AI Use, a Year Later

I last wrote on generative AI and its use a hair under a year ago, decrying its sloppish output, its climate and water impact, and its foundational thefts of intellectual property and creative labor.

Much has changed since. A year ago Zello had a scattering of internally influential early adopters. This worried me, but I wasn’t sure where they’d take us. In my limited and pedestrian product thinking, I couldn’t imagine what use walkie-talkie users would have for AI-labelled, headlining, highly-marketed features.

In January the company made major strategic bets on AI as a productivity multiplier for frontline workers, at the same time using the increasingly popular rhetoric: “this is the future” and its implication, “don’t be left behind.” Like many others,^[1] we also had a push from leadership to increase internal AI use within the company, from development to Sales, Engineering, Design, and Product. As (then) Director of Design, I was directed both to find ways to use generative AI in my own work and my team’s.

On the receiving end of coercive, fear-mongering language like the above, I dig my heels in. Then on a push to eek further productivity from a hard-working team,^[2] I dig my heels in further. Finally, following an employee engagement survey where I certainly wrote more than I should’ve, I found myself treading water East of Madagascar after plummeting through the Earth’s core and out the other side.

Later that month I went out for drinks with our COO, who shared necessary market perspective, and allowed me some insight on our team’s planning talks. This helped me to see the value they envisioned, and while it matters less to the frontline workers I’ve invested in, I see the business opportunity.

Over the following months, generative AI use spread further through the company, its tendrils snaking in both where prominent and remote. Walking from my desk on one end of the office to the kitchen on the other, it’s common to see ChatGPT used for code references, content summarization, and copywriting. Some use it for research and brainstorming, and many of the early adopters use it in lieu of Google’s search engine.

My early objections aside,^[3] I worry about ChatGPT use for the basic tasks expected of our professions. A seasoned engineer can use Cursor to scaffold and even write extensive code — they know how, so they can proof the work — but when a junior follows that example, I worry they’ve lost a learning opportunity.^[4] I then worry about the seasoned engineer’s skills atrophying. These examples extend beyond the quantifiable in code output — I hope in a few years, folks will still write their own emails.^[5]

In my own use, I’ve had to weigh principle against my security and my team. Insisting on Luddism paints a target on my back, and may on my team’s as well. A principled refusal — even one built on both a concrete fear of climate impact and a respect for the tools’ stolen labor — would’ve not only hurt my options, but hurt my team by association, and risked unchecked and unguided use of generative AI tools overlapping with our work. Early in the year we clamped down on the use of image generators for unauthorized brand assets and illustrations, and so revealed the need to preempt further adoption.

Encouraged to find uses for myself and my team, I found tools like Kami to expedite our prototyping, and team members started using Cursor for work on our marketing website. I’ve used it to scaffold and even draft javascript that I didn’t want to take the time to write, and to refactor my own rough code. I built out Origami systems for use with Whisper and ChatGPT, allowing us to prototype with text to speech, translations, and a conversational partner. Working with a live voice communications tool, it’s just cool to prototype a conversation, as opposed to using prerecorded audio responses. When GPT-4o‘s image generation was released, I created a GPT to build off our illustration style to see if we could use it to further our visual design output.^[6] I’ve since then used the agent to generate large illustrated slides when pressed for time on our monthly product updates, and it does so impressively. I later started using it in my UX and documentation work. Were I to have five obvious solutions to a problem, I had tried using ChatGPT as a discussion partner to see if there were a sixth. This rarely yielded meaningful output. Lastly, I’ve started using ChatGPT to format longer bodies of content in ways my colleagues find palatable. Where I previously would have sent a Wikipedia article, or pull quotes from a research paper’s abstract, I’ve used ChatGPT to format content so that it’s more quickly consumed by colleagues. As folks grow accustomed to ChatGPT’s predictably structured output, I’ve found many colleagues preferring it to other reference materials.

As a visual designer, I’ve toyed with 4o’s image generation, and its novelty and coolness trumped my principles of a year ago. I’ve used it to apply emoji and Memoji aesthetics to my sketches, and then used those as inspiration for my own illustrations. I’ve used it to preview different palette, textural, and light treatments on my own near-finished illustrations before I take the trouble to laboriously iterate. Of course I’ve created Studio Ghibli illustrations from selfies, with the caveat, reminiscent of earlier strong-held principles, that I was doing this to see and test for myself, and not to share, therefore diluting Ghibli’s brand and IP. Vaguely echoing my sentiments a year ago, I refuse to use any generated imagery as-is, and I’m only limitedly comfortable using its output to inform my own creative work. Here its steady incentivizing and adoption wheedle their way in: I need to use it professionally, this is a way to do so, and I keep telling myself “left to my own devices, I wouldn’t.”

And then, I did. I’ve been using ChatGPT to challenge my written and spoken French, inputting my own writing and requesting grammatical and stylistic critique. This evolved to more targeted prompts when T-Mobile’s French data roaming has failed to load WordReference^[7] and BonPatron: “What’s a more natural way to say « qqch en français » and why?”

Looking back, what started as a gentle slope quickly became slippery and then sheer as I plummeted down. None of my objections from a year ago have lost ground, but professional encouragement^[8] moved me to humor what I’d found deplorable, and I increasingly gave into the cool factor. Recent past trends — Crypto, NFTs, etc etc — lacked appeal largely in that they were for consumers and true believers. Crypto bros tried applying their passion to everything without a meaningful problem to solve, and NFTs were for hucksters, new-money wannabe art buyers, and of us artists, class traitors. AI, while often a solution in search of a problem, has cool, novel, and fun elements that when allowing me to build something new is attractive.

Whatever fire I once had to push my politics in the workplace has lessened in this second Trump administration, and my desire to argue and fight is nil. I like my job and employer, and despite this growing disagreement, I would truly rather build communication tools for frontline workers than anything for comfortable individuals like me, enjoying high salaries at a standing desk in their air-conditioned office. That said and having taken a moment to pause, I dislike the person I’ve become in regards to my own gradual AI-adoption, and I need to step back and reevaluate. Zealotry is challenged by hundreds of tiny, seemingly inconsequential opportunities to bend, and I’ve found myself wrongly more excited by finding productive and acceptable uses than by enforcing my original stricter principles.