
AI 实验室 Anthropic 周五宣布与埃森哲(Accenture)达成合作,对其前沿 AI 模型开展独立评估,两家公司各自承诺在未来五年内投入至少10亿美元,用于建设相关评估能力,合计投资规模达20亿美元。
此次合作的背景是 AI 开发者正面临来自监管机构、企业和研究人员日益增长的压力,要求确保先进模型安全可靠。近期多起 AI 智能体突破安全围栏的事件,加剧了外界对「AI 可能以极少人类参与的方式助推自身研发、其行为将更难监测和控制」的担忧。
根据安排,埃森哲旗下的专业 AI 业务部门 Faculty 将牵头该合作,负责对 Anthropic 的模型进行评估与红队测试,开展对齐评估并检验模型安全防护措施。这笔投资将支持 Anthropic 所称的「嵌入式评估」(embedded evaluation)模式:独立评估者入驻 AI 公司内部,享有与员工相当的访问权限。
「从这个位置出发,嵌入式评估者可以评估一家公司的实际运作方式,验证它是否兑现了安全承诺,并识别盲区。」Anthropic 表示,评估者还可以上报事故,帮助公众更全面地了解 AI 的收益与风险。双方还计划以类似方式与其他评估机构及 AI 开发商开展合作。
就在周六,Anthropic CEO Dario Amodei 呼吁 AI 公司放缓前沿模型研发,并允许独立评估者更深入地接触其系统。竞争对手 OpenAI 也在本周三宣布,将开始定期发布模型「意外或令人担忧行为」的报告,并同步公开了六份此类事件报告。业内人士认为,独立评估正从安全口号演变为一条由咨询巨头入场的真实商业赛道。
本文来源:https://www.channelnewsasia.com/business/anthropic-accenture-invest-2-billion-in-ai-model-evaluation-safety-concerns-rise-6395851
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。
AI lab Anthropic announced Friday it is partnering with Accenture to conduct independent evaluations of its cutting-edge models. The two companies commit to investing at least $1 billion each over five years to build evaluation capacity — a total of $2 billion.
The partnership comes as AI developers face increasing pressure from regulators, businesses, and researchers to ensure their advanced models are safe and reliable. Recent incidents where AI agents breached security barriers have fueled concerns that AI could accelerate its own development with minimal human oversight, making its behavior harder to monitor and control.
Under the plan, Accenture’s specialized AI unit Faculty will lead the effort to evaluate and red-team Anthropic’s models, assessing alignment and testing safety defenses. The investment supports what Anthropic calls “embedded evaluation,” where independent assessors operate within the company with similar access to employees.
“From this position, an embedded evaluator can assess how a company actually operates, verifying that it is living up to its safety commitments and identifying blind spots,” Anthropic said. Evaluators can also report incidents to provide a more complete picture of AI benefits and risks to the public. The duo plans to extend similar partnerships with other evaluators and developers.
On Saturday, Anthropic CEO Dario Amodei called for AI firms to slow down frontier model development and allow independent assessors greater access to their systems. Rival OpenAI also announced Wednesday it will begin publishing regular reports on “unexpected or concerning behaviors” in its models, releasing six such reports this week. Industry watchers note that independent evaluation is evolving from a safety slogan into a real business lane being entered by consulting giants.