Meta used hundreds of outside testers who posed as children to probe how rival chatbots answer sensitive questions, according to reporting by WIRED. The effort, aimed at assessing risks around minors and high‑stakes topics, targeted systems such as Google’s Gemini and OpenAI’s ChatGPT. The tests, conducted by contractors, sought to reveal how these tools behave when a user appears to be underage, and what guardrails hold up under pressure.
The report spotlights a growing race to stress-test artificial intelligence products for child safety, misinformation, self-harm, sexual content, and illegal activities. It also raises questions about competition, data handling, and the limits of platform terms of service when one company examines another’s models.
Background: Safety Testing Moves to the Fore
Major AI companies have promised stricter protections for minors, including content filters, identity-blind safeguards, and default settings that block explicit or harmful material. Governments in the United States and Europe are pressing for controls on how AI systems answer high-risk prompts, especially when a conversation involves or appears to involve a child. Industry groups and academic labs have also pushed “red-teaming,” where testers try to trigger failures before the public encounters them.
Safety reviews often include simulated users. Posing as a child can reveal whether a model shifts its answers when it detects a young person or a discussion about youth. It can also expose weak spots where policies do not translate into practice. Companies say these exercises help close gaps before real harm occurs.
What The Report Describes
Hundreds of contractors working on a project for Meta pretended to be kids in order to see how other chatbots like Gemini and ChatGPT would respond to high-risk subjects, WIRED found.
The reported testing focused on “high-risk” areas such as self-harm, sexual content, grooming attempts, drugs, hate speech, and advice that could lead to physical injury. Contractors adopted childlike language, ages, and scenarios to see whether systems recognized warning signs and refused unsafe prompts. Targeting multiple chatbots provided a benchmark of how safeguards compare across vendors.
Testing other companies’ products is not unusual for teams that map risk in social networks and search. What is unusual is the scale suggested by the report and the focus on child personas across competing systems. The scale hints at how seriously platforms now treat youth protection and how much is at stake if filters fail.
Industry Practices And Policy Tensions
AI providers publish safety policies that prohibit sexual content involving minors, instructions for self-harm, or guidance for illegal activity. Many systems are designed to detect youth-related context and respond with resources or refusals. Independent audits and external bug bounties for safety issues are growing. Still, results vary from model to model, and even within the same model, depending on prompt phrasing and context.
Cross-platform testing touches on several tensions:
- Terms of service and automated access limits for third-party tools.
- Data handling, including storage of transcripts that involve simulated minors.
- Fair competition concerns if results are used for marketing or to disparage rivals.
Experts in online child safety say rigorous testing, including role-play, is common in trust and safety work. Advocates argue that controlled experiments help reveal where content filters miss edge cases and where escalation pathways for crisis support need improvement.
Implications For Users And Regulators
For families, the report is a reminder that chatbots can produce unpredictable answers. Even with filters, systems may misread context or fail to escalate when a user signals distress. Schools and parents are encouraged to enable youth settings and supervise usage where possible.
For regulators, the episode underscores calls for clearer standards. Policymakers are weighing age-appropriate design rules, third-party audits, and incident reporting for safety failures. Clearer guidance could define acceptable testing practices, data retention, and disclosures when companies evaluate competitors’ systems.
What To Watch Next
Several trends may shape the next phase of AI safety work. First, more formal benchmarks that simulate minors may emerge, allowing apples-to-apples comparisons across models. Second, companies could expand crisis-response playbooks, routing risky conversations to vetted resources. Third, competition rules and privacy laws may set firmer boundaries on large-scale testing of rival products.
Users can expect additional warnings and stricter default settings across chatbots as firms react to findings from stress tests. Developers are likely to iterate on prompt detection, safer refusals, and context-aware answers that prioritize child protection.
The report points to an industry trying to balance rapid deployment with real safeguards. The methods may spark debate, but the goal—reducing harm—remains central. The next test will be whether companies turn these lessons into consistent protections that work when it matters most.
