From malicious space operating systems to murderous terminators, there’s always been a fear of AI robots taking over. In fairness, AI has already hit some industries hard, particularly within manufacturing. (Read: Top 12 AI Use Cases: Artificial Intelligence in FinTech.)
For now, it’s mostly the manual labour that AI has been able to replace. And so, this created a safety blanket for writers. They’ll be alright — they’re creative after all and AI cannot replace that.
Outside of fiction, you can see the panic in headlines such as:
- “Robot automation will take 800 million jobs by 2030″
- “Over 30 million U.S. workers will lose their jobs because of AI”
- “AI warning: Machines will replace humans at the top – and wipe us out if we dare resist”
Dark tidings indeed. Then, news broke of OpenAI’s GPT-2. This is a text-generating AI that’s so good the company deemed it “too dangerous to release”. Publicity stunt or not, this has changed the question from: will AI replace writers; to when will AI replace writers?
From Clunky Text Machines to Intelligent Prose
Last year, Adzooma tested a few of the AI content generators on the market. Generally, the results weren’t good. The content that came out was full of gibberish that lacked a purpose or understanding of what it was about. It might have been original and about your topic so to speak, but it wasn’t publishable.
Things have changed since then. In just 5 months, Adzooma probably would not get the same results from that test. The world of AI grows quickly. The more time you wait, the more likely you’re going to be left behind. And it might be GPT-2 that leads the way.
Is GPT-2 Really that Good?
Well, we will never know about the version of GPT-2 that OpenAI didn’t release. That’s still top secret. However, they did release a less dangerous GPT-2 tool. And frankly, the results are astonishingly good.
For a bit of context, GPT-2 is designed to generate text based on the previous words before it. You can’t give it a theme and set it going, you need to fill it with information first.
This is the explanation from OpenAI’s website:
“GPT-2 is a large transformer-based language model with 1.5 billion parameters, trained on a dataset of 8 million web pages. GPT-2 is trained with a simple objective: predict the next word, given all of the previous words within some text”.
One of the best ways to see it in action is in The New Yorker’s article Can A Machine Learn To Write For The New Yorker? At various points in the article, they will stop and let you see what GPT-2 wrote, given all the text supplied so far.
This was one of some of the text generated:
“OpenAI, an artificial-intelligence company, announced that the release of the full version of its A.I. writer, called GPT-2—a kind of supercharged version of Smart Compose—would be delayed, because the machine was too good at writing.”
Yeah, that’s from GPT-2.
Personally, I wouldn’t write like that. But that content is right and in line with the style from The New Yorker. If I read that with no context, I wouldn’t be able to spot that it was different straight away. Looking at pieces of writing like that, you can almost feel the clock ticking down until AIs are doing our jobs for us.
On Second Glance…
There are a few flaws that we need to mention with GPT-2.
Firstly, it’s not always accurate. The text it generates might sound good, but it has a tendency to make up quotes and names. Language expert Sarah Gregory said: “The data fed into the AI would have shown that quotes are often used so would have provided them to make the text more natural”.
If people believe quotes like these as authentic, then you’ve got a fake news situation on your hands. And yes, I did make Sarah up to prove this point.
Secondly, the text still lacks purpose. It can follow on perfectly well from the text in front. It makes grammatical sense and it sounds like it could be right. But there’s often a feeling of something being off with it.
John Naughton put it best: “At this point, the reader gets the eerie uncanny valley feeling: this is almost, but not quite, authentic. But the technology is getting there.”
Words aren’t there for the sake of it. They’re there for a reason and to achieve a particular goal. The words generated by GPT-2 have a different purpose: to blend in with the existing text. And that difference means it just doesn’t quite fit.
Using to Aid, Not to Replace
AI isn’t perfect, but that doesn’t mean it’s useless.
GPT-2 can be a useful tool for writers. When testing the tool, Patrick House wrote that this became a valuable tool he utilised daily, using it as a sounding board to be “a curator to new ideas and a splint for bad ones”.
Editing can be a lot easier than writing sometimes. And if GPT-2 is giving you something to mould into perfection rather than starting from scratch, you’ve saved yourself a lot of time.
It could even help if you’re just suffering from writer’s block. Put in what you have so far, see what it’s coming back with and then use it to power your own work.
So, Are Writers Safe?
For now, writers are indeed safe. But things change fast. This software is leaps and bounds above the rest in its category. Over time, people will make it better. In another decade, it might be writing articles like this for me.
There are some industries that do utilise AI writers already. The Associated Press (AP) uses machine learning (ML) to automatically create write up posts of minor league games. They don’t have reporters to get to every game, so they feed the game data into their AI to create a write-up of what happened.
You have to remember that nothing will ever be 100% taken over. AI will have its part to play, but in most cases it will help streamline tedious parts of your work, leaving you free to focus on something else. (Read: Top 20 AI Use Cases: Artificial Intelligence in Healthcare.)
We adapt. We work with AI. And we should learn to do it fast, because it will come quicker than we think.