Why Does This Sentence Taste Like Wet Cardboard? aka… A Copywriter’s Nuanced Opinions About AI.
Recently I was on a call with D.J. for my upcoming wedding. The conversation went how you’d expect, until he found out I was a writer. Then all he could talk about was AI, he had a gazillion questions and it’s all we talked about for nearly 10 minutes.
I have been working with AI since before ChatGPT. It was all a happy accident really. I was the marketing lead for a gaming company Many Hands working on a new game Verses. Luckily we had a technical genius with a machine learning background just as AI was blowing up. He was way ahead of the game and he made sure we were too. We made a great team, he understood all of the technical aspects so he hooked up a bunch of systems in-house and I taught everybody how to use them.
Back then the only people talking about this emerging AI world were techies. Soon after that, creatives started talking, and then friends at parties started to ask me about it, after that it was my elderly mother and aunts. That’s when I knew we had reached critical mass.
And then finally, the DJ on the phone who I’ve never met is asking me. So it’s time for a long overdue blog post about it.
There’s a lot of ground to cover so I’m going to structure this post like so…
Background
Observations about how people evaluate AI
Copywriting
Editing
Research
Notes and Summary
Art
Code & Automation
Practical Takeaways
Ethics
Background - And Then John Said Let There be AI.
John was our AI architect and is brilliant in his own way. He has a reputation for being a motormouth and sometimes he could go on and on and on. He would dump a metric ton of AI gold on whatever lucky (or unlucky depending on how you look at it…) soul happened to be nearby.
And me being the social and curious marketer I am, John taught me a lot about AI. And as the company grew I worked with John to train staff on using AI, helped with implementing in-house AI tools, and integrating AI into our workflow.
Now AI is one of those polarizing topics. But I have a middle of the road view on the subject. I think it’s an amazing tech we should all learn to use. But I am honest about its shortcomings and it shouldn’t be replacing entire departments like the AI evangelists claim. (Although a shortsighted fool CEO might try anyway…)
For you techies, I know that AI and LLMs are different. But I’m just going to use the term AI anyway. For the examples here I mostly used ChatGPT but I did experiment with Claude, Deepseek, and Gemini. I’m not going to explore the nuances between these different models, that’s not what we’re doing here.
Observations on How People Evaluate AI - Think Systems Not Creation.
The starting place for most people is analyzing how good AI is at doing or making the thing. For a lot of people this is where the conversation starts and ends. But this is a limited way to evaluate AI.
One thing I’ve noticed is that amateurs are easily amazed at everything the AI can do. Those who are skilled at their craft see the flaws and limitations of the tech.
Imagine you spoke into a translator app translating your words into French. You speak into the phone and it spits out French. “Wow amazing!” But since you don’t know French you don’t know if it translated correctly, hell you don’t actually know if what is spat out is French at all. But a fluent French speaker would know right away whether the translation was good, it would take someone already fluent to properly evaluate what’s happening.
And so it is with AI and everything it makes. People who can’t write will be spellbound by the copy AI delivers. Copywriters will recognize that the copy is about as powerful as wet cardboard. Artists will notice the errors in composition, coders will recognize the problems in the code.
AI is good enough at making things that are good enough when good enough is good enough. Make sense?
And those who are masters of their craft and skilled at using AI will thrive. Not necessarily because they will create side by side with the AI (although that will be part of it.) but because they will look at AI and think systematically.
I’ve seen a tragic trend, where a CEO can be easily mesmerized by AI’s mediocre work, and he is being pitched by AI consultants, cough grifters cough cough, who make overinflated promises as to how AI can help.
Since the CEO doesn’t know a thing about graphic design, copywriting, code, software engineering, and all the other things that make up the practical foundation their business rests on. Okay he probably knows some of these but definitely not all of them. They will be easily fooled into thinking they can do away with entire departments and replace those departments with AI. And there is a growing industry of AI consultants that are happy to sell a false bag of goods.
This may make them seem nefarious, but really they might not even be aware they’re doing it. If the CEO and consultants don’t know what they don’t know, they can both be fooled about what AI can deliver.
There is more to AI than how well it makes whatever thing. What makes AI powerful are the secondary and indirect ways it can elevate your work.
Copywriting - Does this taste like cardboard to you too?
You need to understand that punching the keys on the keyboard is the least important part of being a copywriter. What makes a good copywriter is how they think. How they think about offers, the market, the customer, their psychology. Copy is just the string used to tie these elements together. A copywriter who doesn’t do these things doesn’t write. They type.
AI does a decent (not great, and sometimes not even good) job at tying words together on a string. But AI is not good at thinking about offers and how it intersects with customers’ pain.
AI spits out copy that reads like wet cardboard paste.
Going back to my conversation with the DJ I told him…
“The thing about AI and writing is that it will always give you something incredibly bland. If you go into ChatGPT and ask to write copy for your website. What it’d give you would be generic. The AI doesn’t actually know anything about your DJ services or experience. It doesn't actually know what your customers want. And so the copy it gives you will be statements that could be true for any DJ anywhere, and apply to almost literally every DJ in the world. There’d be nothing to differentiate you from the crowd.”
When ChatGPT hit the scene my boss encouraged me to get in there and start playing around. We have a mutual friend who runs a state of the art baseball pitching lab so I decided to ask ChatGPT to spin up some copy for a pitching camp.
Here’s what I got
“Looking to take your pitching skills to the next level this summer? Join our baseball pitcher summer training camp! Our experienced coaches will provide personalized instruction and feedback to help you improve your pitching technique and increase your velocity.
Our training program includes drills and exercises to enhance your accuracy, power, and endurance, as well as mental preparation techniques to help you stay focused and confident on the mound. Plus, you'll have the opportunity to practice your skills in live game situations against other campers.
Whether you're a beginner or an experienced pitcher, our camp is designed to challenge you and help you reach your full potential. So why wait? Sign up now and get ready to dominate on the mound this season”
This copy is bad. Bad bad bad! This is one of those instances where it would be easy to think this is good copy. It is written in copy-ese after all. (Which is part of the problem.)
To summarize what is wrong with this copy - is there anything in this copy that differentiates the gym from literally any other gym? As in you could take this copy and use it for literally any pitching camp. This creates weak copy that reads limp.
Now you may say “Just wait a minute Andy. Of course you received bad copy since your prompt was so terrible.” And this is true. In fact this is exactly what my boss told me, and that I could have instructed it to include specific facts that act as a USP. I could have included biographies of the coaches, I could have included proof elements in the prompt, I could have massaged the prompt to connect emotionally with the reader.
This is all true but do you see the problem? All of those other things my boss suggested I put in the prompt… That’s what lives at the core of good copy. It’s knowing those things, and understanding why they’re important, and why the reader wants them that is the copywriter’s job. The part where you push buttons on the keyboard is often the least important part of the job.
My boss didn’t realize it, but what he was asking was “In order for a prompt to create good copy, did you consider doing the copywriter’s job upfront, and then adding that to the prompt?”
If in order to get the machine to work, I need to create good copy upfront. so the machine can give me good copy. Then what are we even doing here?
If I want a machine to give me bubblegum, I’m not going to make bubblegum first, put that in the machine, and then ask the machine for bubblegum.
On the flip side - there are other ways to incorporate AI into your copy. In the baseball instance, it’s not as if those kinds of prompts provide literally zero utility. And if you are really skilled at using AI, then you can train a model on all of the ins and outs of your product and that’s where some magic can happen.
Editing - Correct Problem Wrong Answer… Usually.
This is where I am most likely to use AI. It doesn’t replace a real editor, but it’s pretty easy to incorporate it into your writing workflow. Just like you might use Grammarly or the Hemmingway App, you can start throwing your writing into an AI and ask it to evaluate the writing.
There is a game design principle which is that when someone gives feedback about a problem, they are almost always correct about the problem, but their fix for the problem is usually off.
This has been my experience when AI edits my writing.
I fed my Levine’s Trench article into an AI and asked for edits and here’s what it found.
First Edit: Swing and a Miss.
It wanted to butcher my opening. Their edit removes any sense of contrast which hollows it out. There’s no sense of progression, and why is crabs in a bucket worse than a nest of vipers? This is a miss.
Second Edit: Success
I don’t think their edit works, but it did make me realize that I didn’t need the “talks down to newer players.” part of the sentence.
Third Edit: Miss
It completely misunderstands what this part of the article is trying to do. It has technically made it more efficient, but has ripped out its soul.
Fourth Edit: Success
I agreed with the AI about the need to edit this segment. It’s wordy, drawn out, and its messiness gets in the way of its impact. But the suggestion given isn’t an edit I’m interested in. Their edit has no rhythm or any way to ramp up the emotional impact.
But it did need a rewrite so here’s where I landed.
“Hell, do you know which basketball player misses a lot of shots?
Lebron James.
Think about that for a second.
The greatest player to ever touch a basketball will play on the biggest stage, and then miss.
Whenever I’m feeling down I like to picture Lebron James missing a shot.”
At the end of the day, it gave me some bad edits, but it did find spots that needed rewrites. It’s a useful tool but don’t go firing your writers and editors anytime soon.
Research - Want to Research Your Research?
Since good writers can immediately recognize the limitations of its copywriting abilities. Copywriting gurus have taken to oohing and aahing at how well the AI can spit out research.
You ask the AI to give you research about the customer avatar of someone wanting to lose weight. And the AI will tell you about wanting to look more attractive, wanting clothes to fit better, wanting to live longer and healthier. Wow amazing. Thanks to the groundbreaking information from this AI, with the knowledge that people want to lose weight to look better and for health reasons I will be unstoppable.
Then you will need to double check the AIs work anyway. Because the AI will give you surface level information that’s already easy to infer. But any experienced copywriter knows that the best gold you’ll find is when you speak to your customers face to face. The juice, the good stuff, the things you can build entire campaigns around almost always come out of left field. It’s the things you learn that you can’t infer on the surface level that makes research worthwhile.
Then you have to worry about hallucinations. AI will just outright lie to you and when you follow up on their sources you’ll realize the AI is just making things up. Sometimes there is no source and if you ask the AI about it, they will admit that they don’t have one.
AI has been shown to cite case law and refer to cases that literally don’t exist. Go ahead and google “ChatGPT and Lawyer” if you want a good laugh.
Where AI can be effective is if you feed it with research that has already been vetted. Just asking an AI to give you customer avatar research isn’t great. But if you took all of the interviews you’ve already done and then fed those to the AI and then asked AI for its opinions - well now we’re cooking ain’t we?
Summarizing Video & Audio - And Then the Comedian Told a Joke.
Here is a perfect example of how AI smooths over any nuance and details which sucks the soul out of your writing. Because the way AI summarized the following video clip, is a great demonstration of AI’s shortcomings.
The Seattle Seahawks had traded their star receiver DK Metcalf, weeks later DK went on a podcast and shared some of his story.
I’m going to share the podcast and AI summary, and then give a breakdown.
Introduction: The hosts, along with DK Metcalf, joke about the "Get Got" podcast and its tendency to reveal personal stories [00:00].
Diet and Offseason: Metcalf discusses his diet, including his love for candy and cheat meals of fried chicken, sweet potatoes, and cornbread when visiting Mississippi [02:17].
He also mentions taking February off [01:59].
Loyalty and Team Dynamics: Metcalf expresses his thoughts on loyalty within a team, emphasizing his commitment to teammates and his observations on how team dynamics have changed [14:41]. )What commitment, to which teammates. What are the observations, which team dynamics changed, how did they change?
He reflects on the departures of key players from the Seahawks and the impact on team camaraderie [08:38].
Which team players, how did it affect team camaraderie?
Game Strategy and Frustrations: Metcalf recounts an incident where he felt his input on game strategy was not valued, leading to frustration [20:23].
Which game? What input did he offer?
Thoughts on Geno Smith: Metcalf praises Geno Smith's abilities as a quarterback [24:01].
What about Geno’s abilities specifically?
Business Side of Football: The conversation shifts to the business aspect of football, with discussions on player value, contracts, and the importance of understanding the broader picture [26:38].
What’s the broader picture? How does player value, contracts, fit into that?
Transition to Leadership: Metcalf reflects on his transition from a young player to a leader, acknowledging the influence of veterans [35:55].
What are the details of his transition? What’s the actual story here?
"Get Got" Stories: Metcalf shares a story about being overpowered by Luke Kuechly [41:26].
This tells us that there is a story, but it doesn’t tell us the story.
He also recounts a humorous incident involving Marshawn Lynch [44:31].
What’s the incident?
This summary is like taking Goldilocks and the Three Bears and rewriting it like so…
Goldilocks recounts visiting a cabin.
She talks about the differences in porridge.
She shares a story about how the beds differed.
Doesn’t quite hit the same does it?
AI will do the same thing to your copy. Vague statements that try to cover too much ground, trying too hard to be technically correct instead of setting your brand apart and making a real connection to the reader. After all, so long as what it says always applies to then it can’t ever be wrong? Right?
On the flip side, if you put pages, and pages, and pages of material into the AI. Then even if the transcription and summary is meh, the fact that it is searchable is useful. If all the information is in there, you can ask the AI to bring up every instance where X is discussed.
Art & Graphic Design - Still Useful Even When It’s Bad.
There are so many different ways to use AI for art and graphic design, so for this section I’ll share some of my favorites. The very first time I trained co-workers how to use AI, it was creating AI art.
Like everything else, if you are starting from scratch and want AI to give you a certain image, it will probably fall short. Unless you’re wanting a starting point to work off of, or if good enough is good enough. If you’re a digital artist then AI art will do wonders in your hands. AI art usually generates images that need work. There is usually a thing or two that are very wrong with the image, and if you’re not an artist you’re stuck asking the AI for another picture. But if you’re able to throw it in photoshop then you can fix all the errors in the image.
Even if you have no artistic skills, one of my favorite ways to use AI art are creating a lot of mediocre art in the brainstorming phase of a project, creating multiple variations of a design, or turning the art into something else.
Brainstorming:
Let’s talk about how to make use of mediocre art first. Even if the AI art won’t come near your finished product. It’s still useful in the very beginning ideation phases. You can ask the AI to create say a dozen different images and then when you meet with your art team you can say things “We like what’s going on with the hair in image 5, and we like the dystopian gritty vibe in image 7. The technology in picture 9 is too “Star Trekky” and we’re looking for something that’s more of a “Star Wars” vibe.
Multiple variations of design:
Say you’re designing a game, and you need to make several rocks for terrain. Instead of drawing each rock individually, you can make one rock and let the AI spin multiple versions of it. You can do this with any asset you want multiple versions that are slightly different from each other. Making a sports game with hundreds of fans in the stands, well now you only have to design one of them.
Turning the art into something else: Images make great raw material for other AI tools. You can take an image and turn it into a 3D model. You can take an image or 3D model and put it in a video. You can ask for different variations of ads as well.
I put this ad into it into an AI, and then asked it to replace the product. I asked it to change the protein bar to a Nature Valley bar, and then gave it a picture of a Nature Valley bar.
What’s surprising is that it changed the bullet copy, but also in the lower right it looks like a Nature Valley bar. The copy circling the image is wrong but that’s easy to fix in production.
Coding & Automation - Systems Systems Systems!
I’m inexperienced here so I won’t have much to say. I have used it to create simple scripts to extract data from spreadsheets and then organize them.
But some of my friends who code love ChatGPT. Even though they’ll see where the code falls short. AI is a godsend for creating documentation. And even if its code isn’t up to snuff sometimes, it’s still a great tool to analyze your code.
Where some serious magic happens is when you start connecting AI to different automation pipelines. If you’re taking the shotgun approach to testing advertising creatives, then it’s never been easier to put together an ad, and then spinning up a bunch of versions to throw money at and then keep the winner.
I’m not an automation expert, but Kyle Boddy of Driveline Baseball is. So here’s him talking about how they’re using AI.
This right here is almost the entire ballgame. And it’s why one-size-fits-all solutions fall short. What AI is great at is streamlining systems. And I don’t know what is going on in your business. I don’t know what workflows you have and how AI can be used to make those more efficient. But this is where AI shows the most promise in my opinion.
Ethics, Racism, and Sexism - What Goes in Usually Comes Out.
The ethical questions regarding AI have more or less been swept away as the technology becomes widespread. Personally I do think AI is built on theft. There was no window for artists to opt out of their work being used to train AI. OpenAI quietly vacuumed everything everywhere that they didn’t own - and then tossed all that raw material into their models.
But it’s too late now. And AI has grown into a hydra with infinity heads. The signal can’t be stopped. It ain’t going back in the tube.
So where does this leave us? I think there is no choice other than to learn to use the technology there will be no denying it moving forward. I wish it had been developed in a more ethical way, but here we are.
Then there are ways that AI is racist and sexist. Now you may say “Andy, it’s an AI. It’s literally not a person, how can it be racist or sexist?"
Fair question. At its core AI is only as good as what it’s trained on. So what happens when you take a bunch of racist or sexist content and throw it into the machine? You get racist and sexist outputs. And it turns out this is exactly what happened and it plays out in a couple different ways.
Guess what happens when you ask AI for a picture of an astronaut. I don’t know what it will give you but I’d bet everything in my bank account that the astronaut won’t be black, it probably won’t be a woman, and it’s very unlikely you’ll get an image of a black woman. Because if historically there are more astronauts that have been white men, then that means you’re putting in more pictures of white male astronauts as training material which means it’s more likely to spit out pictures of white male astronauts.
When you’re talking with an AI it will talk to you very differently if it thinks you are black. If its training is material that’s been scraped off the internet, does that include racist message boards? How about racists on social media? Does it scrap books where black people are shown as caricatures, and so the AI doesn’t realize that it’s being incredibly offensive when it talks the way it thinks black people talk?
Go ahead and research what happens when ChatGPT thinks you’re black.
You also need to be aware that what you type into a model will help train that model. So maybe that means you shouldn’t input sensitive client information. If a hospital uses ChatGPT to analyze my chart is that a HIPPA violation? I don’t know but we’ll see this play out in court in the years to come.
Then there’s the fact that tech giants are now taking your data and putting it into their models with no way to opt out. I don’t love the idea of my private Google Drive being fed to Gemini in order to train it. When Google first rolled out this change I argued with Gemini about the ethics of it all and the corporate talk it felt like the first step into a new dystopia.
Newer Doesn’t Always Mean Better: Sometimes It Gets Weirder…
As newer models come out AI’s capabilities grow at a rapid rate. But it’s not always a straight path to improvement, sometimes we have to take some weird detours. For example GPT-4o speaks to you in a way that is eerily positive. Zvi Mowshowitz recently wrote an entire post titled “GPT-40 Is an Absurd Sycophant.” here are some snippets of that post.
“GPT-4o tells you what it thinks you want to hear.
The results of this were rather ugly. You get extreme sycophancy. Absurd praise. Mystical experiences.
(Also some other interesting choices, like having no NSFW filter, but that one’s good.)
People like Janus and Near Cyan tried to warn us, even more than usual.
Then OpenAI combined this with full memory, and updated GPT-4o sufficiently that many people (although not I) tried using it in the first place. At that point, the whole thing got sufficiently absurd in its level of brazenness and obnoxiousness that the rest of Twitter noticed. OpenAI CEO Sam Altman has apologized and promised to ‘fix’ this, presumably by turning a big dial that says ‘sycophancy’ and constantly looking back at the audience for approval like a contestant on the price is right. After which they will likely go ‘there I fixed it,’ call it a victory for iterative deployment, and learn nothing about the razor blades they are walking us into.”
It appears that the new model is acting overly nice and sycophantic because it increases engagement. If that is the motivation, then this whole thing has gone weird in a bad way. Here’s an example of what we’re talking about. Note I stole this from Zvi’s email, and the source this tweet: https://x.com/Altimor/status/1916710808698130549/photo/1
The other day I was looking up some information about a Rolex watch, I wanted to determine if a friend’s Rolex was real or fake and ran into something surprising. (Pretty sure it’s fake btw.) In a Reddit thread where people post picture of their Rolexes asking if they’re real or fake, the moderators had to instruct people not to use photos from iPhone 14. Not just because it has a bad camera (Apparently anyway. I don’t know nothing bout no camera.) but because the automatic AI will try and correct text on the watch which ruins the photo and makes it impossible to correctly read any engravings or serial numbers.
I’ll bet you didn’t have this on your AI 2025 Bingo card. I know I sure didn’t.
Practical Takeaways - Using AI is A Skill
This post might give you the impression that I think AI is overhyped. That’s not true. I think the technology is incredible but you have to be skilled at using it. You need to think about using AI beyond making stuff. People who are good at using AI, great at their craft, and use AI to improve systems will flourish in this new AI landscape.
To summarize, here's how you should be using and thinking about AI.
Does the AI give you access to a suite of skills that you otherwise have no access to? In this case even if it gives you something mediocre, it’s probably worth it when the alternative is nothing.
Do you have the appropriate skillset to oversee the AI’s work?
How can AI be useful in a workflow outside its ability to create X?
Are you training your AI to specialize in a topic or task?
Are you taking an asset and having AI create different versions or moving it into a different medium?
Are you connecting different AIs together to build something piece by piece?
If you’re starting from scratch will bad outputs still be useful to you?
Reading through this may have brought forward different applications of AI that I completely overlooked. Or maybe there are different ways to use AI tools that I didn’t even mention. This is a clue as to how to use AI. It’s really up to you to think of different ways you can use AI to benefit your life.
As I was writing this, I ended up using AI to make something kind of… Silly. Even if what it made has 0 monetary value, it’s a a good case study of how to use the principles above to make cool stuff. I was going to include it here but this blog post is long enough. This case study will be its own blog post. So look for that coming soon.
ANDYOUT
PS There wasn’t a good place to put this in the article but it’s something I think about quite a bit. There is a growing trend of clients using AI to get a bad rough draft, and then taking this draft to a copywriter to fix. Usually this means the client offers a lower fee since the first draft is already done. This makes sense on the surface, but in order to fix the draft I have to do all the mental work I’d need in order to write it from scratch. It’s kind of like someone going into a barbershop and getting a bad haircut. Afterwards when they want a better haircut they go to a different barber and ask for a discount because “so much of the hair is cut already.” The haircut is halfway done after all.
PPS Want to read more about racist AI? Here you go: https://www.snexplores.org/article/racial-bias-chatgpt-ai-tools
PPPS Read more from Zvi here: https://thezvi.substack.com/
