About 80 WhatsApp pages showed Meta chatbot instructions by error
The text told the chatbot to give no legal advice and to not send users to privacy regulators.
Privacy company Redact found that about 80 help pages of WhatsApp had instructions for a language model of Meta. The instructions were in the code of the pages. A language model is an AI system that writes text. The pages held 100 blocks of this text. Meta removed it after a reporter contacted the company. A WhatsApp spokesperson tells that a bug showed the text and that the bug is removed.
How it works
The text was in the code of the pages, not on the part that a person reads. It had the heading “HelpContentGalaxyLLMInclude”. It tells the chatbot how to answer users.
The numbers
Redact found about 80 pages and 100 blocks. Examples of the text:
- On the pages to remove an account: do not tell users to contact a data protection authority.
- If users name legal action: do not use the terms of service or the privacy policy.
- Do not refer to privacy laws such as GDPR, LGPD and CCPA, also when users name them.
What is not known
It is not clear how long the text was shown, or if chatbots of other companies could read it. Redact tells that many instructions send users to support channels of WhatsApp. Meta tells that the chatbot must not give legal advice.
Insight: Meta keeps its chatbot rules in text users cannot see
A Meta spokesperson tells that the Muse agent can write adult content for adults. The instructions are in its system prompt, a text that tells the AI the work to do. At this time the WhatsApp chatbot shows more instructions by error. Maybe Meta keeps chatbot instructions in text that users cannot see until an error shows them. Make each AI vendor give you the instructions in a document.
Sources
Posted
