Artificial intelligence

6 ways to spot AI-authored copy

Generative AI can do some amazing things. It’s a painter and musician and coder and, of course, author.

How good it is at performing those roles is a topic up for debate. AI artwork regularly drifts into accidental surrealism, with superflous human limbs and bizarre fusions of objects.

But what about AI-generated copy? While the glitches can be glaringly machine-like in a picture, they’re more subtle in a passage of text. Here’s how to spot them.

Repeat offence

My father was an avid reader and writer. He’d often take a keen interest in the essays that I wrote for school. One of his most useful pieces of advice was to avoid repeating myself.

He was right. Repetition weakens writing. A lack of variety in phrasing can make an article dull. Redundancy labours a point through duplication. Human authors do their best to avoid these.

A computer on the other hand will be unlikely to police itself to nearly the same level. Snippets on a topic will be pulled from here and there and this and that to build an article. There’s a strong probability that key points will be repeated over and over and over*.

*Sorry, blatant repetition, I know.

Yesterday’s news

Generative AI platforms are trained on huge sets of data. Unless the platform in question has live access to the internet, its knowledge base only extends as far as the last update. The platform would not be privy to latest developments on any given topic.

Old news is unengaging at best and misleading at worst. Humans and search engines alike favour high quality, original content. Out-of-date doesn’t necessarily mean no longer correct. It can simply be information that has become so commonly known that further publication is redundant. Customers prefer personalised email to non-personalised!? Hold the front page!

Get your facts right

If you’re using a generative AI tool to produce or aid articles, never take it for granted that the software knows what it’s talking about. Because, technically speaking, it does not know what it is talking about. It algorithmically reproduces and combines content from multiple sources – which can include information that is no longer true, or perhaps has never been.

As a reader, keep an eye out for factual errors and especially contradictions. If it smells fishy, trust your instincts and verify the information elsewhere.

What’s the story?

A good quality article written by a human has a story-like flow. There’s a beginning and a conclusion. Computer-generated articles on the other hand often hit an abrupt end.

And what’s a story without a message? A good story makes you think and feel something. A robotic author literally feels nothing, so why should you as a reader?

Don’t you dare

Language models by default are clinically impartial. A platform won’t automatically spit out a controversial opinion that makes you stop in your tracks. It’ll compile a collection of neutral statements of fact.

You can coax it out of its formal shell of course with prompting. The results are perfect – if you’re aiming for a plasticky have a nice day flavour.

A human’s opinion piece carries real emotion and real sentiment. Even an article that you fervently disagree with can be an excellent read. There’s a human-to-human spark that is missing with AI.

It just feels… off

You’ve probably heard about the uncanny valley. It’s a term often applied to computer-generated or animatronic simulations of human faces. Our brains are acutely conditioned to recognise faces with their every nuance and motion. It would take something very special to fool us.

AI-authored articles often fall into a linguistic uncanny valley. Attempts at personality are injected jarringly, equivalent to writing “LOL” in the middle of a legislative document. Instead of a human voice shining through the words, there’s a perceptible artificiality to those written by a computer.

Image of mannequin faces that demonstrate the uncanny valley effect.
This, but in words.

How much does it matter?

If we read something and enjoy or learn from it, does it matter if a computer wrote it? What if it was only computer-aided? Platforms like ChatGPT can be very useful as idea generators.

Is it ok if the text is a piece of marketing blurb rather than an opinion piece? How about a social media post, or a response to? Can there be any value to fiction or poetry conjured through ones and zeroes?

Ultimately it’s up to each of us as individuals to decide how we feel about AI, but it’s hard to deny that authentic human content is going to become rarer. With that in mind, it can’t hurt to be able to tell the difference.

Artificial intelligence

Humanised email marketing in the age of AI

We live in the age of artificial intelligence. Sort of. More accurately, we live in the age of algorithmic content‑generation.

Computer programmes can write copy, draw pictures and generate code in seconds. Again: sort of. The intelligence aspect of AI is very overstated – these programmes have no actual understanding of what they’re doing. Therefore they are blissfully emotionlessly unaware when errors occur in their output. And those errors can be both glaring and numerous.

This of course is a technology in its early stages. With a human at the helm, to guide and refine, it can already be used to great effect. So what happens as the technology develops, and the need for human input becomes less and less?

Marketing by machines

Computer programmes can analyse customer behaviour and serve up unique one‑to‑one content in email marketing. The scale, speed and accuracy far exceeds the capabilities of any human marketing department.

For years this has primarily meant product recommendations. These often take the form of a block of personalised content within an otherwise static email. As we move into a more sophisticated era of content generation, it’s not far‑fetched to imagine entirely computer‑authored emails tailored to each unique customer’s preferences from top to bottom.

Be contactable

Email is – or at least should be – a two-way communication medium. All too often however, companies send marketing emails from no-reply addresses. It’s a closed door, and tells the customer: our message matters, yours does not.

As marketing becomes increasingly robotic, leave that door open instead. Give your customer the reassurance of accessible human help.

Join the conversation

Social media is the perfect medium to humanise your brand. Reply to comments, good or bad. Show a sense of humour. Let the world see that there’s a human presence behind the corporate facade.

Likewise, don’t let negative feedback or complaints go unanswered. Nothing puts me off a company like cookie‑cutter replies to bad reviews on Trustpilot. Turn negative into positive by demonstrating a human solution when things go wrong.

A matter of preference

Machine-learning is powerful. But an algorithm will never know your customer better than they know themselves. That’s why a preference centre remains an excellent starting point for personalised email content.

A tick‑the‑box preference form gives customers an easy way to tell you their interests. The obvious benefit is more relevant email content. The less obvious but equally important benefit is the message it sends about the value you place on human choice.

Turn customer into creator

You can create content for your marketing emails. Machines can create content for your marketing emails. But you know who else can create content? Your subscribers.

Invite your customers to share photos or other content themed around your products. Perhaps tie it to a competition. Incorporate this content into marketing emails and suddenly they have more of a community feel rather than corporate.

Everyone gets a vote

Continuing the topic of subscriber interaction – why not encourage engagement through surveys? A simple click‑to‑vote system can be tied to a database at the back end. And let’s keep that two-way communication in mind – current results can be shown via images generated in real time.

The benefits are numerous. Customers see that their opinions are valued. Surveys serve as an insight into consumer behaviour. And your emails become an engaging, living thing.

Be individual

Authentic human content is going to become increasingly uncommon… and increasingly valued. A unique brand voice will be more important than ever.

But why stop at brand level? A company is made up of individuals. Opinion pieces by team members or guest content by industry experts can give your emails a captivating human touch.

The value of authenticity

Generative AI is a fascinating development of the digital age. Anyone can become artist or author or musician at the push of a button. And yet when that work is devoid of effort and meaning, it becomes a kind of creative candy not worthy of perusal. Does that matter when the purpose is marketing rather than self expression? Does it matter when the desired output is a catchy pop song rather than a heartfelt ballad? What happens when it becomes impossible to discern between the creations of a human and a computer? This is a technology that raises many questions across all aspects of human life.

Bringing the focus back to email marketing – it’s important to keep up‑to‑date with technological developments. But perhaps the best marketing in the coming years will be that makes a real human connection with customers.

Artificial intelligence

Is ChatGPT your next email developer?

There are two ways to build a marketing email:

  • Hand-coding
  • WYSIWYG editors

We swear by the former, not only for quality but also for speed. But what if there’s an even better, quicker way?

Enter ChatGPT. Much hype has surrounded the AI platform’s ability to code. It can conjure up HTML and CSS in seconds. So too can it generate Javascript functions or back-end PHP or even truly hardcore programming such as C++. Whether or not it does it correctly is a different matter.

Let’s not worry about that just now. We’re here to put AI email development to the test, so let’s find out if ChatGPT can put together a responsive mailing.

The quirky world of email development

If you work in email marketing in any capacity, you likely already know that it requires some unusual coding techniques. There are lots of devices and email services out there, and they have widely different ideas about how HTML and CSS should be interpeted. In order to construct a mailing that looks presentable on all of them, the developer needs to be aware of these limitations and inconsistencies and the arsenal of tricks to work around them.

Has this niche set of knowledge made its way to ChatGPT? We’ll start with a bare bones request.

Prompt:

Code a responsive email template

Result:

Unusable!

ChatGPT has produced a very basic HTML document with some styling, but I wouldn’t call it an email template. It doesn’t include any means of stacking content on mobile, and the structure is based on HTML div elements rather than tables. While divs are the building blocks of a web page, tables remain the most reliable method for email.

On the plus side, it has picked an inbox-friendly width of 600 pixels. And it’s nice to see that accessibility has been implemented via an image description and a proper heading tag.

My request was extremely minimalistic. I need to do my part here too, and that means being more specific about what is needed.

A little lot more instruction

Take two. We don’t want divs, so let’s tell ChatGPT to use tables. There are some basic universal requirements in responsive email, so we’ll nudge it in the right direction regarding those.

Prompt:

Code a responsive email template, using HTML tables for structure. Set the width to 600 pixels on desktop, with a fluid width on mobile. Include CSS classes to enable stacking of content on mobile devices. Include all known email client fixes that are still relevant. Set the page background to a light grey colour, and the email content area to white.

Result:

Better… but still broken beyond repair.

This time it has used tables for structure, so that’s a major improvement. It has also set a breakpoint. That’s the backbone of responsive email code and the point at which mobile-specific styling is triggered. There’s some kind of attempt at stacking code, but I can see at a glance that it isn’t going to work. We’re also missing the usual pile of fixes that make an HTML email possible.

A rethink is needed.

A different approach

Here’s what we’re going to do: hand code a simple email, and then provide ChatGPT with detailed directions in order to recreate it. This is a reverse way to approach our project, but perhaps if ChatGPT has a more defined goal it will be able to produce a usable template.

Our email will have a main image, intro paragraph and a button. Under those will be a couple of secondary features laid out side-by-side on desktop, and stacking on mobile. For the sake of this test, let’s forget about any header and footer.

Image of our intended email layout

Now for our prompt. It’s going to be a long one. Let’s give ChatGPT a fighting chance and focus primarily on structure rather than styling.

Prompt:


Code a responsive email template, with the following requirements:
• 600 pixels wide on desktop
• Fluid width on mobile
• A page background colour of #f1f1f1
• Email content area background colour #ffffff
• A hero section with an image, heading, paragraph of text, and a button
• The hero image should be 600 pixels wide, to match the email content area
• Button should be pill-shaped, with a background colour of #a56e53 and white text
• Under the hero section should be two secondary features
• Each of these must also have an image, heading, paragraph and button
• Secondary feature images will be 290px wide on desktop, to match their containing column, and expanding to full width on mobile
• Hero text and button should be a bit larger than those of the secondary features
• These secondary features should take the form of adjacent columns on desktop, each at 290 pixels wide
• Place a 20 pixel gap between them
• The secondary features must stack into a single column on mobile
• All parts of the email should have 20 pixels of padding on each side on mobile, except for the hero image which can be full width and touching the edges of the viewport
• All body text should follow this font stack: HelveticaNeue-Light, Helvetica, Arial, sans-serif
• All body text should be colour #61524b
• All heading text should be colour #a56e53
• Use lorem ipsum placeholders for text
• Enter all hrefs as # placeholders
• Apply links only to buttons. Do not apply links to images
• Include all known, currently-relevant email client fixes
• Include CSS or HTML comments around each section to explain what it is or does
• Set a mobile breakpoint based on a max width of 639 pixels
• To ensure compatibility with Outlook and other email clients, use HTML tables for structure

Result:

Nice try… sort of.

In order to test this properly, I’ve saved a local copy and manually added my image references. Here’s how it looks in a browser:

Image of ChatGPT's email as seen in a web browser

At first glance, that isn’t too bad. The general layout, colouring and sizing are all correct. So too are the button shapes, and the secondary features switch to a single column on mobile as requested.

But there’s some strange overlapping going on. Our images are offset to the right, and sit partially over the grey background. This in turn causes some unwanted horizontal scrolling on mobile.

Behind the scenes, the true extent of the errors comes to light. It has reverted to a div-based structure, and uses some CSS code that won’t work universally in email.

Nonetheless, for the sake of completeness I’d like to test this as an actual email. It works, more or less, on iPhones and the Gmail app. Webmail is a mixed bag. Outlook however is where it all falls apart:

ChatGPT's email as seen in Outlook 2019

Outlook is the primary reason that email development requires such unorthodox coding methods. A lot of code that works just fine on a website, simply isn’t recognised by Outlook. Here we can see that the adjacent columns have failed and the pill-shaped buttons are reduced to tiny rectangles. To fix that would entail a complete recode.

No need to re-invent the wheel

So far, ChatGPT has failed to code a responsive email from scratch. To add some faux drama, let’s say our make-believe client is becoming impatient waiting for our make-believe email.

It’s time for a last ditch effort. At The Email Factory we already have a tried & tested template. We don’t need a new one. How about we give our base template to ChatGPT and then ask it to complete some content within that framework?

Result:

Now we’re getting somewhere.

But that doesn’t mean success. This time the template works reasonably well in a browser and even in Outlook, although the dodgy buttons are still present. The secondary features however don’t expand to full width on mobile:

ChatGPT's email using our template, as seen on an iPhone

That can however be easily fixed manually. In fact, it may be feasible to fix everything in this code rather than to start again. But I don’t want to do it myself, as that defeats the purpose of this experiment. Instead I’ll tell ChatGPT what needs to be corrected.

A few pointers

Final try. I’ve fed back some information to ChatGPT for it to make the necessary changes.

Result:

A huge step backwards.

Well, that was a big let-down. Instead of applying some finishing touches to the template, the layout has exploded. It no longer stacks on mobile. The code is now full of Microsoft conditional statements – a technique that should only be used sparingly and under specific circumstances. And the buttons? Still ugly in Outlook:

ChatGPT's corrected email as seen in Outlook 2019

Maybe with painstakingly detailed prompting and a lot of patience we could finally achieve a working email. But we’re already far beyond the point of convenience.

The current state of play

In my experience so far, ChatGPT has only done one thing consistently: fail. And I don’t only mean within the limited scope of this one project. I’ve had similar results when trying to generate marketing copy or website code. The output is usually along the right lines but ultimately too broken to actually use.

It’s clear that I set my expectations too high. The tales of ChatGPT’s near-miraculous capabilities were captivating, so perhaps the reality was always going to be disappointing. If there’s a perfect way to illustrate ChatGPT’s close-and-yet-so-far nature, it’s to ask it for an anagram.

Prompt:

Tell me an anagram of "The Email Factory"

Result:

The Fairy Comet Elf

I’ll save you the bother of checking that – it’s wrong. Trying to recreate that manually, letter by letter, results in this:

The Email Fctory e f

Re-evaluating our AI email development experiment

This project is arguably unfair from the outset. ChatGPT is a language model. Just because it can output code doesn’t mean it is a coder, or even knows what programming is.

Even so, it’s widely known that ChatGPT can generate code. So, despite all the mistakes and unusable templates, the fact that it can make a somewhat reasonable attempt is impressive.

Where do we go from here?

ChatGPT and AI in general are progressing at an incredible pace. It wouldn’t surprise me if everything I’ve written about AI email development above is laughably antiquated one year from now.

Perhaps when that time comes, I’ll prompt it to:

Write an article about how you surpass human email developers

Email Design

Can AI design a marketing email?

AI is a big deal at the moment. And by “the moment”, that likely means from this point onwards in human existence.

We’ve previously talked about AI’s skills as a copywriter and its influence on email marketing in general. Now it’s time to look at another major branch of AI content generation: imagery.

The journey begins

Midjourney is one of the most advanced and best-known AI image generators in the market. If you’re not familiar, let me explain – it’s a form of digital magic that needs to be seen to be believed.

Picture something in your head. Anything. Now type it into Midjourney’s prompt bar. Within seconds, it will generate four images based on your input. How about:

a Jaguar E-Type parked next to the Eiffel Tower in the rain

Four AI-generated images of a Jaguar E-Type in front of the Eiffel Tower

Incredible, right? Pictured unmistakably are the vehicle, landmark and weather conditions of my choice. But after the initial wow factor has worn off, it doesn’t take much scrutiny to spot the flaws. In one picture the Jag is making a nuisance of itself in the face of oncoming traffic. In another there’s not one Eiffel Tower, but two. This is a landmark that surely doesn’t need any more replicas!

As a user, there are several options at this point. We can spin again and get four brand new images based on our original prompt. Or we can ask Midjourney to create variations of what it’s already generated. Maybe we’d like to do some manual editing in Photoshop, feed the image back to Midjourney and ask it to work from that. Perhaps we’d prefer to rewrite our prompt and be a little more specific.

But let’s move on. We could play with whimsical pictures of this and that all day, but we’re here to examine Midjourney’s potential visual contribution to email marketing.

An a-eye for design

We’re going to test this from two distinct angles:

  • Overall email layout
  • Individual images

I should preface this by saying that Midjourney clearly isn’t intended for the design of websites or emails. It’s a tool with an artistic bent (and one that gives it a distinct character from its rivals). Nonetheless, let’s see what it can do as a source of design inspiration.

We’ll work with purpose. Our goal is to design an email for a make-believe travel company. Its brand colours are a sunny blue and a beachy tan. Those can sit on a traditional white background, with body copy rendered in an eye-pleasingly contrasting dark grey. AI can’t read our minds (yet), so my request is quite specific:

Marketing email for a travel company. Beach imagery. White background. Dark grey text. Use of colours #6084f7 and #c1aa60

And here’s what it came up with:

Four email layouts created by AI

TO TE MAIAY THI – do you know what language that is? It’s mangled English. And that’s an indicator as to how AI image generators work. The output is a visual echo, based on countless source images that have been fed in. That’s why it’s not uncommon to see distorted intepretations of brand watermarks or artist signatures. As with all AI products, we’re seeing the results of complex algorithms rather than any real intelligence.

But can we make use of these mock-up emails? Of the four images above, option 1 is the most useful. It may have largely disregarded my colour choices, but in terms of layout and volume of content it’s the most plausible as an email. Now we’ll conjure up some imagery to bring it to life.

It’s a kind of image-ic

The big beach image from our layout mock-up is pretty nice:

AI-generated image of a beach

That picture is of too low resolution to use in its current form but thankfully Midjourney is a flexible tool. I’m going to feed the image back to the platform, along with some instructions:

[beach.jpg] panoramic illustration of a yacht next to a tropical beach with palm trees and rocks --ar 11:4

That ‘ar’ tag at the end is a way to define the aspect ratio of our images. We don’t want the default square shape in this case, so I’ve picked dimensions that suit a long strip-like image.

Here’s the result:

Four AI-generated images of beaches

I like option 4. So let’s upscale it to a useable resolution, and I’ll modify the colours a little in Photoshop to make it more on-brand.

Final, modified version of an AI-generated beach image

We have a layout. We have a main image. Now we can move onto the final step.

Assembly required

Let’s put the AI-generated components into an actual email. If this was a real commercial mailing, we’d have generated many more trial layouts and images, and made a lot more edits along the way. Working with AI tools feels a bit like the infinite monkey theorem. Spin and spin again until we finally get the result we had in mind.

But what we have here is good enough for experimental purposes. Here’s the final product:

Is it a revolutionary design? No. Could the main image be better? Yes. The point isn’t perfection but progress.

Some final thoughts on AI email design

AI content-creation tools are becoming more sophisticated and widespread. It’s not far-fetched to see image generators becoming an everyday tool that lets marketers conjure up fresh, on-brand graphics in seconds.

In answer to the opening question – can AI design a marketing email – my verdict is sort of. It can certainly help. But at the current rate of progress, who knows what it’ll be capable of a year from now.