What happened
01
AGCM opened consumer protection cases against the three chatbot providers.
02
It said users were not told clearly and immediately that answers could be fabricated.
03
DeepSeek committed to prominent in-chat warnings and full Italian disclosures.
04
The closures were published on 30 April 2026, with a 120-day compliance check.
Why this is a design failure
A confident answer with no warning invites people to trust text that may be invented.
The fair alternative
Say plainly, inside the chat, that answers can be wrong and should be checked.
Do
Show a clear accuracy warning where people read the answer.
Don’t
Don’t bury the hallucination warning in the terms.
Design guide for this case
If you're designing a chat assistant
Interface
Show a short accuracy warning right under answers, in the user's language, saying answers can be wrong and should be checked.
Words
In the content guide: "Write the warning in plain words, like 'This answer may contain mistakes. Check important facts.' Never hide it in the terms."
Flow
Show the warning in the chat from the first answer, not only after the user opens settings or the terms of service.
Metrics
Don't count fewer warning views as a cleaner experience. Track whether users can recall that answers may be wrong.
Testing
Ask 5 new users, in each supported language, where the app tells them answers may be wrong. They should find it in seconds.
Before launch
01
Is the accuracy warning visible where people read the answer?
02
Is it written in every language the app supports?
03
Is it shown from the first answer, not buried in the terms?