If you live and breathe technical marketing content – like those of us here at Fixate – you’ve probably heard more than murmurs over the past year about GPT-3. GPT-3 is a nifty open-source tool that can generate text with stunning accuracy. The idea of replacing human writers (not to mention programmers) with Artificial Intelligence (AI) that can do the job faster and more cheaply has some managers salivating — and some technical writers shaking in their boots.
Will GPT-3 really upend the world of technical writing and marketing? We don’t think so. It’s a useful tool and bound to make a big splash. But we’re not worried about good content writers and editors being automated away anytime soon.
To prove the point, here are five problems that GPT-3 can’t solve for technical marketers. At least not on its own.
Write Totally Accurate Technical Content
Probably the biggest limitation of GPT-3 for technical content marketers is that AI models based on the tool sometimes end up saying things that sound accurate, but aren’t.
For instance, I asked a GPT-3 based AI model in OpenAI’s playground “What are the most important log files to monitor on Ubuntu?” Here is the response:
- auth.log
- syslog
- dmesg
- kern.log
- apt/dpkg.log
- secure
- boot.log
- Xorg.0.log
- faillog
- user.log
If you know about Ubuntu, you’ll know that most of the items on that list are accurate. But it’s a little off in subtle ways. For one, Ubuntu doesn’t have a user.log file; OpenAI appears to have made that up. For another, there is a dpkg.log file on Ubuntu, but it’s not inside a subdirectory named apt, as the list above suggests.
These are examples of technical inaccuracies that would be easy to gloss over if you lack deep domain expertise, but which informed readers will notice and call out. If you entrust your technical content generation entirely to GPT-3, you’re likely to end up publishing information that is just plain wrong in certain respects; although you may never know it until readers start leaving comments that criticize your company for not knowing its stuff.
Generate Truly Original Technical Content
Like most AI tools, GPT-3 works by ingesting vast amounts of existing data and using it to generate new data.
That approach works well if the new content you want to create is fundamentally similar to content that already exists. If you want to bless the world with yet another article about “The top 10 reasons to move to the cloud” or “5 ways to start your digital transformation journey,” GPT-3 will probably be pretty good at writing the article for you, because it can draw on the vast amount of content that already exists on subjects like these.
But if you’re looking to create something more original — say, documentation about your product, or a how-to article about a new tool that no one has written about before — GPT-3 won’t have much (if any) relevant data to work with.
If you want to create original content — which you probably do if you’re hoping people will actually read what you publish, and it won’t be buried on page 12 of the search results — you’ll need human writers.
Explain Complex Technical Topics
Part of the reason folks are excited about GPT-3 is that the tool can write not just prose, but also computer code. That makes it different from generic text-generation software, and it raises the prospect that GPT-3 could write technical articles that include original source code.
The problem, though, is that there is a huge difference between writing code and explaining code. Generating code is cool (although it’s not exactly novel — no-code tools have been around much longer than GPT-3), but to create an article with real value, you need to walk the reader through what the code does, and why it’s written the way it is. You need, in other words, to explain why you used the particular programming strategies and methodologies you did, not just produce raw code.
GPT-3 is not likely to be very good at these tasks – unless, again, it’s regurgitating content that already exists, in which case the content is not likely to be very interesting in the first place.
Self-Curate Technical Marketing Topics
One thing you learn quickly in the world of technical marketing is that no matter how good your content is, few people will care about it if the topic is boring or irrelevant to your target audience.
This is not a problem that GPT-3 can solve totally on its own. AI models based on GPT-3 can generate good-enough topic ideas based on keywords that you feed to them. But you have to come up with relevant keywords, which requires a deep understanding of your marketing goals and audience needs.
Plan Content Formats and Type
Along similar lines, GPT-3 is not going to help you decide whether the content you want to create will be best positioned as, say, a blog post or a whitepaper. That type of decision requires a deep understanding of the type of audience you are trying to reach, as well as the part of the sales journey that the content is designed to support.
Here again, it’s hard to imagine AI automating this work away. There are too many nuances at play to give this job to anything other than a human.
Conclusion: GPT-3 is Innovative, but Not Revolutionary
To be sure, GPT-3 may help technical marketers in some ways. For example, imagine using the tool to write summaries of articles written by engineers. It might also be cool to explore the use of GPT-3 as a type of “smart” copy editor. It could, for example, use context to determine that you actually meant to write “OpenShift” in an article where you wrote “OpenStack.” That’s the type of nuance that only copy editors with specific domain expertise (which are hard to find) will understand. So GPT-3 may be a useful aid.
In other words, we’re in a world where GPT-3 is to technical marketers and writers what AIOps is to IT teams. It is a way to automate mundane tasks so that knowledge workers can focus on more innovative work. But it’s hardly a full replacement for a skilled team of technical content curators, writers and editors.