Skip to Main Content

How I tried (and failed) to write a press release with AI and machine learning

Posted by Jorden Dakin-White on 9th January 2018

As some PR professionals sharpen their axes and light the proverbial torches for a long-drawn-out protest against AI entering the workplace, others stand by the idea that AI is like the God in the heavens who will anoint their business with the righteousness of the Lord and save their struggling company from the damnation of hell. But, few ask or consider “how can AI help me do my job better right now?”

For the duration of this article we adopt IBM’s definition of AI as “augmented intelligence” instead of “artificial intelligence”, where augmented intelligence is an imposed and often logically defined set of rules for a program to follow that does not emulate human cognition, because it lacks the ability to simultaneously run a problem through an emotional pathway or program.

Now, as long as you haven’t been living under a rock this week you will know many articles have surfaced on how AI is poised to take over the mundane tasks in our jobs. For journalists, this could be writing Associated Press-style statistics-based articles using AI with almost instant publishing, while for the PR industry this could be automating specific aspects of the business or using AI to predict workflows or even write press releases.

At this stage of AI development, it’s important to remember AI can only add value, and it has a limited breadth of capacity and ability.

Some of you may have seen the articles discussing the fantastic failure of AI to write a chapter of Harry Potter, and so it dawned on me, what if we could get an AI algorithm and machine learning program to learn how to write a press release? What would it look like, given that I only have a basic — and I mean bottom-level — understanding of coding?

Here is a quick synopsis of what I did: gathered text from the latest 50 consumer tech press releases from PR Newswire written in English, installed the AI code and TensorFlow, input the sample press release text into the correct file, executed the machine learning script, ran the AI program, and then grabbed the output text of 1,500 words.

The results are appalling, proving augmented intelligence and machine learning still have a long way to go. Here are the top three things I learned (sample text is at the end).

Get a large data sample

Even using outdated predictive AI and machine learning, one thing still holds true: AI and machine learning programs are only as good as the volume of data you can put in it. Since I have a relatively small hard drive on my laptop, less than a terabyte, I was limited to the amount of data my HDD could hold and still execute the processes.

To have a good sample size, try to have at least 500,000 data points, words in this case, and if you can get 1 million you should be in a lot better shape and have a far more readable and reliable output.

Find a better use case

Using predictive AI to write a press release probably isn’t the best use for this technology. Perhaps using something more like what the AP uses for writing statistics- and data-based articles via a template and machine learning designed to pick out the right details could work better.

Or better yet, use AI and machine learning to scour the internet for brand mentions and tonality of coverage similar to the system Meltwater uses.

AI still needs oversight

If anything, this exercise proves an AI program alone can’t write a press release. And AI still has a long way to go before it can do my job better — which is good news for communications professionals.

Even if the AI spit out a beautifully worded press release, chances are it would still need to be copy edited and proofed for errors, because there are slight differences in languages and terms by region — just look at how usage and spelling differs between American English, British English, and Australian English. While any human person could tell the difference, a computer would have no idea.

Looks like our jobs are safe — for now.

Check out our most recent AI experiment and find how good the technology is at writing blog posts…

Sources: Forbes, Medium: Deep Writing, Venture Beat, The Guardian, IBM

Jorden Dakin-White

Jorden joined Wildfire in 2017, bringing with her a mix of knowledge across various sectors of PR, including B2B technology, consumer tech, consumer health and beauty, and food. Jorden hails from the United States and before embarking on a career in PR she worked as a freelance writer and for her university paper. Not content with just working full time, Jorden is currently completing an MSc in Communications from Syracuse University in New York and learning to code in Python 3. When Jorden returns to Seattle to visit friends and family, she can be found sipping coffee and shopping for clothes, makeup, books or , even more dangerously, browsing for new video games. A self-confessed techie and gamer, Jorden’s always on the look-out for a new video game — suggestions on a postcard please!