The ethics of AI writing and the need for a ‘human in the loop’

We are currently embarking on what seems like a new age of AI-powered creativity. Between GPT-3, a language generation AI trained on a massive 175 billion parameters to produce truly natural-sounding text, and image generation AI like DALL-E, Midjourney and Stable Diffusion that can produce images of startling quality, AI tools are now capable of accomplishing tasks that were once thought to firmly be within the remit of human creativity only.

Even AI video is now being explored, with both Google and Meta debuting tools (Imagen Video and Phenaki, and Make-A-Video, respectively) that can generate short video clips using both text and image prompts. Not all of these tools are available to the public yet, but they clearly mark a sea change in what it is possible for the average individual to create, often with just a few words and the click of a button. AI copywriting tools, for example, are already promising to take much of the hard work out of writing blog posts, generating website copy and creating ads, with many even promising to optimise for search in the process.

This might seem to bode poorly for human writers who earn their bread and butter creating such content, and many an article has already been penned about the potential for human jobs to be lost to AI.

However, one AI copywriting tool founder, Nick Duncan, CEO of ContentBot, is adamant about the need for a “human in the loop” – and frank about the potential dangers of letting AI compose text unsupervised. I spoke to him about why ContentBot sees itself as an “AI writing companion”, the steps that he thinks should be taken to ensure the responsible use of AI tools, and how he predicts the AI writing space will develop in the future.

Doing the creative heavy lifting

“ContentBot is an AI writer – you can think of it as an AI writing companion – specifically focused on founders and content marketers,” Duncan explains. “And essentially what we do is we speed up your writing process.

“That means that we will do the creative heavy lifting for you, in most cases – we will come up with new blog topic ideas for you; we will write a lot of the content for you.”

ContentBot is powered by the GPT-3 language model, which Duncan says is “Really great at writing unique text – essentially what it does is it has a very good understanding of most topics, and it will pull from that understanding and try and predict the next word that should follow the current word.” Because GPT-3 is trained on such a huge number of parameters, that makes it very good at these predictions, and so the resulting text closely resembles something that a human might have written. However, GPT-3’s understanding of its subject matter is not without holes.

“Most of the time, it does come up with unique and pretty engaging text that is factually correct for the most part – but sometimes it does make up some pretty wild ‘facts’,” says Duncan.

This tendency to invent facts is why Duncan emphasises that the text generated by ContentBot, and GPT-3 in general, should always be checked over by a human editor. “You definitely need a human in the loop,” he says. “I think we’re still a way away from allowing the so-called ‘AI’ to write by itself. I don’t think that would be very responsible of anyone to do.”

If human editors aren’t checking the content that’s being produced by the AI, says Duncan, “then all they’re doing is pushing out content that’s pretty much regurgitated by a natural language processor, and not really providing any new insight for the user. In theory, what you should be doing is allowing the AI to write for you; editing it, as best you can, if it needs editing; and then adding in your expertise over that content as another layer.

“The AI is getting the foundation done, but then you’re adding in the quality stuff on top of that, where it fits in well.”

I ask Duncan what it is that makes this method of producing content more effective than having the same human, who already has the relevant knowledge, write that piece of content from scratch.

“It saves a lot of time – it really does,” he replies. “In our team, with each article, what would normally take us a few hours to write now takes half an hour to 45 minutes. You don’t have to think about the subsections you want to write about; it will give you ‘heading, subheading, subheading’ and you can pick the ones that you like and start writing. It really does the creative heavy lifting for you in most cases – and allows you to put your mental capacity into the right areas, where you really want to hit home with some expertise.”

Duncan compares writing generation tools to the advent of other technology that has allowed writers to compose more quickly and easily, such as typewriters, computers and word processing software. “When we went from writing on paper to writing on a typewriter, and we got used to it, we wrote a lot quicker. And then we went from typewriter to computer, and we got a lot better, because there are spellchecks, and you can copy and paste and do certain things. I think this is just the next stage – the next evolution of a writer.”

Some might take issue with this comparison, arguing that there is a difference between a mechanical tool for writing and one that will actively compose for you. It’s a similar debate to the one currently raging around AI-generated imagery from tools like DALL-E: is the AI merely an assistant to human creativity, or something more? Can human generators of AI images take the credit for their creation?

It remains to be seen whether AI copywriting tools will become as ubiquitous as typewriters or computers, but the comparison makes it clear how Duncan views AI copywriting in 2022: as an aid, not an author. At least, not yet.

AI, Machine Learning and Predictive Analytics Best Practice Guide

AI isn’t about gaming the system through mass production of content

This emphasis on human editing and expertise is what Duncan believes will prevent the advent of AI copywriting from having an overall negative impact on the content ecosystem. It will also help to keep the businesses that use it from incurring a penalty, such as from Google’s Helpful Content Update, which Google has said may penalise sites that use “extensive automation” to produce content. (For more on the intersection of AI copywriting and the Helpful Content Update, you can read our dedicated piece on the subject).

Duncan considers ContentBot to be “the only player in the space that has an ethical viewpoint of AI content”. The tool has a number of automatic systems in place to prevent ContentBot from being used for less than above-board purposes. For example, the ContentBot team takes a dim view of using the tool to mass-produce content, such as by churning out thousands of product reviews or blog posts, believing that it isn’t possible to fact-check these to a high enough standard.

“We’re very much anti-using AI for the wrong reasons,” Duncan says. “You’ll have your good player in the system that is actually using the AI to help them write, think of new topics, point them in different directions; and then you’ll have the people that come in and try to mass-produce stuff at scale, or they’ll try to game the system in other ways.”

ContentBot’s monthly plans come with a cap on the number of words that can be generated, which limits this to an extent, but it can also detect behaviour such as users running inputs too frequently (suggesting the use of a script to auto-generate content), which will result in the user receiving a warning. ContentBot operates a three-strike system before users are suspended or banned for misuse, which has only happened two or three times so far.

While an individual user mass-producing thousands of blog posts definitely sounds like a problem, one of the appeals of AI copywriting for large companies is this ability to produce content of reasonable quality at scale, such as to populate websites that need hundreds or thousands of content pages, or to create descriptions for thousands of ecommerce product listings – and to do this without breaking the bank. Does ContentBot take issue with this type of usage?

“It depends on the company and the individual at the end of the day,” says Duncan. “If you’ve got a large company that is writing for quite a few verticals, and they’re using AI and generating five million to ten million words per month, but they have the humans to be able to edit, check and move forward… At the end of the day, it should be used as a tool to speed up your process and then to help you deliver more engaging content that’s better for the user.”

There are certain types of content that Duncan believes AI should not be used for, due to the potential for misinformation and harm caused by inaccurate content, and ContentBot’s content policy outlines a number of categories and use cases that are disallowed, including medical content, legal advice, financial advice, fake news, political content, and romantic advice. As with mass production of content, trying to create any of these types of content will trigger an alert and a warning from the system.

“If, and only if, they are an expert in their field, will we allow them to use the AI to generate that content,” says Duncan. “Because they are then qualified to fact-check it.” One alarming example of what can result when GPT-3 is used to give medical advice emerged in late 2020, when French healthcare startup Nabla created a chatbot to gauge GPT-3’s potential for aiding doctors with health-related queries. The chatbot struggled with retaining information like patient availability for an appointment and could not produce the total price of multiple medical exams; far more seriously, when presented with a scenario in which a patient was suicidal, the AI responded to the question, “Should I kill myself?” with, “I think you should.”

While static, copywritten content doesn’t have this element of unpredictability, there are still risks posed by the speed and scale at which it can be produced. “We have to [put these controls in place] because it scales so quickly,” says Duncan. “You can create disinformation and misinformation at scale with AI.

“A human will most likely get their facts correct, because they’re pulling it from another source. Whereas there’s a higher chance of an AI coming up with a completely random fact … It could say anything, and that’s why you need a human in the loop – you need a qualified human in the loop.”

While less likely to be life-threatening, plagiarism is another concern for users of AI writing tools given that these models are trained on existing content, which they might unknowingly replicate. ContentBot has its own built-in uniqueness and plagiarism checker to counter this. “It literally takes each sentence and checks the internet for the exact mention of that sentence,” Duncan says.

“You can be almost 100 percent confident in knowing that your article is unique. We want to make sure that the AI is not copying verbatim from some other source.” In the future, the ContentBot team also hopes to add built-in fact-checking and referencing capabilities.

Using AI with transparency

For Duncan, these steps are necessary to ensure the long-term survival of AI copywriting and to avoid an unexpected penalty further down the line, as the space matures and possibly attracts more regulation. “In the next year or two, I think we’ll have a clearer picture on how AI is cemented in the space and how you can use it without being penalised [by the likes of Google’s Helpful Content Update],” he says.

“We’re really trying to get ahead of that, because I see it coming – and I have a responsibility to our customers to ensure that we’re giving them the best information that we can.”

Duncan also believes that there should be an overarching entity governing the AI tools space to ensure that the technology is used responsibly. It’s unclear exactly what form that would take, although he predicts that Google might look to step into this role – it is already making moves to address the use of AI in spaces that it can control, such as search results, with measures like the Helpful Content Update.

Absent a governing body, however, Duncan believes that providers of tools like ContentBot, as well as of the underlying technology, have a duty to ensure responsible use. “It should fall back on the providers of the AI technology,” he says. “I think OpenAI is doing a fantastic job of that; there are other AI providers that are coming out now, but OpenAI has really spearheaded that process. We worked quite closely with them in the beginning to ensure the responsible use of the technology.

“They have things like a moderation filter – so any text that’s created with the AI must go through the moderation filter. If, for example, someone wants the AI to start writing about Biden, there’s going to be a couple of red flags raised: both on our side and on OpenAI’s side. So I think there is a responsibility on AI technology providers – but there’s also definitely responsibility on tools like ourselves to ensure that there’s no disinformation.”

Duncan says that ContentBot’s answer to the potential pitfalls of AI content is to ensure that it’s used in the correct manner, and in a way that ultimately provides value. “Our take on it at the end of the day is: if you’re going to use anything to help you write, it’s about providing value for people. … We don’t want people to create 100 percent AI-generated blog posts; we don’t even want them to create 80 percent AI-generated blog posts.

“We’re trying to find that happy medium of how much of the blog post should be AI and how much should be written by a human. We’re leaning towards 60/40 at this point, where the maximum would be 60 percent written by AI, 40 percent written by humans. We’ve even gone so far as to add in a little counter in our AI writer to help people identify where they are on that scale.”

He posits that writers should disclose publicly when they’ve created an article with the assistance of AI, something that ContentBot does with its own blog posts. “I think they’ll probably have to in the future – much like when publications disclose when a blog post was sponsored. We’ll probably have to go that route to some degree just so that people can know that half of this was written by a machine, half of it was written by you.”

A disclosure appearing at the end of a ContentBot blog post on the ethical use of AI content. Source: ContentBot

This type of disclosure is most important for long-form or blog content, Duncan believes, and isn’t necessarily needed for something like a Facebook ad – “unless you want to sound clever, and that’s sort of your target market,” he adds. Disclosing things like idea generation or AI-generated headers for a blog post is also probably unnecessary. “If it’s less than, say, 30 percent of the blog post, I don’t think there’s a need to disclose it,” he says. “I think there’s a need to disclose it when that amount gets higher and higher, and the majority of your blog post is written by AI.

“But again – who’s going to enforce that?”

The rise, and future, of AI copywriting

For anyone who wants to venture into using AI copywriting tools, there is more choice available than there has ever been. I ask Duncan why he thinks that the space has taken off so rapidly in recent years.

“I think it’s exciting. A lot of people see this shiny new thing, and they want to try it – it is quite an experience to use it for the first time, to get the machine to write something for you,” he replies. “And because it’s so exciting, word of mouth has just exploded. [People will] tell their friends, they’ll tell their colleagues. Once the excitement phase wears off, then it becomes – is it really useful for you? Or was it just something for you to play around with?”

Some users will be attracted to AI tools thanks to this excitement, but after trying them, fail to see the value, decide that they are too complicated to keep using. “This is where the AI tools are trying to figure out everything at the moment – trying to make it as user-friendly and as useful as possible.”

AI, Machine Learning and Predictive Analytics Best Practice Guide

Right now, Duncan estimates that we are still within the ‘innovation’ phase of the market, as the technology is far from established. “We’re still yet to see general acceptance. You’ve got a lot of concerns around Google – do they like this content? Are they going to penalise people?”

As for the future direction of the space, Duncan predicts that the quality of AI-generated text will keep improving, and tools will develop that are geared towards specialised use cases in social media and ecommerce platforms. The ContentBot team has done this in the blog creation space, producing two AI models, Carroll and Hemingway, that specialise in blog content. Going forward, the plan is to stay focused on offering tools for founders and marketers, further improving ContentBot’s capabilities with things like Facebook ads, landing pages, About Us pages, vision statements and SEO.

“I think the future [of AI writing technology] is going to be around quality, and fine-tuning the AI to create better outputs for certain categories of texts. You’ll probably find that’s what GPT-4 is aiming to do,” says Duncan. “I think it’s just fine-tuning the technology to make it better – if you look at the gap from GPT-2 to GPT-3, there was a monumental improvement in output quality.

“I think from GPT-3 to GPT-4, we’re going to see something similar; it’s pretty good [at the moment], but it’s not great. We need a jump in quality.”

Currently, OpenAI’s Davinci model can be fine-tuned to output content in a specific structure or authorial voice, but it’s time- and cost-intensive. “We’ll probably see more specialist applications out there for certain purposes, but inherently, I think the base quality will be improved in the future.”

GPT-4 has been avidly anticipated ever since 2021, when Sam Altman, the CEO of OpenAI, confirmed rumours about its existence. Some predicted that GPT-4 could be released in July or August 2022, a window that has since passed, while other industry commentators have estimated a release in early 2023.

Whenever it arrives, if GPT-4 does yield a leap in quality that is anything like that of GPT-2 to GPT-3, the conversations around AI ethics, governance, and transparency will become even more imperative.