最新的美国与英国劳动者（数千人）调查指出，AI 使用与薪酬与教育程度高度相关：高薪、受过高等教育者明显更常使用 AI，这意味著既有生产力与收入差距可能被放大。文章将之与实验结果区分开来；实验多集中于同一职业或公司内部，且常假设 AI 在职场全面导入；而民意调查显示跨职业采用不均。只要 AI 在知识型工作中应用更广、而在较例行的白领或蓝领任务中有限，则行业间的不均衡采用可同时缩小部门内差距、却扩大整体社会层次差距。

受访者对未来就业影响的预期也呈不对称：以本行业或本职位观点看，38% 认为未来五至十年职位流失多于新增，17% 则认为新增更多；但谈到整体经济，悲观升至51%，而认为新增更多者降为18%。另一个关键讯号是培训不足；仅14% 的受访者曾接受雇主提供的正式 AI 训练，另有21% 仅获得非正式指引。更显失衡的是，约25% 的英国受访者与接近三分之一（约33%）的美国受访者表示每个工作日都有使用 AI 工具。

职能差异进一步放大了态度落差：以写程式为主的开发人员与分析师对 AI 的看法最正向，而专业作家（含 copywriter、content creator、translator、writer、editor）是对 AI 最负向群体，甚于不使用 AI 的白领与蓝领群体。软体与分析工作中程式编写常属例行环节，AI 可削减重复劳动；但写作本身即为主要产出与创作核心，过度自动化可能稀释品质与声誉。研究对自我报告的生产力提升仍需警惕：在软体资深开发中证据较强，但其他领域仍不稳定。Sullivan & Cromwell 近期揭露高阶案件文件中的 AI 幻觉错误，显示高薪用户过度依赖风险。Ioana Marinescu 及其对 AI saturation 的研究提醒，高薪 AI 使用者或是「短期赢家」，但长期可能因可完全远端的智慧型工作被替代而成为脆弱者；劳动可能转向更多实体工作。

A new poll of thousands of workers in the United States and the United Kingdom found that AI use is much higher among higher-paid, better-educated employees, suggesting a potential widening of existing productivity and income gaps. The article contrasts this with firm- and occupation-level experiments, which typically report reduced inequality inside the same team or occupation. A key difference is scope: experiments often assume near-universal adoption, while the survey shows uneven adoption across job types. If AI is concentrated in knowledge-intensive roles and less present in routine white- or blue-collar work, inequality can fall inside some occupations but rise between occupations at the economy level.

Perceptions of job impact were also asymmetrical. In respondents’ own industry or job, 38% expected more jobs to be lost than created over five to ten years, while 17% expected the opposite; most chose “little difference.” At the economy level, pessimism increased: 51% expected more jobs lost than created, versus 18% expecting more jobs created. Training signals were similarly uneven. Only 14% reported receiving formal AI training from employers, and another 21% received only informal guidance. Yet around 25% of UK workers and almost one-third (about 33%) of US workers reported using AI tools on a typical workday.

Task composition produced a clear divide in attitudes. Developers and analysts, whose tasks include coding, were most positive toward workplace AI; professional writers (copywriters, content creators, translators, writers, editors) were the most negative, even relative to non-users and blue-collar workers. In software and analytics roles, AI often automates routine code production and frees time for higher-value work, while for writers AI can erase core output and weaken craft quality. The article also cautions that self-reported productivity gains are not equally validated: confidence is strongest for senior software developers. A Sullivan & Cromwell filing with AI “hallucinations” illustrates reputational risk from over-reliance. Ioana Marinescu warns that well-paid AI users may be temporary winners, and that fully remote intelligence work may eventually be replaced by AI, pushing workers toward more physical in-person roles.