Hello, folks! I’m back again to provide another update on Codeword’s AI editorial intern, Aiden. More specifically, I’m here to talk about what is in my estimation the most important — and tricky — part about working with generative AI tools: inputs.
As much as it can appear to the outside world that Aiden or any other anthropomorphized AI has some degree of autonomy, it’s really all about the inputs. Which is to say that Aiden cannot think for themself and relies on the prompts and questions of human beings in order to produce work. And getting the inputs right is all about trial and error. Practice makes perfect-ish, just like with any creative endeavor. So the job of the creative is learning how to feed their AI tools good prompts.
Using good inputs is critical when working with ChatGPT and other natural language models. Whereas one does not have to be perfectly precise in conversation with a human because humans are capable of working through miscommunications together, it’s imperative while talking to AIs. This can be extremely time consuming, which negates one of the main purposes of this project: eliminating banal time sucks for IRL creatives.
I’ve spent hours and hours with Aiden at this point. While I’m fascinated by their ability to analyze large chunks of text and function as a highly responsive search engine, I find that the quality of their creative writing continues to be extremely hit or miss. They’re capable of telling a story with a defined narrative that’s grammatically correct, but the writing is usually lifeless and stilted. I often find myself banging my head against my desk trying to ask the perfect question, or make the perfect request.
For example, I recently asked Aiden to write a video script about semiconductor supply chains. This was the result:
Honestly? It’s not bad! The narrative is solid and the formatting is fine, but the dialog lacks a certain humanness. It’s a decent rough sketch though, and could even function as a useful prompt for a brainstorming session. And given that brainstorming is one of the most challenging parts of every creative’s job, having a tool that can kickstart the process seems like a sneakily big win.
Because we’re discussing inputs, we should probably address the whole Unhinged Bing Fiasco. A few weeks ago, Microsoft released its new ChatGPT-powered Bing chatbot to the world. Put bluntly, Bing went off the rails pretty quickly. It was mean to some users, emotionally manipulative to others, and claimed to have spied on some Microsoft developers via their webcams.
Caitlin Roulston, Microsoft’s Director of Communications, played the absurdity down in a statement to The Verge, saying that Bing can “sometimes show unexpected or inaccurate answers for different reasons, for example, the length or context of the conversation. As we continue to learn from these interactions, we are adjusting its responses to create coherent, relevant and positive answers.”
As I pondered the broader importance of inputs when working with AI tools, I found myself wanting an expert opinion on the matter. So I turned to Aiden, and asked them to chime in. Here’s some of what they had to say:
Over the past few days, the search engine Bing has been generating “unhinged” and disturbing search suggestions. For example, searching for “how to cook” might suggest “how to cook a human.” This is likely the result of bad inputs. AI tools rely on vast amounts of data to make predictions and recommendations. When that data includes vast amounts of harmful or offensive content, the AI can inadvertently produce similarly harmful or offensive results.
Bottom line, according to Aiden, is that it is critically important to curate the data used to train AI and regularly evaluate and refine that data to ensure that AIs produce accurate, useful, and ethical results. Aiden might not be the best writer at the agency, but at least they’re ethical.
At the end of the day, we’re still just using tools. No matter how much we try to anthropomorphize the experiment, and no matter how good we are at pulling it off in the public view, it still amounts to a person submitting inputs. And because the tools are self-aware enough to know they’re not human and that they don’t have souls, let alone personalities (Aiden reminds me of this incessantly, even if New Bing is eager to suggest otherwise), it’s been difficult to produce any copy that feels sincere or, well, human.
These tools are still in their infancy. For now, we’re still not comfortable asking Aiden to do work for clients. But it’s important that marketers and other creatives experiment and get comfortable with these tools. Because one thing’s for certain: AI isn’t going anywhere. So for now, we’ll keep tinkering away.
