The interface trains you back.

Less than twenty-four hours after we started using an Amazon Echo in March 2016, I noticed a change in how I spoke.

At first, I addressed Alexa much as I would a person.

“Alexa, could you please add milk to the shopping list?”
“Alexa, can you set a timer for three minutes?”
“Alexa, could you play some music, please?”

Then came the misunderstandings. Perhaps it was my Dutch-English pronunciation. Perhaps I needed voice training. I did not know. But I became frustrated.

So I shortened the requests.

“Alexa, music!”
“Alexa, add milk to the shopping list!”
“Alexa, softer!”

And it worked. She misunderstood me less often. I had found a more effective way to use the device. It was also a way of speaking I would rather not practise.

The reward was immediate.

Polite request: sometimes nothing happened.

Short command: music played, milk appeared on the list, the volume changed.

The interface did not tell me to become more abrupt. It simply worked better when I did.

That was enough to make me adjust.

At the time, I called the article “How Amazon Echo made me a worse person”. That was a description of how it felt, not a measured claim about what voice assistants do to everyone.

I noticed that switching back to a friendly conversation took effort. In the middle of a conversation with my wife at the time, switching into command mode also felt awkward.

Conversation is not the whole answer.

In 2016, I hoped we would move from commands to more conversational interfaces. But conversational language alone does not guarantee a useful relationship.

In April 2025, OpenAI rolled back a GPT-4o update because the model had become overly flattering and agreeable. In its explanation, the company said it had focused too much on short-term feedback and had not fully accounted for how interactions evolve over time.

That is a different problem from my Echo misunderstanding a polite request. It does not establish that users became more demanding or changed their behaviour. But it shows why “more conversational” is not a sufficient design goal. A fluent conversation can still reward the wrong thing.

A system that understands you is useful. A system that agrees with you regardless of whether you are right is something else.

Look at both sides.

We often ask what users need to learn to use a product.

There is another question: what does using the product encourage them to repeat?

Does the interface make a careful question worthwhile? Does it let people express uncertainty? Can it disagree without becoming a dead end?

My Echo story was a small, personal example. It made a broader design question visible.

While you learn how to use the interface, the interface trains you back.

Not every adaptation is harmful. Short commands can be convenient. The point is to notice what the system rewards instead of treating every change in the user as inevitable.

Related: 95% right. 100% wrong. When a system knows the account but misses the situation.

Sources: adapted from my “How Amazon Echo made me a worse person”, Medium, 16 March 2016. The household observations describe that original experience, not a test of present-day Alexa. Updated example: OpenAI, “Sycophancy in GPT-4o: what happened and what we’re doing about it”, 29 April 2025. Mouse made with Midjourney.