OpenAI 日前宣布将在未来几周内加强 ChatGPT 对暴力相关内容的限制，为未成年人用户增加武器和安全系统信息方面的保护。根据新措施，当对话显示用户持续关注暴力或严重伤害他人的想法时，ChatGPT 可以自行结束对话，若多次因此类问题被结束对话，甚至账号都会被封禁。
ChatGPT 原有政策禁止提供帮助用户策划或实施暴力行为的指导，此次调整将限制模型回答有关“过去大规模袭击事件及实施者”的问题时提供的细节，即模型会从安全层面缩减内容甚至拒绝回答。
限制范围包括重现袭击过程中涉及的武器、装备及获取方式，以及限制用户查询可能扩大袭击者知名度的伤害规模详细比较。以上限制适用于所有用户，不区分用户年龄段或提问目的，因此即便是成年用户也会面临相同限制。
对于被识别为未满 18 周岁的用户，OpenAI 还将限制可能被用于实施暴力的武器和安全系统的实用信息，这项保护也会应用于表面上看起来没有显示暴力意图的问题。OpenAI 表示，青少年仍然可以使用 ChatGPT 寻求帮助，了解如何预防或缓和暴力、与可信赖的成年人取得联系。
改进后的年龄预测机制则会结合用户行为等多种信号，持续评估账号是否可能属于未成年人，在适当的情况下系统将自动应用适用于未满 18 周岁用户的相关保护措施。如果用户被误判为未成年人，也可以主动提交身份信息进行验证。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Open AI has recently announced that it will, in the coming weeks, strengthen ChatGPT ' s restrictions on violence-related content and increase the protection of weapons and security system information for minors. Under the new measure, ChatGPT can end the dialogue on its own when it shows that users are constantly concerned about the idea of violence or serious harm to others, and even accounts are blocked if they are closed on such issues several times.
The original policy of ChatGPT, which prohibited guidance to help users to plan or commit acts of violence, would limit the details provided by the model in reply to the question about “the past mass attacks and their perpetrators”, i.e. it would reduce the content from the security level and even refuse to answer.
The scope of the restrictions includes the re-emergence of the weapons, equipment and methods of acquisition involved in the attack, as well as a detailed comparison of the scale of harm that may increase the profile of the attackers by limiting user access. The above restrictions apply to all users, without distinguishing between age groups of users or the purpose of the questions, and therefore even adult users face the same limitations.
Open AI will also limit practical information on weapons and security systems that may be used to commit violence for those identified as being under 18 years of age, and this protection will also apply to issues that appear to have no violent intent. Open AI indicated that adolescents still have access to ChatGPT to seek help in preventing or mitigating violence and in reaching trusted adults.
The improved age projection mechanism, combined with a variety of signals, such as user behaviour, continuously assesses whether the account number may belong to a minor and, where appropriate, automatically applies the relevant protective measures to users under 18 years of age. If the user is miscalculated as a minor, identification information may also be submitted on its own initiative.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.