Are machines good enough to write your content yet? Here’s what we discovered about the quality of machine learning in content creation.
Technology is rapidly changing our day-to-day lives and that’s impacting how we behave, both as consumers and as businesses.
Take Amazon as an example. If you’ve logged in to Amazon recently you may have noticed sections entitled ‘Related to items you’ve viewed recently’ or ‘Inspired by your shopping trends’. That’s a perfect example of machine learning making personalised suggestions based on your past behaviour.
Spotify does it too. They’ll make song and artist suggestions based on tracks that you’ve listened to previously.
But, is that as far as the technology can go?
I’m sure we’ve all been staring at a blank screen, fully into writer’s block and wishing that there was something that uses machine learning to write the words for us.
In the case of content creation, machine learning should work like this:
- We give it data.
- The program analyses this data and “learns” it, using it to predict what to write.
- We get content!
Note the emphasis on should there. The point is that machine learning is growing at an incredible rate, but we don’t know if it’s quite there yet.
To see what’s out there, we tested some of the available options on the market to see how good they are – and whether the content they made was good enough to publish or not.
What is machine learning?
Machine learning is the science of computers improving themselves through repeated tasks and algorithms. It’s one of the most well-known forms of AI and powers a lot of the systems used in digital marketing. A significant benefit of machine learning is allowing systems to “learn” from its data, to find patterns, and make better decisions that humans may have missed.
However, it’s important to note that the data it learns from is dependent on who inputs it and what the data contains. If you input biased data into a system, the output will be biased too. Bias in machine learning is something data scientists attempt to remove to avoid harm society.
No – machine learning is not a threat
Machine learning and AI as a whole get mixed responses. Some people don’t trust it, preferring to rely on humans, who obviously never make a mistake.
Others are threatened by it. They think that if a robot can do their job better and more accurately than them, then what is the use for them now?
And this clearly isn’t helped by headlines like ‘technology has already taken over 90% of the jobs humans used to do’. But you can’t just skim the headlines. When you look deeper, you’ll actually start to notice the jobs that technology has taken over.
Firstly, there are things that we often forget. Take NASA for example. In the days before calculators were invented, they employed people to do the maths by hand. This is how they calculated space trajectories – yet today it would be unheard of to not trust a machine to help out.
It’s more accurate. It’s faster. It’s how we’ve gone from the moon to landing rovers on other planets.
The point is: technology and AI may have closed some doors. But they’re opening others that didn’t exist before. 20 years ago you wouldn’t have dreamed of someone’s job role to be creating online videos on YouTube.
Things change. It’s up to you to embrace it and get on the winning side of AI. Done right, it can simplify tasks, remove inaccuracies and save you time every single week.
Content generators hitting it out of the park
There are some cases that this works a treat. A great example of this is The Associated Press (AP) use of machine learning to automatically create write up posts of minor league games.
How? With data. It uses game statistics from official minor-league statisticians at MLB Advanced Media and Natural Language Generation (NLG) technology to create legible posts that accurately run through the game.
This is great for building local interest and gives them more coverage – as these games were too minor to typically be covered by a real-life human journalist.
You can read some on AP, or the Minor League Baseball website.
These posts are factual. They’re accurate. They do the job they were supposed to. But they lack one critical thing that we might never be able to replace: creativity and emotion.
So, can machine learning create awesome content?
Here’s what we found using these 4 different machine learning generators:
Talk to Transformer
Talk to Transformer is a free content generator tool that tried to finish a piece of text from a snippet.
‘One of the hardest digital marketing challenges small businesses face is getting noticed online.’
Hitting go, this is what we got back:
And yeah, this isn’t publishable.
This story (obviously) never happened, and even though the words sound right, they make no sense.
I mean, in this story, something wasn’t working, so they took their entire website down. Bit of an overreaction, but okay. The consequence of that? Their website might be taken down.
Although saying that, I love the sentence “if you don’t know your market and your industry, you will probably not know much about it”. Brilliant.
This tool is nowhere near ready to replace actual humans yet.
But, it can be used to spark ideas for you to write yourself. Just don’t spend hours using the bot to create fake stories. That’s not saving you time, that’s procrastinating.
Article Generator is a little more in-depth. You supply the keyword you want your content to rank for and select options from the drop-down menu.
Once you’re ready, it will generate complete, fully-fledged articles. This surprisingly also included images to illustrate the point it was making.
For this test, we went for a generic and broad keyword of SEO. The three articles we got back were:
- The Business of Enterprise SEO: Mastering People, Process & Platforms
- Video: Rand Fishkin on the early days of SEO and starting and leaving Moz
Generally, these sound like regular topics you would see on a blog. The third one about Rand Fishkin is more specific and works more as a PR piece, rather than generic content about SEO.
It was an odd coincidence that we got his name supplied back after we just discussed Rand’s fictional collaboration album with Beyoncé [PDF]. Confused? Keep reading to the end and you’ll totally get why this was on our minds.
Ignoring this one, for now, we had a closer look at the other two.
Generally, they all made sense. It included stats, which was incredible. Such as “according to Google Consumer Insights, 84% of Americans are shopping in any given 48-hour period, in up to six different categories.”
Unfortunately, the writing wasn’t perfect.
Some parts are a bit clunky that don’t flow. It’s like someone copied and pasted bits and pieces from all over the internet without any real thought about the article as a whole.
Which is exactly what it is. There’s an option to create unique content on the homepage, but you’ll need to pay for this option. But considering what we got from the free version, we wouldn’t rush to get our wallets out just yet.
There’s no real clear goal or point from each article. They give you stats and figures and pieces of information. But, they don’t tell you what it means, why it’s important or what you can do with it.
This tool is good for generating content ideas and making a start with your posts. But it shouldn’t be used as the final picture.
The information needs to be understood and rewritten in a way to give value to your customers. It needs to be something they will actually want to read, rather than just a cut and pasted piece of text that does nothing for your business.
Next up on the list is Articoolo.
To use this, you need to first sign in and register for an account. Once you do, you’re prompted to describe the topic you want your article on and choose a maximum word count. This only goes up to 500 words though, so it can’t be used for longer form blogs.
You only have 2-5 words to put into the box, so you can’t get too specific either. For this test, we used “Small business SEO”.
It also takes quite a while to generate. But strangely, this filled us with more confidence. After all, good things are supposed to come to those who wait, right?
Once it’s ready, it will give you snippets of what it’s created. But to unlock it, you need to pay.
To see if it was worth it or not, we unlocked this test article.
Unfortunately, the results were incredibly disappointing. It was hard to read, difficult to follow and didn’t make a great deal of sense. In fact, we put into the Hemingway app to test readability, and the whole thing was lit up in red.
That’s a really bad sign. An ‘avoid at all costs’ type of warning signal there.
Nothing in the writing works together. There’s one sentence that promises points below that will “help out with fostering the user experience.”
But does it deliver them? No.
If you want further proof that this article is unusable, just take a look at this conclusion sentence:
“The visitors who’re intrigued in your products or services, where you must grab an opportunity to get your website optimised at reasonable prices, making it anchors for your requirements and needs.”
It’s not worth paying for. The writing is substandard and certainly not good enough to publish.
On the plus side, the website did promise 100% unique content. So, we also ran it through a plagiarism checker just to make sure. And yeah, it held up.
It’s just a shame that the writing didn’t do so well.
Next up on the testing front was Article Forge, which requires a paid monthly subscription to use. It doesn’t just create content either. It also integrates with your WordPress blog, allowing you to schedule and publish right from the same place.
As a paid platform, the options when creating an article are much more advanced and in-depth. Giving you the ability to really customise what content you need before it’s generated.
Unfortunately, some of these options require you to connect to different platforms. Such as the ‘Guarantee article uniqueness’, which requires a Copyscape login.
After playing with a few of the settings, we tested this using SEO’ for the main keyword, following the same route as others.
For sub-keywords, we used broad terms like ‘small business’, ‘website’ and ‘eCommerce’.
At $47 a month, Article Forge was the most expensive option that we tested. And a great reminder that paying more doesn’t guarantee great quality.
Like the other platforms we tested, the content was a letdown.
It made sense. But it lacks structure and value and any real engagement to make it an interesting read.
Halfway through, it mentions a search engine DuckDuckGo, with no context. In a section titled ‘The 5-Minute Rule for SEO’, it changes the focus to Atlanta businesses for no reason. And more importantly, it doesn’t actually give you a 5-minute rule for SEO.
There were stand-alone moments that were good. Like explaining that SEO stands for Search Engine Optimisation at the start of a sentence. Unfortunately, this came after the 8th use of the word, making it too late to make a difference to this article.
And again, it passed the plagiarism checks, but that’s just not enough.
Don’t get fooled into thinking that just because you need to pay for this service, that it’s any good at writing content. You’d be better off with user-generated content if you were looking for an external solution.
Like the others, it’s got a long way to go before you can use it to generate articles that are ready to be published.
One thing we did like about it was the headlines and the ideas that you can generate using a main and sub-keywords. But it’s a lot to pay per month just for ideas, so we’d avoid this for now.
Can robots take over content creation?
From our experiments, we’ve determined that these content generators are useful in places that are based on plain stats and data, such as baseball. But content-wise? They’re not there yet.
They’re lacking personality. They’re lacking emotion and spirit and that one, clear vision that makes content truly great.
Until we get feeling robots with fully loaded personalities, we don’t think you’ll be able to replace content writing any time soon.
But, that doesn’t mean you need to avoid them. Used right, these tools can be pretty good for generating ideas when you’re stuck staring at a blank page.
We’re not the only ones who have been questioning what machine learning can do for us.
Earlier this year, our team attended SearchLeeds to see Britney Muller’s talk on Machine Learning for SEO. She used a Shakespeare generator tool, feeding it Beyoncé albums and SEO articles by Moz’s Rand Fishkin.
The result was his collaboration album (we warned you about this earlier!), which didn’t really pan out in a way that makes sense.
To be fair to her, this was putting data from two widely different sources. It was never going to be perfect. But it’s a good reminder that machine learning needs data. The more data you feed it, the smarter it gets.
It’s the same logic that our very own machine learning advertising platform uses. We call it the Opportunity Engine. Analysing your account 24/7, it uses data to suggest smart opportunities to improve the performance of your campaigns.
That’s not all though. Our powerful automation feature is also built to take hours off your workweek by eliminating manual, repetitive tasks. This means you have more time to focus on the bigger picture.
Want to see how it works? Try Adzooma for free today.
Further resources about machine learning
- Machine Learning Basics (video) by Simplilearn
- This is how AI bias really happens—and why it’s so hard to fix by MIT Technology Review
- A Tour of Machine Learning Algorithms by Machine Learning Mastery
- Machine Learning: What it is and why it matters by SAS UK
- How to Learn Machine Learning, The Self-Starter Way by Elite Data Science
- Use of AI in online content moderation [PDF] by Cambridge Consultants (2.7MB)
- How to Achieve Personalization at Scale with Machine Learning (video) by Raj Ramesh
- Machine Learning in Action for Search Engine Optimization by Andrea Volpini