AI 工具让职场写作趋同,除非你是女性

AI 写作助手正逐渐成为职场日常沟通的一部分。员工用它们起草邮件、润色报告、准备演示文稿。不过,他们得到的回复会因提问措辞的不同而有所差异。
约翰斯·霍普金斯大学的研究人员开展了一项研究,考察不同沟通风格如何影响 AI 生成的专业写作。研究发现,当提示词中包含更常与女性相关联的特征——比如“maybe”“I think”这类模糊限定语,以及“we”“our”这类集体指代——时,生成的回复通常不够复杂、更易读,也更不正式。
这些发现为理解语言模型如何回应不同提示词提供了启示。不过,该研究并未证实这些差异是否会影响职场沟通,或影响用人决策。
“最大的变化是速度。人们可以在几分钟内从一张白纸变成一份可用的初稿,如果你用第二语言写作,或者只是觉得写作很慢,这会帮上大忙。但我不太常注意到的是,这些初稿开始听起来千篇一律:同样的开头,同样工整的措辞,同样略显平淡的语气。”SaaS 初创公司 Responsive 的人力运营负责人 Shipra Kamra 对 AIM 表示。
约翰斯·霍普金斯大学的研究人员测试了 Meta Llama 3.1 3B Instruct、Mistral 7B Instruct、Gemma 2 27B 和 GPT-4。他们在四个模型中都观察到反复出现的差异。这些效应在邮件和求职申请上最为明显,而在辞职信上的发现则较弱。该研究的分析还发现了正式程度上的差异。
研究人员还把提示词中使用的名字改为传统的男性和女性名字。名字线索对结果几乎没有影响,而语言差异依然存在。这表明,在受测试的提示词中,用于表述请求的语言比署名中包含的名字对回答的影响更大。
“很难揭示模型为何会产生某一特定输出。我们推测,在这种情况下,语言信号直接压过了来自名字的信号。我们还推测,这些模型可能经过训练,能够抵御与名字相关的偏见,但无法抵御更为细微的语言信号,”约翰斯·霍普金斯大学计算机科学助理教授 Anjalie Field 解释道。
在另一项初步实验中,研究人员让 GPT-4 比较生成的文档,并评估哪位作者看起来更聪明、更有能力。在涉及 GPT-4 生成回答的 105 次比较中,当评判模型表达出偏好时,它更常倾向于那些由包含男性关联特征的提示词所生成的回答。
总部位于班加罗尔的智库 Synergia Foundation 首席运营官 Sambratha Shetty 表示,这些发现凸显出在评估模型回答时,将语言特征与其他因素一并考察的重要性。
“模型主要回应的是请求的表述方式,而不是作者所声明的身份。这意味着传统的姓名盲审式审计可能会完全漏掉这一问题,”Shetty 指出。
她还提醒,不要把这一结果当作对当前 AI 系统的定论性评估。该研究考察的是特定版本的模型,其中包括一个较旧的 GPT-4 模型,而这些系统的行为在不同新版本之间可能存在差异。
研究人员也承认,他们修改了提示词,以进行受控比较。所考察的语言特征来自主要聚焦于性别关联语言的研究,并非某一性别所独有。
AI 生成的文字并不总是作为最终产出使用。员工在发送前可能会编辑草稿、调整语气或重写部分内容。
“如果一个工具不断把所有人推向某一种‘专业’版本(通常更正式、更强势、更复杂),那个版本就会慢慢变成标准。职场规范从来就不是中立的。它们是由当时掌握权力的人塑造的,而在既有文本上训练出来的模型会学到这一点,并大规模重复它。单独一条建议本身没什么危害。让我担心的是,数百万条建议全都指向同一个方向,”Kamra 观察道。
AI 写作助手是否会在不同组织和应用场景中一致地鼓励某种特定的沟通风格,仍有待考察。清晰度、准确性以及传达想法的能力在各岗位中可能都重要,而表达方式和语气在某些场景下可能尤为重要。
AI 生成的文本也在招聘和员工评估中得到应用。求职者可能会使用写作助手来准备申请材料和求职信,而管理者可能会用 AI 起草反馈或总结员工沟通内容。
Field 指出,此前关于 AI 辅助招聘的研究考察了与求职者姓名相关的偏见,但结果不一。约翰斯·霍普金斯大学的这项研究考察的是另一个方面:提示词中的语言线索是否会影响 AI 生成的文字。
“我们的结果表明,如果你用 AI 来帮助起草申请材料,比如求职信或写给潜在雇主的邮件,而你又没有注意提示词中的措辞,也没有留意 AI 为你起草的文本,那么 AI 可能并没有帮你做出最有力的陈述,”Field 提醒道。
这凸显出,生成的文本仍然需要经过审核,判断它是否准确、有效地传达了一段经历。
不过,该研究并未证明招聘人员会因为这些输出差异而对候选人做出不同的评估。
Kamra 也指出了这一区别。“一封主要由助手代笔的漂亮邮件,几乎说明不了这名员工的技能水平。而且,如果把候选人自己写的文字与经过 AI 润色的文本相比较,那些写作风格或语言背景不同的人可能会被不公平地打低分,”她解释道。
她建议,组织应就招聘和绩效评估中哪些环节使用 AI 制定明确的准则,保留人类对决策的问责,并考察这些工具在不同员工群体中的表现。
对于采用 AI 写作助手的组织,评估其表现可以包括考察它们如何回应表述方式不同的相似请求。这有助于企业了解,提示词的差异是否会导致生成文本在长度、正式程度、阅读难度或准确性上出现差异。
Shetty 建议采用配对提示测试(paired-prompt testing),即组织用直接、委婉、个人化和协作式的语言提交相似的请求。随后可以对比生成的回复,以找出不同写作风格中的规律。
她还认为,审查的严格程度应取决于具体应用。用于日常通信的工具,与用于起草绩效评估、职位描述或招聘推荐的工具,其影响并不相同。
Kamra 强调,在挑选和测试这些工具时,应让员工参与其中。
“员工应当感到自己可以无视某项建议,而且当某个工具以让他们觉得不对劲的方式改变其表达风格时,他们应当有地方可以反映。包容必须成为工具选择方式的一部分,而不是事后才补上的东西,”她说。
要确定 AI 生成写作中的差异是否会影响员工,所需的研究不能仅限于比较模型输出。纳入印度英语和其他职场语言的研究,有助于判断该研究中观察到的模式是否也会出现在不同的职业场景中。
就目前而言,Johns Hopkins 的发现提供了证据,表明在被测试的各个模型中,语言特征会影响 AI 生成的专业写作的形式。但这些发现并不能证明这些差异会导致人类判断或就业结果上的不同。
译文由英文原文机器翻译,术语按行业惯例统一,可在文末展开原文核对。
查看英文原文(AI Tools are Making Us Sound the Same at Work—Unless You're a Woman)
AI writing assistants are becoming part of everyday workplace communication. Employees use them to draft emails, refine reports, and prepare presentations. The responses they receive, however, can vary depending on how they phrase their requests.
A study by researchers at Johns Hopkins University examined how different communication styles influence AI-generated professional writing. It found that prompts containing features more commonly associated with women, including hedging phrases such as “maybe” and “I think” and collective references such as “we” and “our”, generally produced responses that were less sophisticated, more readable, and less formal.
The findings offer insight into how language models respond to different prompts. However, the study does not establish whether these differences affect workplace communication or influence employment decisions.
“The biggest change is speed. People can get from a blank page to a workable draft in a few minutes, which helps a lot if you're writing in your second language or simply find writing slow. What I notice less often is that those drafts start to sound alike: the same openings, the same tidy phrasing, the same slightly flat tone,” Shipra Kamra, Head of People Operations at SaaS startup Responsive, tells AIM.
When Language Shapes AI Outputs
The Johns Hopkins researchers tested Meta Llama 3.1 3B Instruct, Mistral 7B Instruct, Gemma 2 27B, and GPT-4. They observed recurring differences across the four models. The effects were strongest for emails and job applications, while findings for resignation letters were weaker. The study's analysis also found differences in formality.
The researchers also changed the names used in the prompts to traditionally male and female names. The name cues had virtually no effect on the outcomes, while the linguistic differences persisted. This suggests that, in the tested prompts, the language used to frame a request had a stronger influence on the responses than the name included in the sign-off.
“It's difficult to uncover exactly why a model produces a particular output. We speculate that in this setting, the linguistic signal simply overpowers the signal from the name. We also speculate that these models may have been trained to resist name-associated biases, but not more subtle linguistic signals,” explains Anjalie Field, Assistant Professor of Computer Science at Johns Hopkins University.
In another preliminary experiment, the researchers asked GPT-4 to compare generated documents and assess which author appeared smarter and more competent. Across 105 comparisons involving GPT-4-generated responses, the judging model favoured responses produced from prompts containing male-associated features more often when it expressed a preference.
Sambratha Shetty, Chief Operating Officer at Bengaluru-based think-tank Synergia Foundation, says the findings highlight the importance of examining linguistic features alongside other factors when evaluating model responses.
“The models responded primarily to how the request was expressed, not to the writer’s stated identity. This means that conventional name-blind audits could miss the problem entirely,” Shetty notes.
She also cautions against treating the results as a definitive assessment of current AI systems. The study examined specific model versions, including an older GPT-4 model, and the behaviour of these systems may differ across newer versions.
The researchers also acknowledge that they modified the prompts to create controlled comparisons. The linguistic features examined were drawn from research primarily focused on gender-associated language and are not exclusive to either gender.
The Question of Professional Communication
AI-generated writing is not always used as a final output. Employees may edit drafts, change the tone, or rewrite sections before sending them.
“If a tool keeps nudging everyone towards one version of "professional" (usually more formal, more assertive, more complex), that version slowly becomes the standard. Workplace norms were never neutral anyway. They were shaped by the people who held power at the time, and a model trained on existing writing will pick that up and repeat it at scale. One suggestion on its own does little harm. Millions of them, all pointing the same way, is what worries me,” Kamra observes.
Whether AI writing assistants consistently encourage a particular communication style across organisations and applications remains to be examined. Clarity, accuracy and the ability to convey an idea may be relevant across roles, while presentation and tone could carry additional importance in certain settings.
Implications for Recruitment, Employee Assessments
AI-generated text is also finding applications in recruitment and employee assessments. Candidates may use writing assistants to prepare applications and cover letters, while managers may use AI to draft feedback or summarise employee communication.
Field points out that previous research into AI-assisted hiring has examined biases associated with applicant names, with mixed findings. The Johns Hopkins study examines a different aspect: whether linguistic cues in prompts influence the writing AI generates.
“Our results suggest that if you use AI to help draft application materials, like cover letters or emails to potential employers, and you are not careful about the language in your prompt or monitoring the text the AI drafts for you, the AI may not be helping you make your best case,” Field cautions.
This highlights that generated text still needs to be reviewed for whether it accurately and effectively communicates an experience.
The study does not, however, demonstrate that recruiters assess candidates differently because of these output variations.
Kamra also notes that distinction. “A polished email that an assistant mostly wrote tells you very little about the employee's skills. And if a candidate's own writing is compared to AI-smoothed text, people with a different style or language background can be marked down unfairly,” she explains.
She recommends that organisations establish clear guidelines on where AI is used in recruitment and performance assessments, retain human accountability for decisions, and examine how these tools perform across different employee groups.
What Businesses Need to Examine
For organisations adopting AI writing assistants, evaluating their performance could involve examining how they respond to different ways of phrasing similar requests. This would help businesses understand whether variations in prompts lead to differences in the length, formality, reading level, or accuracy of generated text.
Shetty recommends paired-prompt testing, in which organisations submit similar requests using direct, hedged, individual and collaborative language. The resulting responses can then be compared to identify patterns across different writing styles.
She also argues that the level of scrutiny should depend on the application. A tool used for routine correspondence has different implications from one used to draft performance evaluations, job descriptions, or recruitment recommendations.
Kamra emphasises involving employees when selecting and testing these tools.
“Employees should feel able to ignore a suggestion, and they should have somewhere to say so when a tool is changing their voice in a way that feels wrong. Inclusion has to be part of how tools are chosen, not something added afterwards,” she says.
Establishing whether differences in AI-generated writing affect employees will require research beyond comparing model outputs. Research involving Indian English and other workplace languages could help determine whether the patterns observed in the study also appear in different professional contexts.
For now, the Johns Hopkins findings provide evidence that linguistic features can influence the form of AI-generated professional writing across the models tested. They do not establish that these variations lead to differences in human judgement or employment outcomes.
正文保留英文原文,标题与摘要由机器翻译。
原文发布于 Analytics India Magazine(AI Tools are Making Us Sound the Same at Work—Unless You're a Woman),标题、摘要与正文由机器翻译,术语按行业惯例统一,未经人工逐句校订。