You’re Probably Using AI Wrong

Using AI is an exercise in meta-thinking. At Every, that often means writing about what it means to write with AI. Among our recent pieces, we've talked about how to build an AI writing partner, the benefits of using AI voice-to-text tools, and how to preserve a unique writing style while using AI. Today we're re-surfacing a perspective from Rhea Purohit, who after struggling to work AI into her day-to-day routine found success when she began using it not for high-volume productivity, but to amplify meaning in her work and life.—Kate Lee

Was this newsletter forwarded to you? Sign up to get it in your inbox.

I tried to use AI to be efficient—and I was disappointed.

I’m a writer. LLMs can generate coherent passages of text faster than me. Getting AI to write for me is arguably the most efficient way I could use it. But each time I prompted an LLM to do this, the writing was flat, bland, and impersonal, so much so that I didn’t think it was worth using, not even as a first draft.

I kicked, screamed, tore a few strands of hair out. And then I realized that the problem wasn’t AI—it was me.

More specifically, my narrative around how AI fit into my work and life.

Most consumer AI apps share a common promise: They help you do things faster, or with less effort. The resulting narrative frames AI’s primary role as boosting efficiency. Some people might find meaning in efficiency itself, but many others don’t, which can inadvertently discourage them from experimenting with AI.

If you, like me, are part of that second group, I have a solution for you: Reframe your narrative around AI. Stop using AI to be more efficient. Instead, use it to increase the meaning you derive from work and life.

As humans, we are drawn to activities that give us meaning. In some domains of life—like, say, mountaineering—this may seem obvious. As an example, take the Eiger, a 13,000-foot high peak in the Swiss Alps.

The north face of the Eiger—a vertical wall of ice and limestone—has earned the moniker Mordwand, German for “murder wall,” because of the number of people who have died trying to climb it.

A third of the way up the jagged mountain, at around 9,000 feet, is a door.

On either side of this door is a different reality. Inside the Eiger, trains rumble through tunnels, plying passengers to tourist attractions. Outside, the door opens onto a narrow ledge where numb-fingered climbers pause for a moment of respite. (The two realities collide during emergencies, when stranded climbers need to be rescued.)

Two centuries ago, the only way to get to the top of the cloud-covered mountain would have been to scale it. Trains—an advance in locomotive technology—now provide an efficient alternative. Even so, you wouldn’t bother trying to convince a mountaineer to catch a train to the summit. Their decision to climb the Eiger has nothing to do with efficiency; it stems from the meaning they derive from the experience.

Humans are drawn to activities that bring them meaning. This is common sense when it comes to mountaineering, but my hunch is that we lose sight of it when we try to use AI.

AI beyond productivity

I like the process of choosing the words I commit to posterity on the internet. I don’t just like it, I cherish it. That’s why, when I outsourced this task to AI in an attempt to be more productive, I was left unimpressed, and gradually, I stopped trying to use LLMs altogether.

I wanted to have a better relationship with this new technology, and the first step to developing that was rewriting my narrative around AI. I stopped thinking of LLMs as tools that would write for me. Instead, I started using them to help me with a part of writing I don’t enjoy: being stuck.

In drafts, I like to articulate my arguments with examples. However, I sometimes struggle to come up with good ones. Pre-AI, I would turn to my partner, explain the problem in a garbled frenzy, and expect him to brainstorm with me. What followed was an erratic, volatile process. At times, the approach worked like magic, but other times, when he was busy or unable to understand the required context, it would fail miserably.

Now, I turn to Claude instead. LLMs are never busy, and they make do with the inarticulate context I include in my prompt—so when I struggle with finding an example, I paste the unfinished draft into Claude and leave a blank like this: “____” where I want the example to feature. Then, I prompt the model: “Fill in the blank in 10 different ways.” The output usually gives me fodder for new ideas. (And if it doesn’t, I just prompt it again!)