AI content, or ContentAI, is the application of artificial intelligence to the broad field of content creation and editing. It involves machines used to produce or enhance text creation. Artificial Intelligence promises to provide insights, deliver predictions, make recommendations, create new content and streamline communication with consumers, but is this already being delivered?
Let’s look at the important breakthrough from OpenAI, known as GPT-3 (General Pre-Trained Transformer). GPT-3 was created by OpenAI, a San Francisco AI research and deployment company founded by Elon Musk, Sam Altman and other high-profile heavyweights in 2015.
As you may have guessed, GPT-3 is the third generation of the Generative Pre-Trainer Transformer. New iterations will surely come, showcasing ever more powerful engines and broader training sets.
How GPT-3 From OpenAI Changed the AI Landscape
OpenAI trained the GPT-3 engine in 2019 by feeding it information from millions of websites, books and other resources. In technical terms, it is an autoregressive language model that uses deep learning to produce human-like text. Think of it as a supercharged version of your phone’s autocomplete feature.
GPT-3 can be used to:
- Generate content snippets
- Disambiguate pronouns
It can even accomplish some basic common-sense reasoning and arithmetic. GPT-3 is a significant breakthrough because it contains 175 billion parameters in over 96 layers, giving its neural network unprecedented power. Training the model cost over $4,600,000.
GPT-3 is not alone in this space. Other contenders are making rapid progress, such as:
- BERT (Google)
- XLNet (Google/CMU)
- RoBERTa (Facebook)
- DistilBERT (HuggingFace)
- CTRL (Salesforce)
- ALBERT (Google)
- Megatron (Nvidia)
Ruder.io keeps an up-to-date comparison of models. You can also try similar models online through an aggregation service such as Neuro. While the GPT3 beta signup is closed you can always get your feet wet by trying a the GPT-J (similar to GPT-3) for text generation.
What Can GPT-3 Do for Text Generation?
Assisted copywriting tools may only speed up expert copywriters' workflow if they learn to harness the tool’s power. Some may be frustrated by a change from creative writing to tweaking prompts, randomness and creativity values to get meaningful results.
Copywriters would need to become skilled at writing effective prompts: a skill now referred to as prompt engineering. It is striking how minor differences in prompts create dramatically different results. A prompt for “Create a poem about cats and dogs” will yield a vastly different result from “For today’s assignment, create a happy poem about cats and dogs.”
In our experience, prompt variations would generate text ranging from astonishing to slightly wonky to outright incoherent. The problem is that there is no surefire way to guarantee meaningful results every time. In addition to the prompt, the engine allows tweaking two key variables: temperature and top-p. Both control the degree of randomness and engine creativity in the output and have a major impact on the results.
It is important to understand that these engines are core technologies that can be integrated with all sorts of products in the content landscape. Many marketers of a wide range of tools will be sure to highlight that their shiny, new version is improved by AI.
AI Writing Services
Writing services include Jarvis and CopyAI, which provide a public interface to interact with GPT-3 using fine-tuned parameters that can—with the right prompts—generate impressive output.
You can check out some demos and reviews of CopyAI or Jarvis to get an idea of hoy they may assist, frustrate or entretain expert writers in their work. One thing I did notice is that given good training materials, the writing services create headlines matching the traits of what common web content uses.
Is It Just Generating Nonsense?
Aside from the marvelous achievement of machine-create text that ranges from somewhat coherent to brilliant, they’ve been humorously labeled as a new generation of bullshit generators by Raphael Milliere in his insightful article in Nautilus Magazine. Sometimes, compared to an enthusiast intern on their first day, the engine will write about everything with access to robust research without keen skills to differentiate between what’s worthwhile, accurate or inherently relevant.
In the book “On Bullshit,” Harry Frankfurt presents an enlightening framework to understand the historical significance and the nuances of the ever more common way of communicating ideas where the speaker combines truths, half-truths and incorrect statements without necessarily knowing the difference.
Is AI Replacing Expert Writers and Copywriters?
After exploring the landscape, my short answer is: not for now. The long answer is that AI will be raising the bar for existing copy by either making it easy to generate believable content or expanding the productivity of professional copywriters. Tech circles commonly mention the 10x engineer, a team member who is so skilled and productive they can provide ten times more value than average. These tools are more likely to expand copywriters' productivity by ten times before replacing anyone completely.
A key value of an expert writer is good judgement and that is an area where AI needs to improve dramatically. There’s a whole movement around #ResponsibleAI where researchers aim to trigger outlandish and offensive output that highlight engine bias based on their broad training data sets.
Trust Will Become the Key Differentiator
In a way, GPT-3 can be considered a weapon of mass creation. Its massive creativity is tempered by its own limitations. While it is clear that high quality output will require several editing and rewriting steps before being production ready, the creation of low-quality content – for whatever purpose – is definitely enhanced.
But when these tools become commonplace, and everyone can generate somewhat believable text, who will you trust?
Why are companies so excited about this technology?
We're seeing a fundamental qualitative improvement in text output while proving that that great progress can be made by sheer scaling of models: exponentially increasing training data and configuration parameters size.
New generations will improve some of the drawbacks and will enable new businesses to spring from autonomous or assisted text generation.
Companies are investing in adopting the Content AI technology – in its current state – as early innovation drivers. Upgrading models or tuning the current ones requires companies with knowledge of the technical posibilites and limitations.
We are currently doing Proof of Concepts and test implementation of GPT3 and other Content AI for our clients. It is an exciting time for the ContentAI space. Reach out to us if you need a technical partner for this implementations.