五角大楼希望在五年内获得3030万美元的资金，用于开发一种无需接触人体即可检测谎言的人工智能系统。该计划被命名为“Polygraph+”或“Polygraph Next”，并已列入国防反情报与安全局（DCSA）的2027年预算申请中。
The Pentagon wants $30.3m over five years to build an AI lie detector that reads the body without touching it. The programme, called Polygraph+ or Polygraph Next, appears in the Defense Counterintelligence and Security Agency’s 2027 budget request.
《Inside Defense》率先报道了这一计划。《MIT Technology Review》的阿米特·卡特瓦拉（Amit Katwala）向相关专家咨询了该技术的可行性。国防反情报与安全局负责为联邦政府进行背景调查，但并未回应《MIT Technology Review》关于详细信息的请求。
Inside Defense first reported the plan. MIT Technology Review’s Amit Katwala asked experts whether it could work. The agency, known as DCSA, runs background checks for the federal government. It did not respond to MIT Technology Review’s request for details.
根据预算文件，该机构在2027年请求获得642万美元的资金，随后在2031年之前每年需获得565万至645万美元的资金。国会尚未批准这一预算申请。
What the budget says DCSA asks for $6.42m in 2027, then between $5.65m and $6.45m a year up to 2031. Congress has not yet approved the request.
预算文件指出，该计划的目的是“升级联邦政府的测谎技术及可信度评估手段”。具体措施包括：通过非接触式方式收集人体生理数据（如心率、呼吸频率等），并结合人工智能与机器学习技术进行数据分析；同时开发自动评分系统及辅助决策工具。
The programme aims to “modernize federal polygraph and credibility assessment technologies”, the budget document says. It lists “standoff sensing”, or taking physical readings from a person without attaching a device, along with AI and machine learning scoring, automated scoring and decision aids.
这些工具将用于审查员工背景以及检测“内部威胁”。国防反情报与安全局还计划使用云计算和数据分析技术来优化相关软件。
The tools would be used to vet staff and for “insider threat detection”. DCSA also plans cloud storage and data analysis tools to refine the software.
据《MIT Technology Review》报道，在国防部长皮特·赫格塞斯（Pete Hegseth）的领导下，五角大楼越来越频繁地依赖测谎测试来查明信息泄露的源头。9月份，《纽约时报》曾报道称，联合参谋部约有50名官员接受了测谎测试。
Leak hunts and prototypes Under Defense Secretary Pete Hegseth, the Pentagon has turned more often to polygraph tests to find the sources of leaks to the press, MIT Technology Review reported. In September, The New York Times reported that about 50 officers on the Joint Staff had been tested.
五角大楼此前也曾尝试过类似的技术：2023年，其国防创新部门选定了两家公司来开发相关原型。Presage Technologies公司表示他们可以利用普通摄像头测量心率和呼吸频率；Altec Research公司的原型设备则能监测头部运动、面部皮肤温度及毛孔活动。
The Pentagon has tested this kind of technology before. In 2023, its Defense Innovation Unit picked two companies to build prototypes. Presage Technologies says it can measure heart and breathing rates with ordinary cameras. Altec Research’s prototype tracks head movement, facial skin temperature and pore activity.
然而，从科学角度来看，测谎技术本身并不能直接检测谎言——它只是记录人体的生理反应，由考官根据这些数据来判断被测者是否说谎。2003年，美国国家研究委员会曾指出测谎技术的准确性“充其量也很有限”。
The Pentagon has tested this kind of technology before. In 2023, its Defense Innovation Unit picked two companies to build prototypes. Presage Technologies says it can measure heart and breathing rates with ordinary cameras. Altec Research’s prototype tracks head movement, facial skin temperature and pore activity.
美国测谎协会表示，这种测谎技术的准确率在80%到94%之间。尽管如此，任何基于该技术的筛查程序仍可能产生许多错误（即：即使测谎结果准确率很高，也仍可能存在误判）。据《麻省理工科技评论》报道，五角大楼共有280万名员工。
The science problem A polygraph does not detect lies. It records breathing, pulse, blood pressure and sweat, and an examiner reads the changes. In 2003, the US National Research Council called the evidence for its accuracy “weak at best”.
英国诺森布里亚大学的法律学者基里·科茨奥格鲁（Kyri Kotsoglou）认为，在测谎设备中加入人工智能技术是“两败俱伤”的做法——这种做法不仅加剧了测谎结果的不确定性，还进一步削弱了测谎技术的可靠性。科茨奥格鲁对《麻省理工科技评论》表示：“即使你掌握了世界上所有的测谎记录，你也无法确定这些记录的真实性。”
The American Polygraph Association says the test is 80% to 94% accurate. At that level, a screening programme could still produce many errors. The Pentagon employs 2.8 million people, MIT Technology Review noted. Kyri Kotsoglou, a legal scholar at Northumbria University in the UK, calls adding AI to the polygraph “the worst of both worlds”. It adds uncertainty on top of invalidity, Kotsoglou told MIT Technology Review.
与科茨奥格鲁共同撰写相关研究的法学教授玛丽昂·奥斯瓦尔德（Marion Oswald）也指出：“即使你拥有所有相关的测谎数据，你仍然无法确定这些测谎结果是否准确。”
“Even if you have all the records in the world from polygraph tests, you don’t know whether those polygraph tests are right or not,” Marion Oswald, a law professor who has written with Kotsoglou, told MIT Technology Review.
鹿特丹伊拉斯姆斯大学研究欺骗行为的索菲·范德泽（Sophie van der Zee）认为，人工智能可以将多种测谎指标综合成一个更难以被欺骗者操纵的评估结果。她对《麻省理工科技评论》说：“目前仍然没有一种能够像‘匹诺曹的鼻子’那样（通过明显的外部信号）判断一个人是否在说谎的可靠方法。”
Sophie van der Zee, who studies deception at Erasmus University in Rotterdam, said AI could combine several measures into one score that is harder to game. “There is still no Pinocchio’s nose,” van der Zee told MIT Technology Review.
《麻省理工科技评论》还指出，此前开发的基于人工智能的测谎系统并未取得长期的成功。欧盟曾资助过名为iBorderCtrl的边境监控项目，该项目试图通过视频分析来检测人们的欺骗行为，但最终也失败了（就像美国的AVATAR边境监控项目一样）。
Earlier AI lie detectors have not lasted, MIT Technology Review reported. The EU funded iBorderCtrl, a border pilot built on a system that scored deception from video. It faded away, like the US AVATAR border project.
目前，其他利用人工智能来识别人类行为的工具已经在实际应用中：例如美国监狱中使用的AI语音分析系统、英国商店中的面部识别警报系统，以及美国警方使用的AI摄像头监控系统等。
Other tools that use AI to read people are already in use, from AI voice analysis of US prison calls to facial recognition alerts in UK shops and AI camera searches by US police.