If you would not show it to the customer, do not show it to the model

Researchers found more than a hundred hidden blocks of text on WhatsApp’s help pages, written for AI systems rather than people. Two of them steered chatbots away from pointing users to data-protection regulators. WhatsApp says they were instructions for its own support bot, exposed by a bug, and that they are fixed.

I take the company at its word on the bug. The lesson is still worth stating. As more customer service runs through assistants, the instructions behind them become part of the service. If an instruction would embarrass you on a public page, it does not belong in the prompt either.

Our Help Center assistant is told to answer only from published help content, to say when it does not know, and to hand off to a person. It is never told to keep a customer from a regulator, a refund route or a complaint. Our Privacy Policy names how to reach us and the authorities that oversee us, and the assistant is free to repeat it.

For any business deploying a support bot, here is a simple test: print the system prompt and ask whether you would be comfortable if a journalist read it. If the answer is no, rewrite it before someone else finds it.

— Edgar Melo Costa, founder

Founder’s notes are opinion. They are not investment, legal or professional advice.