The AI penalty: employees penalized for using AI in good faith

      Companies are encouraging employees to rely on AI while attributing the credit to the technology itself. Researchers refer to this as the AI penalty, and workers claim it is affecting their chances for promotions and salary increases.

      Aubrey dedicated over a year to a project aimed at optimizing an expensive medical manufacturing process. Upon completion, her manager requested that she present it to senior leadership as if the AI chatbot, Claude, were responsible for the work. She reached a compromise by promoting the bot while emphasizing her own contributions. Despite this, her manager interrupted her presentation to declare that she had completed the entire project in just a minute with AI. Subsequently, her annual review was only lukewarm, and her boss acknowledged that this incident negatively impacted her assessment.

      She is not the only one facing this issue, as noted by Business Insider’s Shubham Agarwal. Deepak, an IT developer at a Fortune 500 company in India, began to openly acknowledge the coding agents he utilized for the sake of transparency. However, soon his managers began to credit the AI for his achievements, which he believes hindered his promotion prospects.

      Many white-collar professionals find themselves in a conflict between employers who demand greater AI use and those who penalize AI users. This has led some to conceal their reliance on AI.

      This inclination to hide their use of AI seems rational. Christoph Riedl, a professor at Northeastern University, conducted a meta-analysis of 13 studies across various occupations and identified a consistent trend: managers tended to rate work lower once employees acknowledged the assistance of a chatbot, assuming that the AI had done the majority of the work. Riedl refers to this phenomenon as the “AI penalty.”

      He discovered that the primary way to avoid this penalty is to maintain control over the key task and clearly delineate one’s contributions. However, this is challenging as managers increasingly implement blunt methods to monitor AI usage in a job market already unsettled by automation.

      Measuring creativity through tokens is ineffective. The most straightforward metric is the token, the fundamental unit processed by an AI model. Tracking tokens reveals how frequently an employee interacted with a chatbot and the volume of text exchanged, without providing insight into the AI’s actual contributions. Employees quickly learned to manipulate this system, producing irrelevant queries to appear as proficient users, prompting companies to take steps to limit this behavior.

      Recently, Amazon discontinued an internal leaderboard that ranked token usage. "Please don’t use AI just for the sake of using AI," stated senior vice president Dave Treadwell during a company meeting.

      Even more refined metrics can be misleading. Coding assistants like Claude Code attribute a co-author label to the code they assist in generating but do not indicate which lines are theirs or how much they were influenced by human input.

      Riedl notes, “If AI use is disclosed without specific details about how it was used, the manager’s default assumption seems to be that it was used in a way that reduces agency.” In simpler terms, supervisors presume the AI took the lead. Hence, he emphasizes that the specifics of how AI was used are crucial.

      Some researchers are attempting to clarify the distinction between human and machine contributions. Graham Neubig, a computer scientist at Carnegie Mellon University, co-founded OpenHands, an open-source coding platform that annotates any lines written by AI, ensuring reviewers examine them closely.

      An IBM team took this concept further by creating an AI Attribution Toolkit, inspired by the systems scientists use to attribute credit in academic papers. This tool allows users to record how much a chatbot generated versus what a human verified, and subsequently generates an attribution statement.

      High-level recognition of AI contributions is insufficient, according to Jessica He, one of the toolkit’s developers.

      The deeper issue is social. Several studies indicate that disclosing AI usage, even truthfully, diminishes colleagues' trust and leads them to perceive the user as lazy. Oliver Schilke, a professor at the University of Arizona, found similar results in his research. He describes this as a paradox: those who act honestly end up facing negative consequences. He advocates for standardized rules regarding AI attribution rather than leaving it to individual workers’ assumptions.

      Thomas Prommer, an engineering executive at Adidas, has witnessed mandatory attribution backfire. He notes that his engineers completely stopped utilizing AI, stating, “because they didn’t want their best contributions footnoted as ‘co-written by Claude’.” Instead, he found success when the focus was on crediting the results rather than the tools used.

      The implications extend beyond missed opportunities for raises. Earlier this year, Amazon was criticized for holding employees accountable for mistakes made by an AI agent, resulting in layoffs. Deepak remarked, “The praise goes to AI, but we are responsible for reviewing its content.” Alessio Artuffo, CEO of the learning platform Docebo, contends that simplistic attribution is misguided.

      The critical issue is not the method of production but whether the individual behind it is capable of defending and rectifying it. If