Today I was writing an interview summary in Greenhouse, the recruiting tool, when I noticed one of those little “pixie dust” icons I have come to understand means “AI here!” It made me curious, so I carefully saved my work elsewhere then clicked the button.
Because I sit on my company’s AI governance committee, I have spent a lot of time over the past 18 months or so wondering what that button does in its many manifestations in all the apps where it appears. Sometimes it suggests some stuff you might want the LLM to do for you, like help plan a project, and other times it just does some sort of rewrite based on … ideas? … about “helping” you with your writing.
I know some ESL folks who like the pixie dust button when it stands for “writing aid,” because they want help with their writing. Because they’re on their second or third language, while they may not be experts about English they have a proven skill in language acquisition, so they’re not naive about what the button means. Not nearly as naive, at least, as people who speak one language and don’t give a fuck about that one, let alone any others. So they scrutinize the output. One of them consults unfamiliar words in the dictionary before accepting them, but appreciates the help with basic structure.
I also know people who want help with their effectiveness. Like the ESL people I know, they’re coming into that interaction less from the perspective of “I barfed some shit into a form and the robot made it good.” They’re coming at it with a use case like, “my boss is think-skinned, defensive, reactive, and mildly paranoid so how can I tell them something they don’t want to hear without triggering an episode?”
Having typed in what they want to say, then prompting for how they want to be received, when a more diplomatic revision comes out of the little pixie dust slot, they read it carefully, think about using it, then maybe massage it a little, or just use it as a suggestion. In other words, they apply their judgment, but benefit from the slight frame shift to imagine better alternatives.
One person I’ve talked to who uses AI in that mode has had a lot of success with the approach. It takes a little extra time because she never just copy/pastes the output, but it has gotten her good results: She truly struggled to express herself in a way that her thin-skinned, defensive, reactive, mildly paranoid boss could hear, and now she’s getting better outcomes in those interactions.
“I feel like I’m cheating,” she said. “Like I’m not doing my job.”
I said, “you’re a senior director for one of the largest Medicaid consumers in the state, and your job is to communicate effectively enough that your 100 employees can get shit done without you triggering a diplomatic incident or your boss losing her shit because you decided fuck it, she doesn’t need to like what she’s reading.”
So I don’t really judge the pixie dust button or the edits it sometimes suggests to be an unalloyed social ill.
It was in that spirit I clicked the Greenhouse button.
I was hoping it’d be something novel or different. Like, maybe it would offer some insight particular to this kind of writing, or would analyze my commentary for unconscious bias, or even warn me that I was about to commit a hiring practices violation to a corporate system of record. Instead it just mostly repeated my writing but replaced a few turns of phrase I chose very intentionally with more cliched ones that altered my meaning.
I don’t know how many hires I’ve done, nor how many hundreds of interviews I’ve conducted. Lots. Lots and lots. When I’m the hiring manager, I’m a close reader of interview notes. I pay attention. When I’m on someone else’s panel, I write carefully and want to be clear.
Today was the first day that I acutely felt the impact of AI fucking with two things I care about: Hiring and communicating. It took my words and Business Englished them into something other. It altered meaning, and to no good end. Nothing was more clear. There were no syntactical, grammar or punctuation issues to call out. It just, I dunno, blorped them. There’s no word for it. It didn’t, like, insert sexist slurs or generate a picture of a Black Nazi. It just sorta gnawed the input like a slobbery, toothless dog and left a saliva-covered pile of verbiage normalized to some inscrutable standard. It was sort of like how when you have shake reduction on a camera, you’re supposed to turn it off when you put the camera on a tripod because it counter-intuitively risks blur: Feed the LLM good enough English, just not statistically congruent English, and it’ll just sort of cock its head then spit on your sandwich because it’s supposed to do something when you click the pixie dust button.
So my mind went immediately to some recent examples I’ve caught go by in other parts of my job where someone was plainly hucking their shit into a form and pressing the pixie dust button. The tells were the tells we’re all learning to spot – the sorta flowery lilt, needless explication, and obtuse summarizing; the gratuitous length that reminds me of early efforts to use CG in superhero movies, where it’s just obvious that it’s CG because human bodies just can’t articulate like that, and would rip their own arms off or snap their own necks moving with that kind of speed. Like, nobody in their right mind, asked for a quick outline of a modified five step process you could fit on half a sheet of paper, would spend the time it took to extend that to five pages.
But as much as it was a waste of my time, it was easily addressed because it was obvious and gratuitous. The perp owned it, reminded me of their insecurities about writing, and agreed that they’d lost sight of the problem they were trying to solve because for a few seconds it felt really good to not struggle quite as much to make words come out. I don’t think it was meant to be a knife-twist that they said they fed it some of my writing as a guide.
And for a lot of business writing tasks, I worry less. If my helpdesk manager wants to send a quick nastygram to three engineers who won’t send back their old laptops because they’re too lazy to get their shit off of them and put them in the return box, I don’t care. “My director is going to tell your director, then he said I can lock that fucking thing and delete all your shit” is going to be clear no matter what the pixie dust button does.
But I benefitted from Puppet’s outstanding emphasis on interviewing and hiring a long time ago, and every detail of those things has mattered to me ever since. I’ll cut corners, let myself have bad days, fudge, procrastinate, take unwarranted mental health days, accidentally nap through the first 30 minutes of a QBR, and generally be a slob about a lot of other things, but interviews matter. Period. They’re an absolute. When I have one coming up, I pull my shit together and go do it. When I sense someone else isn’t invested, or they’re phoning in their feedback, or they didn’t pay attention during the kickoff, I judge.
Greenhouse provides a button to fuck with something that matters. A kind of writing that matters. It was just a stupid fucking five-pointer to some dev, plopped in the backlog by some product owner who might not even see the sense in it but knew they needed a third bullet for the AI features list, mandated by some director of product who got it loud and clear from someone above them that the market demands moar pixie dust buttons.
Anyhow, it was sort of clarifying. I’ve been a little paralyzed when I get the weird AI documents, because sometimes they’re non-objectionable. Maybe too long, or too wordy, or with odd little nuances that aren’t wrong but are off, but I ask myself “for everyday dumb bloodless business writing meant to detail a process, explain a system, or document some steps, is this okay?”
Usually yeah. It’s fine.
But this got me off the fence a little. I know some people like to announce total bans on AI assistance. I think that’s risking losing sight of some things, and I actually can’t because we’re under a mandate at work to use exactly the kinds of integrated assistance these things offer via all the Google Gemini integrations. (I did overhear a team mentioning that something they were given to work with by their leader would need careful review “because it came straight out of AI and probably has mistakes,” and that made me cringe.)
So I just told my folks that I’ve needed a reason to follow the thirty percent rule a little more, anyhow – ask for review when you think you’re thirty percent done – and that I want the 30 percent they put in front of me to be just them, not a bot. I just want to know they thought about the core of the thing they’re doing, and that they’ve chosen a direction we can discuss further if needed, before pressing the pixie dust button. If that runs afoul of the “use this shit” mandate, because we’re also supposed to be offloading our thinking, I’ll see ya on the soup line.
If I could burn that button out of Greenhouse though, I totally would.
Coda, bright spot: I listened to a C-level exec from a major software company say that the biggest challenge they have with AI features is that uptake is really poor. People don’t like a lot of it, or technology leaders are resistant. That sort of tracks with the tools I’ve got in my portfolio at work. AI was pitched as an expansion or upsell a lot last year. This past several months, three pretty big ones I manage just said “fuck it, you win,” turned the features on for me, and rolled them into the main SKU without charge. That’s some money and time down a hole.