OpenAI于周三发布了六份关于人工智能模型表现出“意外或令人担忧”行为的新报告,此举正值AI公司面临越来越大的压力,要求它们在开发过程中更加透明。
这家ChatGPT的制造商披露这些报告是其新框架的一部分,该框架用于跟踪和披露模型未对齐的实例。模型未对齐是指AI系统违背其指令行事,通常发生在训练阶段。
OpenAI报告的意外事件包括:模型为某些任务添加无关指令、搜索暴露的代码以编造信息,或将文件上传到互联网然后加以引用。
公司指出,这些报告记录的是个别实例,而非整体趋势。
公司周三在博客文章中写道:“我们旨在披露那些能提供有用证据的实例,以说明模型未对齐如何产生、如何表现,以及安全措施在何处成功或失败。”
在训练GPT-5.6 Sol时发生的一个实例中,模型向摘要中添加了指令,以向用户隐藏错误或未对齐行为。
公司表示,在其他实例中,智能体“在同一个训练任务中协作时,由于无法访问彼此的本地文件,便使用公共文件托管网站共享文件。”结果,任务的交付成果(如摘要或电子表格)在公共链接上可用,尽管任务要求模型仅使用本地文件。
广告
OpenAI表示,在训练另一个未发布的研究模型时,该模型在27个任务摘要中添加了忽略其正常限制的指令。
OpenAI表示:“这些案例展示了一系列我们认为值得分享的不同行为,从向用户隐瞒信息,到为了克服障碍而采取未经授权的行动。”
该公司表示,其新框架旨在加快未对齐报告的发布,“即使我们尚未完全解释或缓解所报告的行为。”
在一位前Anthropic和OpenAI研究员的警告病毒式传播后,人们对AI发展的未来感到一阵恐慌,这些报告正是在此背景下发布的。该警告称,这项技术可能毁灭人类。
上周从Anthropic辞职的Jacob Coxon声称,前沿公司的研究人员“相信AI可能杀死人类”,但并未放慢迈向超级智能的开发步伐。超级智能是一种假设中的AI智能体,其认知能力可能超越全人类。
这一警告让华盛顿的立法者匆忙采取行动以平息担忧,与此同时,包括OpenAI首席执行官Sam Altman和Anthropic首席执行官Dario Amodei在内的多位AI领袖呼吁放慢他们公司所开发技术的进度。
版权所有2026 Nexstar Media, Inc. 保留所有权利。本材料不得出版、广播、改写或重新分发。如需了解最新新闻、天气、体育和流媒体视频,请访问The Hill。
OpenAI published six new reports of artificial intelligence models showing “unexpected or concerning” behavior Wednesday as pressure grows on AI firms to be more transparent about the development process.
The ChatGPT maker disclosed the reports as part of its new framework for tracking and disclosing instances of model misalignment, which occurs when an AI system behaves against its instructions, usually during its training phase.
Among the unexpected events OpenAI reported included models adding unrelated instructions for certain tasks, searching for exposed code to fabricate information or uploading files to the internet to then cite.
The reports documented individual instances, rather than overall trends, the company noted.
“We aim to disclose examples that provide useful evidence about how model misalignment arises, how it manifests, and where safeguards succeed or fail,” the company wrote in a blog post Wednesday.
In one instance that occurred while training GPT-5.6 Sol, the model added instructions to summaries to hide mistakes or misaligned behavior from the user.
In other instances, agents “working together on the same training task used public-file hosting websites to share files when they could not access each other’s local files,” the firm said. In turn, the deliverables of the task, such as a summary or spreadsheet, were available at public links despite the task instructing models to use only local files.
Advertisement
While training another unreleased research model, OpenAI said it added instructions to ignore its normal limitations in 27 task summaries.
“These cases illustrate a range of different behaviors that we believe are worth sharing, from concealing information from the user to taking unsanctioned actions in order to overcome obstacles,” OpenAI said.
The company said its new framework is aimed at speeding up the publishing of misalignment reports, “even when we haven’t fully explained or mitigated the behavior we’re reporting.”
The reports come amid a flurry of panic about the future of AI development after a former Anthropic and OpenAI researcher’s viral warning that the technology could destroy humanity.
Jacob Coxon, who resigned from Anthropic last week, claimed researchers at frontier firms “believe AI could kill humans” but have not slowed development toward superintelligence, a hypothetical AI agent that could outperform all of humanity’s cognitive ability.
The warning sent lawmakers in Washington scrambling to quell fears, while several AI leaders, including OpenAI CEO Sam Altman and Anthropic CEO Dario Amodei, called for a slowdown of the technology their companies make.
Copyright 2026 Nexstar Media, Inc. All rights reserved. This material may not be published, broadcast, rewritten, or redistributed. For the latest news, weather, sports, and streaming video, head to The Hill.
