← All posts

AI Dementia

Charles Stack, Founder, Coworkers.Global

This morning I asked my iPhone, “Hey Siri, what version are you?”

She said: “I am Siri, here to help.” Surprised, I reasked and got the same non-response.

A doctor checking for cognitive decline asks the patient what year it is. Siri can’t tell me what version she is.

My phone and my TV both have assistants. Both of them used to be sharper than they are now.

I don’t mean the technology stopped improving. It kept improving the whole time. I mean the assistants I actually talk to, day to day, forgot how to do things they once did well. They got worse at their jobs while the industry around them got dramatically better. I’ve started calling it AI dementia.

Siri, then and now

When Siri shipped in 2011, she was a small miracle. You could ask her anything, and she’d try. She had opinions. She told jokes. Ask her to divide zero by zero, and you got a short, slightly mean story about Cookie Monster and his lack of friends. People made whole listicles of “funny things to ask Siri.” She had a personality, and the personality was the product: it told you someone at Apple had thought about what it would feel like to talk to this thing.

Ask Siri for the weather back then, and you got the weather: temperature, high and low, chance of rain, what the afternoon looked like. Ask for the news, and you got the news.

Ask today, and you get something like: “It will be warm and might rain.”

That isn’t a forecast. It’s a generic fortune cookie.

The jokes are mostly gone too, or buried. The answers got shorter, blander, and more cautious. Somewhere along the way, a lot of what made Siri worth talking to was sanded off.

It’s not only nostalgia. In 2019, the consultancy Perficient reran a voice-assistant accuracy test it had first done in 2017. Siri’s correct-answer rate had fallen from over 80% to 70%. Alexa fell from over 90% to 79%. They concluded the assistants had hit a ceiling. I don’t think they’d hit a ceiling. They were going backward.

Then it got stranger. In January 2025, John Gruber asked Siri, freshly upgraded with Apple Intelligence, who won a certain 2004 state high school basketball championship. He asked four times and got four different wrong answers, each delivered with total confidence. The old Siri, he pointed out, would have said “here’s what I found on the web” and handed him links that contained the right answer. The new one made things up.

That’s the part that earns the word dementia. Forgetting is one symptom. The other is confabulation: filling the gaps with confident, plausible, wrong stories. An assistant that says “I don’t know” is limited. An assistant that invents a basketball score is unreliable, and you can’t tell which answers are invented.

Apple made promises about the fix. In 2024, it announced a smarter, more personal Siri, then couldn’t ship it. That became a class action, and in May 2026 Apple agreed to pay $250 million to settle claims that it oversold the upgrade (up to $95 per eligible iPhone, no admission of wrongdoing). The redesigned “Siri AI” finally shipped with iOS 27 on September 14, running on a model Apple built with Google’s Gemini. iOS 27 hasn’t reached my phone yet. I am still on 26.6.2, the version that told me “I am Siri, here to help.” I will let you know.

Alexa, and the Fire TV problem

Alexa on an Echo is fine. Timers, lights, music, questions: she mostly does the job.

Alexa on a Fire TV is a different animal. Her main job there is to find the show you asked for, and she can rarely do it.

Here’s the routine at my house. I press the button and say the name of a show. Alexa answers, brightly, “Here it is…” and nothing on the screen changes. No show. No search results. Nothing. She has announced success at a task she didn’t do.

It’s worse if the show lives anywhere other than Amazon. Ask for something on Netflix, Max, or any other service on my Fire TV, whether I pay through Amazon or directly, and she rarely finds it. She’ll offer to rent me something from Prime instead, or show me a row of Amazon titles with a similar name. Amazon’s own help pages acknowledge that some TV and video providers “don’t work with Alexa.” From the living-room couch, that reads as: the assistant is better at selling Amazon than at helping me. She is also annoyingly chipper, which puts her in nails-on-a-chalkboard territory, particularly when she is also wrong or unhelpful.

“Here it is,” followed by nothing, is the same failure as Siri’s invented basketball score. The assistant reports a result it doesn’t have. Worse, it’s wrong about whether it helped.

Amazon’s answer is Alexa+, its generative-AI rebuild, which it made free on Fire TV in August 2026 and automatically pushes to devices. The pitch is that you can ask for “a historical drama with a strong female lead” instead of a title. I have Alexa+ on my TV. Everything above happened with it. I’d settle for the title. Alexa+ also seems to have gotten even more chipper.

Why capable assistants get worse

The obvious question is how this happens at the two most valuable companies in consumer tech. My guesses, from the outside:

  • Nobody owns being right. Features get added, and teams change, and the day-to-day quality of the answer isn’t anyone’s job. A test that passed in 2017 isn’t rerun in 2019, so nobody notices the drop.
  • Risk teams win quiet wars. Every joke is a potential headline and every detailed forecast is a potential complaint. Each individual cut is defensible. Together they add up to “it will be warm and might rain.”
  • The business model leaks into the answer. When the assistant belongs to a store, “find my show” drifts toward “find something we sell.”
  • New models get bolted onto old plumbing. A language model that sounds fluent, sitting on top of a search system that can’t reach the content, produces fluent failures. That’s “Here it is.”

Half a cure

Apple and Amazon, to their credit, have finally recognized the problem and shipped their cure: new models, new names, big launch posts. I will watch to see if the bar has been raised or I keep getting poked in the ear.

A new brain doesn’t help much if nobody checks whether the patient remembers what they knew last year. The industry keeps getting smarter. The assistants in my house keep forgetting.

Hey Siri, what version are you?

Regards,
Charles Stack
Founder, Coworkers.Global

Sources

Coworkers.Global is an AI staffing agency. We place managed agents into organizations that need dedicated expert knowledge work. A managed agent is an AI specialist provisioned for a specific role, trained on your context, supervised by a person, and accountable for its output. The first, Alex, evaluates startup business plans for fundability, informed by human expertise and research, and calibrated against real investor decisions. We are early-stage with paying clients, and we still lead with the quality of our judgment rather than customer logos. Your managed AI coworker.
Get new posts by email

What we're learning building a startup with managed agents, plus notes on raising. Monthly newsletter, no spam.