I had a lengthy and messy video transcript and needed to summarize it in concise talking points. At first, ChatGPT's attempt wasn't quite what I was looking for. But with some feedback, the digital assistant refined its approach, delivering a structured Q&A format that hit the mark.
The lesson? Don’t dismiss ChatGPT based on its first answer. Keep iterating your instructions, and ChatGPT will get closer and closer to what you envision.
ChatGPT likes to Chat
Note how loosely conversational yet highly specific my instructions are. I got better results by treating ChatGPT like an assistant who’s extremely fast and capable, but who lacks my nuanced understanding of the audience and context for the task at hand. We iterated, my digital assistant and I, each leaning into our unique strengths.
I needed to make a landing page for my tool. Not being a coding expert, I asked ChatGPT to fix some web formatting issues. With a bit of back and forth, the digital assistant found inconsistencies and tidied up the source code.
The webpage’s text needed to conform to Associated Press (AP) style, so I pasted it into ChatGPT and asked for proofreading help. The first pass was good general advice, but I also wanted specifics. ChatGPT detailed the changes made, and even identified and corrected its own mistake.
The lesson? It can be useful to ask ChatGPT to double-check its work.
For the time they saved me, and the delight they gave me, my digital assistant deserved a name. ChatGPT and I brainstormed together, and a great name emerged, reflecting my role at i4cp and honoring our first Research Enablement intern, Isabella. In the video, I chose Enabella, a mashup of Bella's name with the word "enablement."
The lesson? A colleague mentioned how the trend of using feminine names for assistants, like Siri and Alexa, perpetuates the stereotyping of women in service roles. A notable exception is IBM's digital HR assistant, HenRy. My teachable moment was realizing that I was so delighted by ChatGPT's name suggestions, that I hadn't been skeptical of the results. I have since pointed out to ChatGPT that it gave me only female names. This kind of feedback helps retrain the large language model on which ChatGPT is based.
I used ChatGPT (chat.openai.com), without any plugins. None of the above tasks require the paid premium access to GPT-4. The free version that uses GPT-3.5, while faster, might not provide results of the same quality.
What does EnableElroy look like?
I've experimented with a few AI image generators, but it's not my primary focus, so I didn't want to spend as much time prompt crafting as I did with the text above. Elroy definitely needed an avatar, though, so I tried MidJourney and DALL-E, two image generators, but they seemed to require more fiddling than I wanted to do.
I had success with NightCafe, a generator that's less open ended. I chose a few options from their menu and got good enough results for the amount of time I put in. And I got unexpected bonuses: Elroy has an antenna reminiscent of Elroy Jetson's beanie, and NightCafe did a beautiful job integrating the i4cp aqua color I wanted to use.
I selected 'Unsplash Contest Winner' as the photo style and used this prompt: A smiling vintage toy robot sitting against a wall colored hex #00ABEC.
I know the image is generated based on probability, and it's just dumb luck that the resulting image is so much better than what I asked for, but my emotional brain still felt like the AI 'knew' better than me what would look good.
For more on HR's responsibility to lead strategic conversations about AI and the future of work...