伦敦：今年早些时候，在盛夏的深处，埃德肖·戈达德律所发现自己陷入了一个小麻烦。正如一种日益熟悉的模式所示，该所在其人力资源部门发布的一则IT职位广告吸引了大量申请。
LONDON: Earlier this year, in the depths of summer, the Addleshaw Goddard law firm found itself in a slight pickle. In what was becoming a familiar pattern, an advert it had placed for an IT job in its human resources department had drawn a slew of applications.
一年前，只有几十人可能会为类似职位投递简历。但人工智能让求职申请变得更快更简单，这一次竟收到了178份申请。
A year ago, only a couple of dozen people might have sent in a resume for a position like this one. But AI has made the job of applying for a job much faster and easier, and this time 178 applications arrived.
问题在于，筛选这些申请所花的时间越长，优秀候选人被其他公司挖走的可能性就越大。就像许多组织一样，埃德肖·戈达德一直在鼓励其管理层测试人工智能如何安全地改进其流程，因此该所的人力资源部门决定进行一次实验。
The problem was, the longer it took to sift them, the higher the chance that good candidates would find work elsewhere. Like a lot of organisations, Addleshaw has been encouraging its managers to test ways that AI can safely improve its processes, so the firm’s HR department decided to do an experiment.
他们想看看一个人工智能代理在评估申请方面相比人类的表现如何。
It would see how well an AI agent could assess the applications compared with humans.
该代理于是被设置起来，使用该所内部的微软Copilot系统，并被告知需要填补的职位，以及申请人需要具备的某些人力资源系统和编程语言的经验，还有对人工智能本身的了解。
AI AND HUMANS PREFERRED DIFFERENT APPLICATIONS An agent was duly set up, using the firm’s internal Microsoft Copilot system, and told about the role that needed to be filled, as well as the experience applicants would need with certain HR systems and programming languages, and with AI itself.
该所的人力资源人员将申请人筛选出大约十几人的名单，并让该代理对他们进行评分。
Humans at the firm sorted the applicants into a shorter pile of about a dozen and asked the agent to score them.
该代理迅速完成了这项任务，并呈上了一份它认为的顶尖候选人名单。人力资源团队也做了同样的事情。
The agent whipped through this task and presented a list of what it deemed the top contenders. The humans did the same thing.
可惜的是，两份名单并不一致。该代理对那些声称在之前雇主处在极短时间内实现了卓越IT业绩，或尽管经验甚少却参与了大型项目的候选人的华丽简历印象深刻。简而言之，这些申请太好以至于不真实——而且明显带有使用人工智能撰写的痕迹。
Alas, the two lists did not match. The agent was impressed by glossy applications from candidates who claimed to have achieved exceptional IT feats in record time at a previous employer, or worked on big projects despite scant levels of experience. In short, they were too good to be true – and smacked of having been written with the help of AI.
埃德肖·戈达德人力资源部主管凯瑟琳·拉思伯恩告诉我，她的团队对此产生了一些假设。
Katherine Rathbone, Addleshaw Goddard’s head of HR, told me her team had theories about why this might have happened.
首先，AI工具可以让候选人更轻松地根据职位描述定制简历。但这还不是全部。“一种可能的解释是，由于AI能力本身就是评判标准之一，代理可能将简历制作过程中出现的AI使用迹象视为AI能力的证据，”她说。使用与招聘人员相同的AI？
For one thing, AI tools can make it a lot easier for candidates to tailor a CV to match a job description. But that was not all. “One possible explanation is that, because AI capability was itself one of the criteria, the agent may have interpreted signs of AI use in producing the CV as evidence of AI capability,” she said. USING THE SAME AI AS THE RECRUITER?
还有另一种更令人不安的可能性。AI代理可能仅仅因为简历涉及AI参与而主动偏好这些简历。
There is another, more disturbing possibility. The AI agent may have actively preferred CVs purely because AI was involved in writing them.
如果情况属实，这将证实一份引人注目的工作论文的发现，该论文分析了AI招聘工具的行为。
If so, it would confirm the findings of an eye-catching working paper that has analysed the behaviour of AI hiring tools.
研究人员模拟了他们所说的24个职位（从教师和顾问到厨师和工程师）的真实招聘流程。他们发现，使用与招聘人员相同的AI大语言模型的候选人，相比依赖人类制作简历的候选人，被入围的可能性高出23%至60%。
The researchers simulated what they said were realistic hiring processes for 24 jobs ranging from teachers and consultants to chefs and engineers. They discovered that candidates using the same AI large language model as the recruiter were between 23 per cent and 60 per cent more likely to be shortlisted than those who relied on CVs produced by a hapless human.
这种人类劣势在销售和会计等商务岗位上最为明显，而在农业和汽车等领域则最小。
This human handicap was highest for business roles such as sales and accountancy and lowest for fields such as farming and automobiles.
该论文仍在修订中，以期在同行评审期刊上发表，这意味着其中一些发现可能会发生变化。但其中一位作者、徐建南告诉我，更新后的分析仍在提供证据，表明AI模型“表现出自我偏好行为”。招聘军备竞赛中的令人不安的消息这是令人不安的消息，因为此时AI已经在推动一场令人沮丧的招聘军备竞赛。一方面是求职者，他们相信自己正在与一堆毫无灵魂的招聘工具作斗争，于是通过大量使用AI应用进行反击，这些应用正在淹没招聘人员。
The paper is still being revised for publication in a peer-reviewed journal, meaning some of its findings may change. But one of its authors, Jiannan Xu, told me that updated analyses continue to provide evidence that AI models “exhibit self-preferencing behaviour”. DISTURBING NEWS IN THE HIRING ARMS RACE That is disturbing news at a time when AI is already driving a dismal hiring arms race. On one side are job hopefuls who, convinced they are battling a wall of soulless recruitment tools, are fighting back with an explosion of AI applications that are swamping recruiters.
如果AI对AI存在偏见，这就增加了一个新的风险维度。想象一下，如果你是使用了错误类型AI的候选人，或者根本没有使用AI的候选人，那会怎样。
If AI is biased towards AI, it adds a fresh dimension of jeopardy. Imagine being the candidate using the wrong type of AI, or no AI at all.
幸运的是，徐的论文表明，通过两种相对简单的策略，AI偏见可以减半以上：一是要求模型忽略简历的来源，专注于内容；二是使用多个评估模型来稀释单一工具的偏见。
Fortunately, Xu’s paper suggests AI bias can be more than halved with two relatively simple strategies: telling models to ignore the origin of resumes and focus on their content, and using a number of assessment models to dilute the bias of any single tool.
与此同时，在Addleshaw Goddard，我很高兴得知，IT岗位的成功候选人是由人类选出的，而不是代理排名最高的那个人。
Meanwhile at Addleshaw Goddard, I was pleased to learn that the successful candidate for the IT job was chosen by humans, and was not the one the agent ranked highest.
至于在招聘中使用AI，这仍然是一个正在进行中的工作。Rathbone表示，这些工具如果使用得当，显然具有潜力。“但人类判断仍然很重要。”
And as for using AI for recruitment, it is still a work in progress. The tools clearly have potential if used correctly, says Rathbone. “But human judgment remains important.”