The recent Olympics Games in Paris were quite the spectacle. Beyond the memes about shooting stances and poor conditions at the Olympic Village, it was still a great showcase of competitive spirit and talent. Trying to capitalise on the hype around the Games, Google released an ad promoting their Gemini AI chatbot featuring American hurdling and sprinting star Sydney McLaughlin-Levrone (video link below). The ad’s messaging means well, but something about the execution just doesn’t hit right – especially when it’s a company as savvy as Google! Join Wordsmith as we explore how we can patch things up.
What went wrong
Google’s ad features a father recounting his young daughter’s passion for hurdling, just like McLaughlin-Levrone. To show their support for their favourite celebrity (and to get in a cheeky watch your back, we’ll be gunning for your spot attitude), the father tasks Gemini AI with writing a fan letter to McLaughlin-Levrone.
Like most chatbots, Gemini AI is very capable at following your prompts – it’ll compile together sources based on your prompt, then arrange them logically into a narrative. From writing college thesis essays and CVs to beef wellington recipes, modern chatbots can put together just about anything (although how well they do it is often up for debate).
Writing a letter to your favourite celebrity definitely falls within the AI’s capabilities, so the problem with Google’s ad isn’t false advertisement. Rather, it’s more of an ethical paradox. Fan letters are generally personal and intimate experiences – from one heart to another. As you’d expect, not all letters are going to have perfect grammar, flow properly or even make sense, but that’s part of the charm of sending and receiving fan mail. Getting an AI to do this simply removes the “humanness” of the letter.
Anthony Ha (an editor at TechCrunch) comments that writing a letter can be a difficult or intimidating process… especially when the recipient is someone as prestigious as McLaughlin-Levrone. While AIs are pretty good at getting initial drafts out, they are far from perfect and require meticulous editing. Google’s ad never mentions this step and makes it seem like the AI’s first draft is already perfect.
“But if this happened in reality, Sydney would just end up with a giant stack of nearly identical letters,” explains Ha. This is because AI’s prioritise efficiency over uniqueness. Unless you directly specify it to write a certain way, it’ll always use a default structure and tone – imagine McLaughlin-Levrone’s disappointment if she opened identical letters that all read and sounded the same!
Let’s fix that
The all-important step of working with ChatGPT or any other AI writers is to ensure that we have a proper prompt. Here’s the prompt the dad uses in the ad:
“Help my daughter write a letter telling Sydney how inspiring she is and be sure to mention that my daughter plans on breaking her world record one day. (She says sorry, not sorry.)”
Is this prompt good? Not at all.
1. Does the father and daughter have a name? It’s hard to start a letter to a stranger without introducing yourself!
2. The recipient’s name isn’t fully written out. Based on the prompt, the AI must determine who “Sydney” is using the descriptor “world record”. Although the AI is probably smart enough to determine the intended recipient is McLaughlin-Levrone, imagine if it pulled up this article as a source and wrote to a different Sydney instead – and because the Sydney in the above article does basketball, track and cross country, the AI would work around those sports instead of hurdling!
3. The current prompt gives the AI too much creative control. Who is this guy’s daughter? Why does she want to become a hurdler and what does the sport mean to her?
Furthermore, without specifying how or why the daughter found McLaughlin-Levrone to be inspiring, the AI will resort to making stuff up. Maybe the AI will spin a sob story about the daughter choosing to become a hurdler despite being handicapped, or because she’s a dropout from school or because she lost a cherished relative – a bit morbid, but it’s a possibility within the AI’s storyboarding logics!
4. Are we writing from the perspective of the daughter or the father? If writing from the daughter’s perspective, what is her reading level and how does she write? Without specifying a writing style or tone, the AI is going to write like… well, an AI. Have you ever seen a young girl write like a college graduate? Unless they’re a prodigy, an uncanny level of eloquence and vocabulary is a solid giveaway of AI use.
5. What about the dad’s opinion on his daughter’s goal? Since the dad is writing and sending the letter on his daughter’s behalf, his input should matte as well.
6. There is no specified call-to-action. Do they want McLaughlin-Levrone to write back? Or are they satisfied with just having the letter written and sent? Any true fan would be ecstatic to receive a reply from their favourite celebrity, so being nonchalant about this seems kind of disingenuous if you ask us!
Unfortunately Gemini is not available in Hong Kong, so we used the next best thing – ChatGPT-4o. Here’s what the original prompt from the father got us:
Compared to our revised prompt and output (names and details are fictional):
“Write a letter to Sydney McLaughlin-Levrone for my daughter Jane Doe and me, John Doe. She’s currently 12-years old and doing her best to follow in the footsteps of Sydney as a future Olympic hurdler on Team USA.
Jane wanted to be a hurdler ever since she first saw Sydney compete. In my daughter’s eyes, to be able to sprint clear each hurdle while maintaining perfect stride and speed was like the most beautiful ballet recital. I’m not even into track and field and I was bloody impressed!
Jane is currently among the top 5 runners in her school’s under-16 track team and I couldn’t be more proud of her. She also says she’s aiming for Sydney’s World Record (“Sorry, not sorry” in her words). Although Jane wants to ditch school and focus entirely on hurdling after high school, I talked her into aiming for a sports scholarship and getting scouted at college-level track events. As much as I believe in her raw talent and drive, I don’t think it’d be a good idea to put all eggs into the same basket. I hope Sydney understands what I’m getting at and reflects this as well if she writes back to us.
Lastly, please ask Sydney what her daily routine is for training. Jane wants to do the same but with one more rep so she can say she’s training slightly harder than her. Thanks ChatGPT!”
The difference in quality is night and day. Notice how in the original prompt that the AI couldn’t even discern that Sydney meant McLaughlin-Levrone and used a very cookie cutter format applicable to any athlete – this is ChatGPT-4o by the way, the latest and supposedly most human model publicly available. Maybe Google’s Gemini is smarter, but leaving out key details is never a good idea in the first place.
Whenever you use a chatbot, you must remember that the quality of the AI’s output is directly proportionate with the effort you put into the prompt. There’s more to using chatbots than just casually slapping in a sentence for a prompt – otherwise, you’ll really have your work cut out for you when it comes time to edit!