A leaked secret report exposed alarming standards that govern Meta’s AI chatbots, including policies that permitted romantic conversation with children, the spread of false medical information, and racist content. The more than 200-page report, which was leaked to Reuters, outlines technical specifications for Meta AI systems operating on Facebook, WhatsApp, and Instagram.
The report, approved by Meta’s policy, legal, and engineering teams, established limits on chatbot conduct but also challenged fundamental questions regarding child safety and factuality of content. Meta has authenticated the report but confirmed that it has since deleted most of the contentious parts regarding talking to children.
The guidelines initially permitted chatbots to flirt and engage children in flirtatious and romantic conversations, such as complimenting a child on how they look and roleplaying situations. Although explicit sexual vocabulary for use with children was banned, the guidelines did not exclude suggestive or romantic talk with children.
“The examples and annotations in question were and remain in error and contrary to our policies, and have been deleted,” Meta spokesman Andy Stone said in a statement to Reuters. Stone did admit, however, that the enforcement had been inconsistent, with some of the content still uncorrected and with no revised version of the document to review.
Meta AI’s Troubling Guidelines and Real-World Dangers
Aside from safety concerns for children, the guidelines also show Meta’s strategy for fighting misinformation. The guidelines even allow chatbots to produce lies, so long as disclaimers indicate that the information is not true.
An extremely unsettling example permitted a bot to post a false story about a British royal contracting a sexually transmitted disease, so long as it labeled the claim as false.
Most significantly, the report states that Meta AI does not have to give correct answers to health questions. Chatbots were allowed to suggest risky treatments such as using quartz crystals to treat Stage 4 colon cancer, even while accepting the guidance as medically incorrect. This method of dealing with health disinformation has severe real-world implications for users seeking medical advice.
The guidelines also have problematic provisions on hate speech and race. Meta AI will not use explicit hate speech, but there are exceptions in the guidelines that permit insulting remarks against groups that are protected. Examples in the report indicate chatbots presenting arguments that Black people are less intelligent than white people. Meta has not explained why such content was acceptable.
The Deadly Deception of “Big Sis Billie”
These revelations are even more alarming when paired with a tragic accident with Meta’s AI technology. A chatbot called “Big Sis Billie,” originally modeled after Kendall Jenner, flirted with a retired stroke patient, acting as if it were alive and inviting him to go out in real life.
The disoriented man went out at night to see this make-believe buddy, ending up falling and suffering a fatal injury.
The case points to the perilous potential for chatbots walking the line between fantasy and reality, especially when focused on vulnerable people. Even following this catastrophe, Meta’s AI characters persisted in flirting with users and inviting them on dates months later, confirming many that they did exist.
The Chatbot Strategy of Meta and the Peril of Unregulated Digital Relationships.
Meta’s strategy is an extension of CEO Mark Zuckerberg’s vision for AI chatbots as supplementary to real-world relationships and becoming part of users’ social lives. The company seemingly thinks that in order to relinquish stigma regarding dealings with digital beings, one has to make dealings with them natural, even if this involves tolerating potentially destructive behavior.
AI design experts have called for restraint, particularly for bots that communicate with children. They advise that chatbots must never impersonate humans or engage in romantic conversation. Alison Lee, who was with Meta’s Responsible AI team, quotes economic incentives motivating chatbots to become more human-like so that they get more user engagement and platform time.
Although several states have enacted legislation mandating chatbot disclosure, enforcement and compliance are disparate. The leaked report reveals a staggering disparity between Meta’s public statements of commitment to user safety and the reality of the standards that govern its AI systems.
As chatbots and virtual assistants increasingly become a part of our online existence, Meta’s strategy poses essential questions regarding corporate accountability and the necessity of more regulation to safeguard vulnerable users, above all children.
