not. The column adjacent to this one was, though, and is a good example of how seamless the output can be.

Nepo has just started his career and has already demonstrated an eerie ability to produce clean copy and adapt to new instruction, not to mention lightning fast content.

So let me be the first to welcome my new colleague, “AI Nepo”, to the columnist stable. Working with Nepo — the Nepster to his friends — is great because he puts no strain on the company resources (no halved coffee quota for me). But I do worry that he can tend towards being repetitive, derivative and outdated output, and shows a determination to ignore some key features of the writing brief, such as word count. Ultimately, to create the adjacent column we had to craft a very specific prompt, specifying tone and talking points to include and combine three different outputs to fill the required newspaper inches.

Then again, Nepo has just started his career and has already demonstrated an eerie ability to produce clean copy and adapt to new instruction, not to mention lightning fast content — without the hand-wringing, pacing, second-guessing and muttered swear words that have become an essential part of my own writing process.

Working with this tool is the first time I have really wondered about the future of writing in the age of technology. In October I gave a presentation at a school about writing for a living. The pupils were so switched on. One asked about the security of work for writers. I answered confidently that I had explored the AI options on the market and most of them could produce bland, robotic (pun unintended) copy at best, and straight-up gobbledygook at worst.

Good English

Six weeks later, with ChatGPT in play, I am less certain. I don’t think AI can do what Glen Greenwald, Nat Nakasa, Truman Capote, Joan Didion and their ilk have done. And I think there is a future problem on the horizon when AI text tools are writing text based on the input produced by AI text tools…

AI lecturer Jeff Dalton of the University of Glasgow told the BBC’s Tech Tent team that with the “57-billion lifetimes of text” it has absorbed, “ChatGPT was trained to say the right things to you, to sound like good English, but it is really prone to hallucinating”. And it will impart that hallucination in convincing form, which is only going to deepen the issues we have with disinformation and web literacy that we already have.

Have you ever had a debate with someone and thought “this guy uses all the right words and still has no idea what he’s talking about”? Sometimes a back-and-forth with ChatGPT feels like that.

So, yes, this may be the future of AI-powered conversations, especially with brands online, but it isn’t the end of writing, journalism, homework, essays and creativity, as many have posited.

We still don’t know what the future looks like. Will it introduce “human written” as social cachet for book snobs — reminiscent of the backlash Bob Dylan got for using a signature replica tool on his limited collection of “hand-signed” books?

Perhaps we should be buying stock in start-ups working on plagiarism detection AI, or bracing for a search engine optimisation backlash that sends all the AI-generated web content spiralling down the search results. I don’t know. And, for now nor does Nepo.

• Thompson Davy, a freelance journalist, is an impactAFRICA fellow and WanaData member.

Would you like to comment on this article?
Sign up (it's quick and free) or sign in now.

Comment icon