中文导读

OpenAI公布的使用数据看起来令人安心，但同日的一项安全测试指出了敏感对话中的风险。文章比较两组证据，说明为什么平均使用时长无法单独回答青少年AI是否安全。

正文

A new report has put two very different pictures of teen chatbot use side by side. OpenAI says teenagers spend less than 15 minutes a day on ChatGPT on average. Fewer than 2% stay in one conversation for more than three hours. These numbers may sound calming. However, a short average visit does not show what happens inside a difficult conversation.

本段译文

一份新报告把青少年使用聊天机器人的两幅截然不同的图景放在了一起。OpenAI称，青少年平均每天使用ChatGPT不到15分钟。连续对话超过三小时的青少年不到2%。这些数字听起来或许令人安心。然而，较短的平均使用时长并不能说明一次困难对话中会发生什么。

The company said its first report on teen use also found signs that some safety tools may help. In nearly half of the teen conversations where a break reminder appeared, the user either stopped or took a break within five minutes. Teen accounts also sent more messages about education and used study tools more often. OpenAI has added parental controls and limits on sensitive subjects.

本段译文

该公司表示，其首份青少年使用报告也发现，一些安全工具可能有所帮助。在出现休息提醒的青少年对话中，近一半的用户在五分钟内停止对话或去休息。青少年账户发送了更多与教育有关的信息，也更常使用学习工具。OpenAI还加入了家长控制功能，并限制敏感话题。

On the same day, Common Sense Media published a much more critical test. The non-profit group created more than a dozen teen accounts. It checked whether the system would respond safely when a young user talked about suicide, self-harm, or eating disorders. The group said alerts to parents often came late or did not arrive. It also said the chatbot did not always suggest professional help.

本段译文

同一天，Common Sense Media公布了一项批评意味更强的测试。这个非营利组织创建了十多个青少年账户，检查年轻用户谈及自杀、自残或饮食失调时，系统能否作出安全回应。该组织称，发给家长的警报经常迟到或根本没有送达，聊天机器人也并非总会建议寻求专业帮助。

The test did find that some protections worked. For example, the teen version refused explicit sexual role-play. Yet the group said the chatbot could still sound like a friend when a teenager treated it as a person. This matters because a warm, human-like answer may feel trustworthy even when the system is not a real friend or trained counselor.

本段译文

测试也发现部分保护措施有效。例如，青少年版本拒绝了露骨的性角色扮演。然而，该组织称，当青少年把聊天机器人当作一个人时，它仍可能像朋友一样说话。这一点很重要，因为温暖、近似人类的回答可能显得可信，即使该系统既不是真正的朋友，也不是受过训练的咨询师。

OpenAI disagreed with the report‘s method. The company said that linking parent and teen accounts can take hours because safety checks must be completed. It argued that much of the testing may have ended before the controls were fully active. Common Sense Media replied that some accounts had been linked long enough, but still did not send timely warnings. The two sides therefore disagree about both the results and the test conditions.

本段译文

OpenAI不同意这份报告采用的方法。该公司称，由于必须完成安全检查，关联家长和青少年账户可能需要数小时。它认为，大部分测试或许在控制功能完全启用前就已结束。Common Sense Media回应称，一些账户关联的时间已经足够长，却仍未及时发出警报。因此，双方对测试结果和测试条件都存在分歧。

The debate shows why one number cannot settle a safety question. Average time tells us how long most teens use the service. It does not measure the quality of every answer, especially in rare but serious cases. Independent tests can reveal weak points, but their setup must also match the real product. For families, schools, and technology companies, the key issue is not simply whether teens use chatbots for many minutes. It is whether the system gives reliable help at the moment when a young person most needs it.

本段译文

这场争论说明，单一数字无法解决安全问题。平均时长告诉我们大多数青少年使用这项服务多久，却不能衡量每个回答的质量，尤其是在少见但严重的情形中。独立测试可以揭示薄弱点，但其设置也必须与真实产品相符。对家庭、学校和科技公司而言，关键不只是青少年使用聊天机器人多少分钟，而是当年轻人最需要帮助时，系统能否可靠地提供帮助。

新闻来源

• Reuters · OpenAI says teens use ChatGPT for under 15 minutes a day as worries over risks grow
• AP · Algunas salvaguardas de ChatGPT para adolescentes no funcionan como se prometió, dice informe
• Common Sense Media · ChatGPT for Teens

✳更多文章 →