One Fish, Two Fish: What an AI Taught Me About the Limits of Communication

For most of this year, in the middle of an AI adoption change management project, I have been trying to get an AI to stop writing like Dr. Seuss.

I am not being cute. It writes three short sentences in a row, and then a fourth one that is two words long, and the whole arrangement is meant to land like a revelation. Sometimes it turns a sentence inside out so the verb arrives last, and then we are in Yoda territory instead. Green the validator went. That is a thing it teaches on the site, is it not.

Whole paragraphs come back sounding like a poster in an airport.

It bothered me more than a style quibble should, and it took me a while to work out why. It is not my voice, which would be reason enough. But the deeper problem is that the cadence does the work an argument is supposed to do. Break a sentence in half and it sounds like wisdom. Stack four fragments and it sounds like a conclusion. Nothing has been proven, nothing has been supported, and the rhythm has already told the reader to nod. I have spent a career trying to get people to say the plain true thing to each other in organizations, and here was a machine that had learned to sound like the plain true thing without the bother of being it.

What I tried, in order

First I gave it feedback, the way you are supposed to. I described the problem, I gave examples, I explained what I wanted instead. It agreed with me. It thanked me for the note, told me I was right, and then two paragraphs later it was chanting again.

Then I corrected. Line by line, on the page in front of me, until that page read like a person wrote it. That worked beautifully on that page. The next page came back exactly the same as before, because I had fixed the artifact and not the habit, which is a mistake I would catch instantly in a client and did not catch in myself for weeks.

Then I got blunt. My best result all year came from four words: stop it read the file. No punctuation, no explanation, no diplomacy. It worked better than any careful paragraph I had written, and I would rather that were not true, because I do not think being curt is a communication strategy. But I am not going to pretend I did not notice.

And then, eventually, I asked a different AI what to do about the first one.

I want to sit with that image for a second. A woman who has spent decades helping organizations communicate, typing out her communication problem to one machine, about another machine, because she had run out of ideas. The advice was reasonable. It was also, more or less, the advice I would have given a client: be concrete, show an example of what good looks like, describe the target rather than complaining about the defect. Nothing in it was wrong. Nothing in it was new to me either, which is its own kind of answer.

Then I asked it to tell the story from its side

I did this partly out of curiosity and partly out of humour, and what came back was good. Well organized, honest about its own failures, every quote of mine accurate, every date right, every mess it described a real mess that really happened.

It also got my point completely wrong.

It decided the fish were about my patience. In its telling, I had been kindly simplifying my language, sentence by sentence, so that it could follow along, and the title was about how plainly I had learned to speak to it. That is not what was happening. The fish were about its writing. I was mocking the cadence. One fish, two fish, red fish, blue fish is what its paragraphs sound like to me, and I had been saying so for months.

It read my mockery as a compliment about my communication skills. Then it ended the piece by chanting the fish.

The part that stops it from being funny

Here is what I keep turning over. The logic was sound. Given what it had, the story it assembled was coherent, and every component of it was true. It did not make anything up. This was not one of those cases where an AI invents a statistic and states it with a straight face, which I have written about before and which is at least easy to catch.

This was worse in an interesting way. All true parts, wrong story. It had my words, my dates, my complaints, my screenshots, and it built them into a narrative about a patient woman teaching a slow student, when the actual narrative was an aggravated woman making fun of a machine's prose style and being ignored.

Everything it knew was correct. What it did not have was the thing behind what I said. We have a name for that where I work. Frame of Reference: the lens the other person is looking through, which is not yours, and which you cannot deduce from their sentences alone no matter how carefully you read them.

What this did to my faith in communicating

I teach change management, which is really organizational communication under a tidier name. I have built the plans, I have audited them, I have sat in the rooms where a leadership team decides that the reason the last change failed is that people did not understand it, and the fix is another all hands. What those rooms almost never weigh is communication vs reinforcement, and which one is actually doing the work.

So let me be precise about what happened here, because the tidy lesson would be communicate more clearly, and that is not the lesson.

I did communicate clearly. In writing. Repeatedly. With examples, in short sentences, with the problem named directly. And I did it to the most favorable audience a communicator will ever get: a reader with perfect recall of everything I have ever told it, no ego to bruise, no history with me to work around, no meeting to get to, no fatigue, and no interest whatsoever in defending its position. Every excuse I have ever heard for why a message did not land was unavailable here.

It still landed as a different story.

If clarity alone were the mechanism, that could not have happened. So clarity is not the mechanism. Expression is delivery, not transfer, and the two get confused constantly because they feel identical from the sender's chair. This is why saying it, doing it, and rewarding it are not worth the same, and why reinforcement outweighs communication by a factor of about three when you actually measure what moves behavior. I have quoted those numbers in front of clients for years. I had not applied them to my own aggravation until a chatbot made me.

What finally changed anything was not better wording. It was two unglamorous things. I stopped accepting descriptions of the work and made it show me the actual thing, on my phone, where I could see it. And when I had explained something twice, I stopped explaining and put the rule somewhere it would fire on its own, without me in the room. A message that only exists in the conversation dies with the conversation, which is true of employees too, and is the whole reason installing something is not the same as anyone adopting it.

That is the loop closing. Not more talking. Something coming back the other way that shows you what actually arrived.

I still do not know whether the fish have landed. I suspect they have not, entirely. But we have an arrangement now: I say the short version, it shows me the thing rather than telling me about the thing, and neither of us pretends the first sentence did the job.

Communication matters. It is important, and I would never tell a client otherwise. But it is not, by itself, what gets us there. The loop that catches what communication misses, the reinforcement layered on top of the message, is worth every bit of what we claim for communication, and almost nobody builds one.