<img height="1" width="1" style="display:none" src="https://www.facebook.com/tr?id=2006193252832260&amp;ev=PageView&amp;noscript=1">

7 Min Read

What Marketers Need to Know About AI Content Generation Today

Featured Image

Wondering how to get started with AI? Take our on-demand Piloting AI for Marketers Series.

Learn More

Editor's Note: This is a sponsored blog post from Marketing AI Institute partner MarketMuse.

For many content marketers, artificial intelligence means leverage. It’s a way to do the things they’re already doing manually—content research, competitive analysis, content gap analysis, content optimization, etc.—in less time and with higher quality output. 

AI is a force multiplier across the content production process, but one part that has never quite gained traction is in the content writing process. The reason? AI technology just hadn’t reached a point where it could be reliably used for generating high-quality content. 

Recent developments, however, have changed things. AI content generation is a reality, and it’s starting to gain a place in the tech stacks of content marketing teams that want to apply the leverage AI gives them in other parts of their content process to the actual writing itself. 

Why Are Marketers Excited About AI Content Generation?

Everyone wants to rank on page 1 of Google (or the search engine of their choice) for the keywords and topics they care about. But ranking well isn’t always the same as writing content that makes people say, “these people really know their stuff!” Content marketing teams need to thread the needle to produce content that engages their target audience, facilitates customer journeys, and drives meaningful traffic from search engines. 

Over the last few years, AI solutions have sprung up to help content marketers tackle tedious tasks, such as keyword research, content optimization, building content briefs, and so on. These solutions have freed up countless hours for content and SEO teams, but actually writing quality content remains a task for humans alone. 

Writing is fundamentally creative, and not even the most advanced AI today can act as a one-to-one replacement for even a writer of average ability. But AI content generation is at a point where it can even automate some of the tedious work that writing for digital content, and search-focused content, in particular, can involve.

Ranking well in search and getting traffic from qualified users who are actively seeking information on your core topics is mostly a matter of how well your content matches the searcher’s intent. In practice, this means that most writing for search involves reading through several of the top-ranking pages on a term you want to rank for and incorporating similar information into your own article.

Many writers have to read multiple documents to understand what they need to cover in an article to match the user intent of the term or keyword they want to rank for. Then, they need to use that research to write a draft that elegantly weaves the related topics and structures that similar content use. From there, they need to edit it, layering in their own expertise, ensuring it’s brand-compliant, and aligning it to their broader content strategy.

Where AI can come in is in that first step in producing written content. AI can analyze a topic, understand the related topics that need to be mentioned to comprehensively cover that core topic, and then generate a usable draft for a writer or editor to hone it into a publication-ready piece. 

It’s moving the starting line of the writing up from a blank page to a draft that is comprehensive and ready for human creativity.  

But not all AI-generated content is equal, and understanding the solutions out there will be critical for content teams looking to bring AI into their creative processes. 

What GPT-3 and Other AI Content Generation Solutions Can’t Do 

GPT-3 is all the rage these days. All over traditional and social media, there are boundless examples of early use cases people are discovering with GPT-3. 

It’s hard to deny that the things people are doing with GPT-3 are fascinating. Still, if you’re looking for a plug-and-play AI content generation solution to produce usable drafts of content, you’ll be disappointed. Even OpenAI’s CEO Sam Altman tweeted that “The GPT-3 hype is way too much.”

The Guardian recently published an article titled “A robot wrote this entire article. Are you scared yet, human?” The article’s premise is apparent from the title: GPT-3 wrote an entire article worthy of publication in a major new outlet.

Yet, the pushback was immediate. AI experts who read the fine print in the Guardian article realized that the outlet was contributing to the hype more than illuminating the true capabilities of GPT-3 as it stands today.

Experts pointed out that the Guardian essay was actually eight different 500-word essays stitched together by editors, who picked the most usable parts from each piece. Out of 4,000 total words, the editors could salvage 500 of them for the final essay. That means, on average, each of the original eight articles contained about 60 words - or 12% of the total words - of usable content.

This is very different than saying a robot wrote the whole article as is.

The fact is that GPT-3, while impressive, is not equipped to write logical articles. As Technology Review observed, “although its output is grammatical, and even impressively idiomatic, its comprehension of the world is often seriously off.”

Essentially, GPT-3 can take a prompt and generate some seemingly coherent text. Still, when you actually read it, it’s clear that it’s just words strung together, creating a tapestry of some truths, some falsehoods, and a lot of nonsense. This is due to the way it’s built.

GPT-3 uses unfiltered data from Common Crawl, Wikipedia, and other sources like social media platforms, forums, site comments sections, and even other AI-generated text. Its creators were not very selective about the type or quality of data used for training the model. Well-written and edited articles represent about 3% of content on the web. That means only 3% of the training data for GPT-3 consists of high-caliber articles. 

Essentially, if your goal for your content is to join the 3% of well-written web articles, you shouldn’t look to GPT-3 as a solution.

The limits of GPT-3 as a means of reliable content generation are clear. If you’re looking for an AI content generation solution, GPT-3 is more of a curiosity, not a serious addition to an AI tool stack. 

MarketMuse First Draft and AI Content Generation

GPT-3, from a content creation perspective, is still a solution in search of a problem. For content teams that need to increase their content publication cadence, create customer journeys, and build greater efficiency into the content production process, GPT-3 is not the answer. 

In fairness, GPT-3 was not specifically constructed for that purpose. 

While the model boasts 175 billion parameters, its training dataset does not filter out low-quality content, user-generated content (forum posts, social media posts, etc.), offensive content, and so on. Anyone who has used the internet for a while knows that there’s a lot of that out there! Not the kind of thing you want creeping into your content marketing efforts.

MarketMuse First Draft is an AI content generation solution that, like GPT-3, uses natural language generation to produce written content. The difference between MarketMuse First Draft and GPT-3 is that First Draft was explicitly built for content teams that need to publish high-quality content and increase their content publication cadence. It’s not an academic effort, but a practical application of natural language generation for content creation. 

Three core factors go into making MarketMuse First Draft a powerful AI content generator.

The Curated Dataset

At this point, the First Draft dataset is constructed by collecting a few thousand well-structured articles on a specific subject. Just like the data used for the base model training, these need to pass through all our quality filters. 

The articles are analyzed to extract the title, subsections, and related topics for each subsection. The data feeds back into the training model for another phase of training. This takes the model from a state of being able to generally talk about a subject, to talking more or less like a subject-matter expert.

The Generation Process

First Draft generations are built off of MarketMuse Content Briefs. These are detailed outlines that provide titles, subtitles, and related topics that each section of the article should include. The related topics are based on a topic model. 

The topic modeling technology underlying MarketMuse allows a user to enter any topic and have AI analyze thousands of documents on that topic. From there, it extracts the related topics from each document, analyzes and sorts them by relevance to the core topic, and shows you what it means to be “about” that core topic.

Because the content is generated based on a topic model, and you control which topic models to include in your article, you can tightly control the content generation. Instead of working from a single paragraph prompt like GPT-3, First Draft writes each section of the article based on the topic model. This is how the AI avoids going off-topic and gives you usable content ready to be shaped into a finished piece. 

No rambling articles, no irrelevant or offensive content that needs to be completely taken out, no Frankenstein-ing multiple generations together to get something readable.

Also, First Draft’s model can be trained to emulate your tone and style or that of a publication you like. 

The Accessibility

Want to use GPT-3? You can sign up for a waitlist, and then once you get in, you can have a trained developer access the API and help you with content generation. For a price, of course. 

How many content teams have a developer at their beck and call? Not many.

Not only do you have to pay to access the API, but you also have to pay writers and editors to shape the output into usable content. Given how long it would take to do what the Guardian did and use multiple drafts to get one decent piece of content, it’s hard to see this as a better option than writing from scratch.

What makes First Draft even more compelling as an addition to a content marketing tech stack is that it’s accessible from a usage and cost perspective. You don’t need a developer to have your own generations run, and you don’t need to join a waitlist to get API access. 

Once you order or build your own MarketMuse Content Brief, you have everything you need to order a First Draft. First Drafts, like Content Briefs, have a set cost, so you can keep your content production costs under control and keep them predictable. 

With AI content generation, the content writing process doesn’t start with a blank page. It gives the content writer that first step in building a strong first draft, which can then be honed with the writer’s expertise and editorial judgment. 

Fully automated content creation is still years away if it ever truly arrives. Right now, AI technology can bring massive leverage to the content creation process. 

But it still can’t replace the human creativity needed to point all of this technology in the right direction. It can’t understand your specific customer needs the way a dedicated content marketing team would. It can’t replace the expertise and viewpoints of people who have worked in your industry for years and have insider knowledge that AI can’t possibly attain yet.

MarketMuse Logo-1

MarketMuse is AI-powered content planning and optimization software. Click here to learn more.

Related Posts

What's the Difference Between Natural Language Generation and Article Spinning?

Stephen Jeske | September 25, 2020

Learn the difference between natural language generation and article spinning and why you should avoid article spinners at all cost.

Why AI-Generated Content Won't Break the Web

Stephen Jeske | October 28, 2020

See a case study on an AI-generated blog, the predicted performance, and the surprising results.

5 Valuable Benefits of Natural Language Generation

Stephen Jeske | October 19, 2020

Natural language generation offers a number of benefits to those organizations willing to take the lead and employ this cutting-edge technology.