A machine can write a better wedding toast than you can. It can't write yours. On AI, the beauty of imperfect words, and why people still learn to talk.
Shawn Smucker recently published a short piece called Please Use AI. The title is a joke. [1] The piece is a list of things you should definitely hand to a machine: your meal plan, your next camping trip, the toast at your child's wedding. Next to each one he describes what you'd miss by asking a person instead. Call your friend who loves to cook, and you'll end up hearing about her father's cancer diagnosis. Text your friend who has fished every river in Pennsylvania, and you'll end up meeting him for a beer.
The part that stuck with me most was the wedding. Smucker imagines a parent who changed hundreds of diapers, who fed this child in the middle of the night, who cried when the child came home late because they were sure something terrible had happened. And he points out that we apparently don't want that parent's words at the wedding. We'd prefer "the sterile words of a machine that never lived, never had an original thought, never felt the pain of miscarriage or broken relationships or the joy of a friendship restored."
That line is the whole argument. The machine's toast will probably be better written. It will have a clean structure and a nice callback at the end. The parent's toast will ramble. Their voice will crack in the wrong place. They'll tell a story that goes on too long and forget the line they practiced.
And everyone in the room will remember it. Not in spite of the flaws, but partly because of them. The cracked voice is what twenty years sounds like. You can't separate the imperfection from the person, because the imperfection is the evidence that a person was there. A machine's toast has no flaws because it has nothing at stake. It never lived the thing it's describing.
Smucker ends on this idea: "hearing another human voice and somehow realizing the beauty of life is found in all of these subtle imperfections." I think he's right, and I think it's one of the most important things to understand about AI right now.
Given, not sourced
A reader named Elizabeth Harwell added a story in the comments that makes the same point from another angle. She'd called her grandparents for a peanut butter pie recipe. Her grandmother is nearly blind and her grandfather is nearly deaf. He read the recipe card over the phone while her grandmother shouted corrections in the background. She made the pie exactly as it was shouted to her, without checking it against a recipe online. It was, she wrote, "lovelier for the way it was given, and not sourced."
Given, not sourced. When you ask a machine for a recipe, you're treating the conversation as a way to move information from one place to another. Often that's exactly what it is. If I need to know the boiling point of water, I don't need to hear it from my grandmother.
But a lot of human communication was never mainly about information. The recipe was the excuse. The phone call was the point. Elizabeth could have found a better-tested pie recipe in ten seconds. What she couldn't have found online was her grandfather squinting at a card while her grandmother yelled at him from across the room.
Here's a test. Why does a handwritten note mean more than a printed one? It doesn't contain more information. It's usually harder to read. It means more because it cost something. Someone sat down and spent a piece of their life on you, and you can see the evidence in the ink. The effort isn't the wrapping around the message. The effort is part of the message, and sometimes it's the most important part.
That's what goes missing when a machine writes the toast. The problem isn't the quality. The problem is that the cost went to zero, and the cost was what you were trying to say.
The extreme case
I think about this a lot because of what we do at Parrot. We help people learn to speak new languages. And by one measure, that should be a dying business.
You can already hold up your phone and talk to someone in almost any language on earth. Translation earbuds exist, and they get better every month. If language were only a way to move information, no rational person would spend hundreds of hours learning Spanish. They'd buy the earbuds.
And yet people keep showing up to learn, every day. When you read why they're doing it, almost nobody says they want to transfer information more efficiently. They say things like this, from one of our reviews: "FINALLY I can express myself to my Mexican MEJOR AMIGA!" Or this one: "My Colombian native boyfriend has told me how shocked he is at how well my Spanish is."
These people didn't need a translation. They already had one. What they wanted was to be the ones talking.
Language learning is the purest version of Smucker's argument, because the beginning is so clumsy. Speaking a new language as an adult is one of the most humbling things you can do. You have the vocabulary of a four-year-old and the self-consciousness of a forty-year-old. You say things wrong. People slow down for you. You feel stupid.
And then something surprising happens: people love it. Watch a tourist try, badly, to order in the local language instead of pointing at the menu. The waiter's face changes. Not because the order was clearer. It was less clear. It's because the tourist paid a price to meet him halfway, and both of them can feel it. [2]
A translation app would have gotten the order right. It would have gotten nothing else right.
Near the end of his piece, Smucker writes about "the way that anything worth doing feels a bit clumsy and painful, especially at first." That's a pretty good description of learning a language. The clumsiness isn't a defect in the process. In a way, it is the process. The struggle to find a word is what makes the word yours once you have it, and what makes it mean something when you say it to someone.
So are we against AI?
No. We use AI at Parrot, and we'll use more of it. It would be strange to build a company in 2026 and pretend it doesn't exist.
But there are two very different kinds of tools, and from the outside they can look the same. One kind helps you do something. The other does it for you. A coach and a stand-in. A good coach makes you need him less over time. A stand-in makes you need him more.
The test is simple: after you use the tool for a while, can you do more on your own, or less? A tool that helps you practice until you can talk to your boyfriend's mother is the first kind. A tool that talks to her for you is the second. Both use the same technology. They point in opposite directions.
I'm not worried that AI will make people stop learning languages. I'm worried that it will make people forget why they wanted to.
What Parrot is for
Parrot looks like a language app, and it is one. But language is the means, not the end. The end is the conversation: with the friend, the partner, the grandmother, the person across the counter. Language just happens to be the biggest and most obvious wall between people, so that's where we started.
Our goal is to get a billion people fluent in a language. Another way to say that is a billion people able to have a conversation they couldn't have before. Some of those conversations will be clumsy. That's fine. The clumsy ones are often the ones people remember.
So yes, please use AI. Use it to get good enough that you don't need it. Then call someone and say it yourself, badly.
Notes
[1] I don't think Smucker is really telling anyone to stop using AI for meal plans. The piece is about what you trade away without noticing.
[2] Economists call this costly signaling: a signal is believable because it's expensive to fake. Speaking someone's language badly, but trying, is hard to fake. That's what makes it worth something.