AI copywriting: An overview

Many in the marketing and advertising industry would no doubt argue that truly ingenious copy can only be thought up by a human. Really excellent copy is funny and attention-grabbing, or nuanced and empathetic, or plays subtly with our expectations and the meanings of words. Surely all of these things require a human touch?

Unfortunately, the kind of copywriting that requires this bolt-from-the-blue inspiration is not often needed, rather there are vast masses of email subject lines, search ads, production descriptions, and social ads that only require a little bit of savvy and imagination, and a lot more boring grunt work.

Because of this, and the fact that artificially intelligent algorithms are getting smarter by the day, automated copywriting tools are slowly encroaching on a territory that was once “human-only”.

Despite fears of an AI takeover, at present only a handful of companies in the marketing and advertising space offer tools for automated copywriting and language optimisation. AI is also being used in industries such as journalism and finance to produce short, factual reports, but this briefing will focus primarily on discussing AI copywriting applications in the marketing industry.

So, how does this technology work, what are its benefits, and what are its limitations? What concrete results can it produce, and are the fears about job losses grounded in reality?

Let’s take a look.

How AI copywriting works

The exact workings of automated copywriting tools will differ depending on their complexity and purpose, but at base, all of them involve converting data into “human”-sounding language (or “natural language”).

The input data needs to be structured, meaning formatted in a way that a machine can interpret – this is the same principle behind adding structured data markup to webpages for search engine optimisation, as search engines are also machines that need structure in order to interpret data.

In order for automated copywriting software to produce readable, human-sounding content, humans need to program in a set of rules that dictate how the content is formatted. From The Ultimate Guide to Natural Language Generation, a resource created by Automated Insights, the company behind natural language generation platform Wordsmith:

“The output of every current NLG solution is powered by the narrative design (also referred to as “template”, “intent”, or “narrative type”), which is constructed by the end user of the NLG solution or by the provider of the software.

“Within this narrative design are rules, also referred to as “conditional logic”, that trigger different outputs based on the data set behind the content. The most powerful NLG solutions allow users to edit these rules and narrative structure to best fit their needs and have those changes reflected in the output instantaneously.”

The most sophisticated AI copywriting tools available to marketers and advertisers have the capacity to learn from, and mimic, a specific brand’s tone of voice. Creating this type of tool requires an in-depth understanding of linguistics as well as programming and AI. Parry Malm, CEO and founder of Phrasee, which uses AI to generate and optimise email subject lines and ad copy, told the Econsultancy blog that the company has an entire team of computational linguists.

“We’ve also built an in-house programming language for language. Our linguists construct language generation models tailored to each brand,” Malm said.

N.B. We profiled one of those linguists for our Day in the Life series on the Econsultancy blog, and you can read more about her role and what it entails here: A day in the life of… an AI language technician.

This is the human side of automated copywriting: it requires people with an understanding of copy, language and tone to shape the AI’s output, check it, and fine-tune it to ensure that it’s intelligible and natural.

Creating an effective brand tone of voice in the first place also requires a human touch. In a future where automated copywriting is much more widespread, the role of a copywriter could potentially shift to that of “tone designer” or “voice designer”, creating a brand persona that AI can then mimic and replicate. We’ll touch more on this, and the future of automated copywriting, a bit later on.

The applications of AI copywriting in marketing: Some case studies

How is automated copywriting being used in marketing and advertising? Let’s take a look at some of the key players in this space, how brands are using the tech, and the results that they’ve seen.

Phrasee and Wowcher

As previously mentioned, Phrasee is an AI-powered copywriting tool that composes email subject lines, Facebook and Instagram ads, and push notifications in a brand’s unique tone of voice, and then automatically tests them until it finds the optimal wording.

Founded in 2015, Phrasee originally started out focusing on just subject lines, before expanding to Facebook and Instagram ad copy and push notifications in 2018. The company works with brands across a variety of sectors, including ecommerce, travel and leisure, hospitality, and banking.

One of the companies using Phrasee’s tool to optimise Facebook ads is deals company Wowcher. An existing customer of Phrasee’s email subject line solution, Wowcher became the first brand to trial Phrasee’s new Facebook ad tool before it was available to the public.

As a company running a large number of campaigns simultaneously at any given time, Wowcher has unique needs when it comes to ad copy. The brand started out with a very constrained budget to begin with, testing four AI-composed ads against one human-composed one, and slowly scaled up its investment as the campaigns began to show results.

The AI-composed ads ultimately delivered Wowcher a 31% reduction in cost per lead, and also managed to raise the brand’s Relevance Score – which measures the quality of any given Facebook ad – to an average of 9, a score that indicates very high quality and relevance, prompting Facebook to boost the ad more on its platform.

The AI was even able to master the correct usage of emoji in ad copy, an area in which it might have the edge over humans, who are more likely to agonise over the exact shade of meaning conveyed by emoji and how appropriate they are to employ.

“Humans spend a lot of time trying to interpret the contextual validity of emojis, because it’s a new way to express sentiment,” Parry Malm told the Econsultancy blog. “However, our AI stack doesn’t care about that, and doesn’t have preconceived bias about what’s right and what’s wrong. It looks at emojis like it would any other part of speech.”

Persado and Hostelworld

Another company in the AI-powered copywriting space – and possibly the most well-known – is Persado. Founded in 2012, the company launched a “self-service” platform in 2015 that allowed companies to optimise their email subject lines (and later push notifications and display ads) using Persado’s software. In 2016, Persado gained widespread attention when Goldman Sachs led a $30 million funding round into its technology.

Persado offers various forms of AI-optimised creative, including SMS messages, email subject lines, Facebook ads, push notifications and display ads, as well as “unmatched insights on how and why specific words and phrases impact marketing campaign performance” (per its product page). Like Phrasee, Persado’s technology can replicate a brand’s individual voice, and the brand boasts that its AI can also write in 25 different languages.

Hostel-focused booking platform Hostelworld made use of these capabilities when it partnered with Persado to scale up its marketing messaging. The brand targets an audience of 18-35-year-olds, and its customers expect personalised messaging. However, executing this across the more than 170 countries that Hostelworld books in is a challenge to say the least.

“Manual testing provided some insights, but with more than 500 audience micro-segments, they needed a solution that could scale as they continued to gather data about their customers,” writes the Persado blog.

Hostelworld used the ‘Persado Message Machine’ to test massive volumes of personalised messaging in an effort to determine what resonated best with customers. The brand ran more than 200 experiments and sent close to 75 million messages, resulting in an 86% overall uplift in click-through rate, and 12% average lift in open rate.

In particular, one experiment with email subject line messaging found that using customers’ first names in a subject line increased response rates by an average of 15%. Hostelworld also tested wording designed to evoke different emotions in subject lines and found that ‘Achievement’ resonated best with its customers, netting a 17% uplift in open rates compared to the control.

Wordsmith and Orlando Magic

Wordsmith is a natural language generation platform created by technology company Automated Insights. It specialises in turning data into an “insightful narrative” and works with clients including Associated Press and Yahoo! Sport to generate automated reporting.

It also partners with American professional basketball team Orlando Magic to create AI-powered customer communications. In a webinar with Automated Insights, Anthony Perez, EVP Strategy at Orlando Magic, explained that Orlando Magic is a data-driven and innovative company that relies on ticket sales as a key form of revenue.

Orlando Magic used Wordsmith’s automated copywriting to create personalised emails targeting fans who it knew were trying to resell their tickets, based on the data it had about how likely they were to be successful. The company wanted to let its fans know about an alternative scheme it had developed for ticket resales which allowed fans to trade their seats in for ‘Magic Money’, a virtual currency that they could use to purchase upgrades and other experiences for their next game.

The campaign generated one of Orlando Magic’s highest open rates for an email marketing campaign, and upon asking customers who engaged with the email to rate its helpfulness, the company found that four out of five customers rated it as ‘helpful’.

Orlando Magic later expanded its use of Wordsmith into writing news and notes on the team’s players for its mobile app that were updated daily with new content on a player’s performance throughout the season – something that would have been impossible to achieve with human copywriters.

The company also used Wordsmith to develop personalised push notifications which would target app users with remarketing messages based on products they’d looked at in the ‘Magic Marketplace’, encouraging them to purchase.

“We see a really great opportunity here to drive that one-to-one messaging at scale,” said Perez.

Alimama’s AI copywriter

Alimama, the digital marketing arm of Chinese ecommerce and technology giant Alibaba, unveiled an AI copywriting tool in 2018 that the company claims can pass the Turing test.

While it’s unclear whether Alimama has actually put its tool through the Turing test or whether it simply means that the AI copywriter is indistinguishable from a human to the casual observer, the tool has definitely had plenty of material from which to learn. Its deep learning and natural language generation abilities have been trained on millions of high-quality product descriptions from Tmall and Taobao, two of China’s largest ecommerce websites.

Alimama claims that the tool is capable of producing 20,000 lines of copy per second, and allows brands using the tool to adjust the length and tone of the copy, choosing from options such as “promotional”, “functional”, “fun”, “poetic” and “heartwarming”. It is currently being used by merchants across Alibaba-owned sites such as Taobao, Tmall, Mei.com and 1688.com, who make use of it on average a million times per day.

Dire predictions have already been made about the technology’s ability to put ecommerce marketers out of work, though Alimama insists that the tech is designed to augment humans, not replace them.

“For merchants, from today onwards, AI can take care of a portion of their copywriting needs. And it significantly changes the way [copywriters] work: They will shift from thinking up copy—one line at a time—to choosing the best out of many machine-generated options, largely improving efficiency,” Alimama said in a statement.

It’s a classic argument in favour of automation: AI isn’t putting humans out of work, it’s taking care of the boring parts and freeing them up to play to their strengths, like creativity and empathy. However, not everyone is convinced by this idea.

Mel Henson, an author, consultant and copywriter specialising in ecommerce websites and retail catalogues, makes an excellent point at the end of an article for Smart Insights when she points out that junior copywriters often cut their teeth on ‘boring’ tasks such as these. “If those kinds of writing opportunities disappear, how do the next generation of copywriters perfect their craft?” she wonders.

With that said, Henson doesn’t despair for the future of the copywriting industry, but simply emphasises that it needs to adapt. “The industry as a whole will need to rethink training and qualifications for copywriters. If that helps humans stay smarter than robots, it’s got to be a good thing.”

AX Semantics

AX Semantics advertises itself as a “Natural Language Generation tool for editors, not developers” that “creates meaningful written content” including product descriptions, news articles, business reports and documentation. It was founded in 2009, and is based in Stuttgart, Germany, with customers from 15 different countries across Latin America, Europe and the US, per an October 2018 blog post.

Like Persado, it is also multi-lingual, offering 110 different languages that “can be based on polyglot data” – so for example, the input data could be in Spanish, while the output data could be in French.

The tool uses data as the basis for the content it produces, and lets its users define the output, including speech style, wording, text structure and keywords. These rules are then used to generate the final piece of content.

While AX Semantics seems like a slightly more basic tool than some of the others we’ve seen here, with no grand claims of being able to emulate a brand’s unique tone, it still has the ability to save brands a lot of heavy lifting by automating their content at scale and updating it in real-time.

The ability to “translate” automated content between multiple languages is also a major boon: a 2014 study by Common Sense Advisory surveying 3,000 consumers in 10 non-Anglophone countries found that 75% prefer to buy products in their native language, while 56% will spend more time on sites that are in their own language than sites that are in English.

AX Semantics generates product descriptions for German retailers Euronics and Otto that are optimised for search and conversions. It also works with an online parenting website, NappyValleyNet, to produce content for its schools guide – however, a read of the sample texts on AX Semantics’ website is a little unconvincing, as it appears some input data is missing and some sentences don’t read well.

OpenAI: Is the hype justified?

It’s impossible to explore the subject of AI copywriting in 2019 without mentioning OpenAI, a non-profit backed by Elon Musk, which recently unveiled an artificially intelligent text generator with an unparalleled ability to generate “human”-sounding writing from just a few lines of text.

The generator, known as GPT-2, prompted a storm of hysterical headlines when it was released in February, due to the fact that researchers behind the project announced that they would not be releasing the full model due to “concerns about malicious applications of the technology” – that is to say, fake news.

“When is technology too dangerous to release to the public?” queried Slate, while Ars Technica proclaimed, “Researchers, scared by their own work, hold back “deepfakes for text” AI”. The Guardian’s Hannah Jane Parkinson wrote, “Brace for the robot apocalypse”.

In a common-sense piece for The Next Web, Tristan Greene pointed out that OpenAI’s decision to withhold some of its research (the non-profit made a smaller version of the GTP-2 model available for researchers to experiment with, as well as a technical paper) was not unheard-of, and that the AI was mostly remarkable due to its having been trained on a much larger dataset (8 million web pages, selected from outbound links on Reddit that received at least three Karma).

Greene writes: “Here’s what the headlines should’ve looked like: ‘OpenAI improves machine learning model for text generator.’ It’s not sexy or scary, but neither is GPT-2.”

The author lays the blame partly at the doorstep of OpenAI, for being unclear about the reasons why it had decided to withhold part of the research, and partly at the doorstep of the media, who made the story about OpenAI’s decision to withhold its full model – not the researchers’ progress. The hyperbolic headlines also did a disservice to the thoughtful reporting that was behind most of them (Slate’s article, for example, was broadly in agreement with The Next Web’s stance that the whole situation had been significantly exaggerated).

However, when it comes to advances in AI, exaggerated news headlines are more or less a given – as Greene himself said, “the media never met an AI development it couldn’t wave pitchforks and torches at.”

The potential for accessibility

This isn’t to say that GPT-2 isn’t a significant development for natural language generation. One of the notable achievements highlighted by OpenAI’s blog post is the fact that GPT-2 doesn’t require “task-specific training data”: in other words, it doesn’t need to be trained on news reports to produce a news report, or email subject lines to produce an email subject line. Whatever sample of text it is fed, it will adapt to and continue in the same style.

“GPT-2 outperforms other language models trained on specific domains (like Wikipedia, news, or books) without needing to use these domain-specific training datasets,” the researchers wrote. The AI was also able to accomplish language tasks like summarisation, translation and ‘reading’ comprehension, again without needing task-specific training.

Will Knight of the MIT Technology Review, one of the publications granted access to the full GPT-2 model, comments that, “in contrast to most language algorithms, the OpenAI program does not require labeled or curated text. It simply learns to recognize patterns in the data it’s fed.”

These are both significant steps forward for natural language generation, and hint at a future in which automated copywriting could become much more accessible, without the need for extensive training on specific types of text, or even structured data. Instead, one AI could turn its (metaphorical) hand to any kind of content generation, once fed a few lines of sample text.

With that said, there are reasons to believe that we’re still a way off from seeing this scenario become a reality, or from needing to worry about GPT-2 making human journalists and copywriters redundant.

First of all, GPT-2 is impressive, but far from perfect as it stands. It is able to generate large amounts of realistic-sounding text from a few opening lines, but a close reading (and fact check) will easily reveal that the details are invented, if not physically impossible. It is unlikely that GTP-2 would be able to create a marketing best practice guide from a few lines about SEO or hold up particularly well with more technical subject matter. Thus far, it still seems best suited to more limited copywriting tasks with a narrow, predictable subject matter – though it can accomplish those extremely well.

Secondly, the theoretical capabilities of technology are always years ahead of what is widely – and commercially – available. We can already create amazingly immersive experiences with AR and VR, but the technology is still uncommon in our homes. Last year, Google Duplex stunned and unnerved the tech industry with its uncanny ability to mimic a human booking an appointment, but the service is only just rolling out to consumers on a limited basis (on certain devices, in English, in 43 US states) – at which point it will become more evident what the capabilities and limits of the technology actually are.

GTP-2 has illustrated what’s possible with natural language generation, and we will all be keeping a close eye on developments with OpenAI from now on. Its existence gives us a better idea of what a future with widespread automated copywriting might look like – even though that future may still be a way off.

What does the future hold for copywriting in the age of AI?

Automated copywriting has proven its ability to achieve impressive things in the marketing industry, including creating highly personalised messaging at scale and finding the exact language choices that will lead to an uplift in engagement or conversion, all while maintaining a brand’s unique tone – and more importantly, a ‘human’ tone.

However, what’s also evident is that only a small handful of companies currently offer this technology, and while increasing numbers of brands are beginning to realise what can be achieved with it, it is a way away from taking over the industry. The argument could be made that automated copywriting in its current form only has a limited set of applications – but in many ways, adopting AI in just one area of copywriting is a more achievable goal than completely overhauling a company’s processes and workflows to automate all copy. Yet relatively few brands have taken that first step – a step which is relatively familiar, the integration of cloud-based software with an email service provider and CRM. This is simply optimisation, after all.

The reality is that companies take time to catch up to the pace of innovation. More than 15 years on from the advent of smartphones, many companies still struggle with basic mobile best practice. While this doesn’t mean that radical change isn’t coming to the way that marketers approach copywriting, it does mean that it won’t happen overnight.

Some trends can be identified from the automated copywriting solutions currently available that hint at what the future might hold, and how it could affect the role of a copywriter in marketing campaigns.

First of all, automated copywriting still requires the input and supervision of a human being who knows the craft, whether it’s adjusting variables on a content-by-content basis, programming the frameworks that will result in a natural-sounding output, or creating a unique and fitting brand tone of voice for an AI to replicate. As previously mentioned, we could see the role of a copywriter shift to focus on the creative work of designing brand tone, with AI then taking care of the legwork of replicating that tone across large amounts of copy.

In the case of GTP-2-style content, a human copywriter might take on the role of fact-checker and sense-checker; or add and correct technical details on a content “framework” created by the AI.

Secondly, entirely new roles are already emerging that combine copywriting and linguistics skills with technical expertise – such as the previously-mentioned role of ‘AI language technician’. A widely-cited PwC report released in 2018 predicted that AI will create as many jobs as it displaces due to improvements in the economy and productivity, just as previous major technological leaps like the advent of steam power and computing have done.

The digital age has created hundreds of jobs that never previously existed (including ‘digital marketer’), and it isn’t unrealistic to expect that the age of AI will do the same. It seems highly likely that the discipline of copywriting will gradually become more data-driven and technical – just as marketing, along with countless other disciplines, has done.

While some may consider this a loss, the examples from this briefing have shown just what can be achieved with AI when it is applied to copywriting: scientific testing and optimisation, fine-tuned personalisation, frictionless translation between languages, and previously unimagined scale, to cite just a few examples.

The automation of copywriting and content creation will ultimately be of net benefit to marketers and advertisers (as well as journalists, financiers and countless others). While there might be a shift in the skills required of a copywriter – which the industry will certainly need to adapt to – copywriting will still be as effective, valued, creative, and in demand as it was before – if not more so.

The post AI copywriting: An overview appeared first on Econsultancy.

Source link