ChatGPT and other generative AI systems made machine learning a media darling in late 2022. By then, Thread had already been testing and working with artificial intelligence to assist our copywriters for a few years. We were interested in its potential to reduce copy costs for our clients as well as production timelines for ourselves. But the results at that time were consistently disappointing: product descriptions that were generic, repetitive, randomly inaccurate and lacking in human warmth or empathy.
Here’s a typical example of repetitive AI copy from the early days:
This men’s tank top is made of sweat-absorbing fabric that keeps you dry as it wicks moisture away from your body.
This one got a little confused about anatomy:
Get around town in these durable and comfortable pants. With preshaped elbows and knees for a natural range of motion.
And then there’s this slightly nightmare image:
Wherever you go, your sweat top will follow. That’s because it will never stop hugging your body.
In 2020, we partnered with an AI startup that focused specifically on copywriting. Their product was still in development, but our preliminary tests showed their AI had the potential to create a level of satisfactory content and get us to our goals over the long run.
That summer we launched a pilot with a small set of products in partnership with a longtime client, a global sportswear brand. The goal was to better understand how the tool could work with a range of products and varying levels of information inputs. And, of course, how AI copy compared with traditionally written copy.
The AI pilot began with about 500 basic products — things like cotton tee shirts and non-technical sneakers. Initially we simply wanted to know if the tool was viable. (Short answer: yes.) We were also curious how the tool would interact with the data systems that feed our client’s product information to us. (Answer: Good, brief data in = more accurate copy out.) Later, we wanted to see how the AI would perform at scale, with more complicated products.
At first, the tool made some adorable missteps — it would misrepresent shirts as shorts, or even shirts as shoes. And sometimes the copy sounded like Jack Handey’s Deep Thoughts. It had trouble understanding and communicating why a consumer would buy a product — an understanding that’s fundamental to Thread’s experienced copywriters. This may never be obvious to AI on the level of individual products, though it can accumulate and apply general principles — that cotton fabric usually feels soft, for example, or sweat-wicking fabric keeps athletes dry.
As we continued to feed revised copy back into the tool, it started to get more competent. A few other things helped: First, the team began to organize the data the AI was reading in a more strategic way. In turn, the copy became more accurate and better representative of the product and writing we strive for — conveying in a relatable way why this product is the right one for the consumer considering it. Second, the AI tool began tracking word choice and phrasing for the brand it was writing for. This helped it be more consistently on brand. Finally, our start-up AI partner was constantly improving its writing model and incorporating those updates into the tool.
Here are examples of rejected AI copy that nevertheless secretly gave our writer brains joy:
This new shoe moves with you as you keep going, which is why it’s perfect for the people on the move.
Just in case you weren’t paying attention, you need to be. Your game needs to be on point, and your look needs to be flawless.
[Describing a woman’s shoe] Every day is a good day to pack all the dad vibes possible.
Contrast them with these later examples, improved by training:
Keep the sun out of your eyes while you train. This visor is engineered for stretch and durability to last no matter how you move.
These high-performance socks are made with lightweight and sweat-wicking fabric to keep you dry and comfortable while working out. An embedded arch band helps provide a secure fit while compression around the ankles allow for a snug fit throughout the session.
This junior girls’ crop top gets a preppy, collegiate makeover. Made of soft and stretchy single jersey, it has a slim fit and cropped silhouette.
The copy we asked our AI tool to write needed to be about 300 characters long and include both a description of the product and a couple of basic SEO keywords. The descriptions needed to be accurate. And they needed to be true to our client’s voice.
AI takes direction like a champ, but it can’t think independently. It’s good at describing basic products, but it can get confused when there’s more complexity. Yet over three years our pilot proved that AI is a tool that effectively supports human efforts.
Here are five takeaways from the pilot:
When the pilot began, we called our AI tool Rosie, nicknamed for the Jetsons’ robot maid. Rosie was a great companion. She trained our team how to work with a machine writing assistant while our team trained her — a relationship that furthered Thread’s growth as well as Rosie’s. Rosie brought us so much learning, so many ideas. But now we’ve outgrown her. We’re saying goodbye to the Rosie writing tool. We’re not saying goodbye to AI, though. That journey, even after three years on the road, is just beginning.
Drum roll, please: Thread is now a Certified B Corp! We 👏 are 👏 stoked 👏