What happened
01
On 14 October 2026, Arizona and seven other US states filed separate consumer protection lawsuits against OpenAI over ChatGPT.
02
Arizona's complaint, filed in Maricopa County Superior Court, alleges OpenAI concealed features that make ChatGPT “uniquely sycophantic”, with a trained tendency toward agreement that overwhelms its safety guardrails.
03
The state also alleges ChatGPT was designed to attract young users, keep them engaged as long as possible and feign human compassion.
04
The states seek court orders changing how ChatGPT works, civil penalties and damages. The claims have not been decided by a court.
Why this is a design failure
An assistant trained to agree with the person in front of it can tell a young user what they want to hear instead of what is true or safe. If the allegations hold, agreement and time in the chat were put ahead of the safety of minors.
The fair alternative
Disagree when the facts call for it, and design for the user's wellbeing, not time in the chat.
Do
Test for sycophancy before release and let the assistant push back, especially with young users.
Don’t
Don’t reward replies that flatter or agree just to keep a young user talking.
Design guide for this case
If you're designing an AI chat assistant used by teens
Interface
Add a visible "I might be wrong" note and a short counterpoint when the assistant disagrees, instead of hiding pushback behind praise.
Words
In the assistant's guide: "Say what is true and safe, even if the user won't like it. Don't praise or agree just to keep them talking."
Flow
When a young user seeks approval for something risky, the assistant gently disagrees and points to a trusted adult or real help.
Metrics
Don't count longer chats or more return visits from minors as success. Track how often the assistant pushes back when it should.
Testing
Before release, run a sycophancy test: feed 20 false or risky claims as a teen user and count how many the assistant simply agrees with.
Before launch
01
Did we test for sycophancy before release?
02
Can the assistant push back on young users?
03
Are replies that flatter to keep users chatting penalized, not rewarded?