Bosses trust AI instead of employees, delaying promotions and raises

When Aubrey finished a big work project earlier this year, her manager emailed her an unusual request. In a presentation to senior executives, could she highlight how Claude, an AI chatbot, helped her?

Aubrey, a New York-based healthcare analyst, has spent countless hours developing new ways to speed up the manufacturing process of expensive medical products. She used Claude in a small role, but her manager wanted her to sound as if the AI had come up with the idea and executed it automatically.

“I spent over a year gathering issues, drafting alternatives, and learning the impact of changes, and my manager wanted to blame it all on AI,” says Aubrey, who asked that her last name be omitted for fear of retaliation. She chose a middle path. In the presentation I had planned, I exaggerated the role of the AI but said I did much of the heavy lifting. But in the middle of her presentation, her manager interrupts her and announces that she built everything in one minute using AI. A few weeks later, Aubrey received her less than enthusiastic annual evaluation. Her supervisor did not directly mention the incident during the review, but later said it was a factor when she asked.

Deepak, an IT developer at a Fortune 500 technology company based in India, recently found himself in a similar situation. More than a year ago, he began regularly evaluating the automated coding agents he had deployed to perform menial tasks for transparency, but he quickly suspects that higher-ups at work began to attribute all of his positive contributions to the AI, which delayed his expected promotion.

Deepak and Aubrey are not alone. White-collar workers are caught in a terrible pickle. If we acquiesce to our bosses’ demands for more AI, will they trust that AI will do the job for them? Are we building the guillotine of our careers? As a result, many employees are hiding their AI usage and wondering how much, if any, they should be evaluating their efforts.

In an age of mass layoffs driven by AI, whether or not we give credit to AI may feel as important as whether or not we do so.

Christoph Riedl, a professor of information management at Northeastern University, says it’s entirely legitimate for people to be hesitant about disclosing AI assistance. In a recent meta-analysis, Riedl and co-authors examined 13 studies across a variety of jobs and positions to assess how managers treated employees after they disclosed their use of AI. The conclusion was clear. When employees revealed AI support, managers did not consistently evaluate their contributions to projects. Administrators believed technology would do most of the heavy lifting. One of the few ways people avoided this “AI penalty” was to maintain ownership over their core work and explain exactly how they contributed to the task. But while this may be easier said than done, the way employers track AI usage can obscure how much influence humans retain over its creative output.

Currently, most companies rely on tracking tokens, which are the basic units of data processed by AI models. By seeing the number of tokens used by employees, managers can see how often the chatbot is queried, how much information is exchanged, and the length of each interaction. However, it does not provide insight into the creative contribution of AI to the project. So anyone can ask a company-issued chatbot endless questions about the weather or their personal life and still appear to be a professional AI user. It didn’t take long for companies to realize that it was counterproductive and could thwart frivolous token maxing. Last month, Amazon shut down its internal leaderboards that tracked AI token usage. This is because they forced staff to perform tasks that did not necessarily solve the problem.

“Don’t use AI just for the sake of using AI,” Dave Treadwell, Amazon’s senior vice president, told staff during a company-wide meeting.

Even more advanced methods can be cumbersome. AI coding assistants like Claude Code go so far as to automatically add co-author signatures to the code they write, without explicitly pointing out which lines were auto-generated or how extensively the human author was involved.

“Our analysis shows that when the use of AI is made public without specific details about how it was used, administrators’ default assumption appears to be that the AI was used in a way that reduces agency,” Riedl told me. In other words, the boss assumes that the bot must be the driving force behind that new product feature, quick software fix, or long report text. “So the details of how the AI was used appear to be very important.”

There are several attempts to better understand the exact balance between human and AI contributions. Graham Neubig, a computer science professor at Carnegie Mellon University, co-founded OpenHands, an open-source AI coding platform that adds footnote-like attributes to lines of code generated by AI. Neubig felt it was important to tag code as AI-generated to reduce the level of trust and increase scrutiny by reviewers.

The team at IBM has created a more detailed way to track contributions. The AI Attribution Toolkit is inspired by CRediT (Contributor Role Taxonomy), a standardized system used by scientists to outline the exact contributions of each author in published papers. The AI Attribution Toolkit form allows you to enter how much of the work was auto-generated, whether a chatbot created the content from scratch, and whether certain elements were reviewed by a human. This tool generates attribution statements that users can add to documents, code, etc.

Jessica He, one of the toolkit’s designers, said a high level of approval for AI use is not enough for both the people consuming AI-assisted content and the AI users. She added that people may interact with a work differently depending on whether AI is used to generate new ideas or simply refine language, and that “if the use of AI is restricted, users may feel that attribution infringes on their ownership rights.”

As companies and researchers seek to better attribute credit and responsibility for work done by machines, people are grappling with the very human assumptions made by their bosses and colleagues. Studies show that even with the best of intentions, disclosing AI can erode trust among co-workers and lead to people using AI being perceived as lazy.

Oliver Silke, a professor of business and sociology at the University of Arizona, whose research has found that simple acts of disclosure can erode public trust, agrees that one of the central contradictions of this new era of work is the tension between companies’ urge to deploy AI to improve efficiency and the social costs of implementing AI. So far, Schilke said, the burden has fallen on individual users, who have to decide how much to involve AI and how much of that involvement to reveal, “creating a paradoxical dynamic where those who do the morally right thing have to be penalized for transparency.” A better alternative, he adds, is a collective AI governance discipline that includes tools such as an attribution toolkit.

Adidas engineering executive Thomas Plomer has noticed a similar pattern within his team. Attribution obligations to AI sound fair, but they quietly killed engineers’ initiative. They stopped reaching for AI tools. “We didn’t want our best contribution to be footnoted as ‘co-authored by Claude,'” he says.

“The signal the AI sent was, ‘With the help of AI, there will be less work, so people hid it or avoided it,'” Plomer said. What worked was evaluating the results rather than the tools. Regardless of how much of the work is done by AI, those responsible will ultimately be praised and held accountable.

Many people, including Aubrey, are starting to wonder if AI is just a program like Excel, or if we should be celebrating AI at all. Silke disagrees with this idea. While traditional tools like Excel are designed to carry out human instructions within a limited and predictable range, AI can generate sentences, code, and ideas, often with little human input. Schilke adds that rather than just assisting with implementation, “they can make a substantial contribution to the form and content of the work, changing the meaning of disclosing its use.”

“The question becomes whether clear intellectual contributions can still be attributed to human authors, or whether agency was primarily outsourced to machines,” Silke said.

For many workers, there will be greater anxiety about AI mistakes. Earlier this year, it was revealed that Amazon had fired an AI agent, blaming it on humans for mistakes.

“AI is praised, but it is our responsibility to check it and if there are any errors, that too will be recorded as our responsibility,” Deepak adds.

Alessio Artuffo, CEO of learning platform Docebo, says simple attribution is the wrong framework. The question is no longer about how exactly the work was created, but whether those responsible for it can defend it, improve it, and take responsibility when it fails.

Ultimately, researchers and organizational experts agree that if companies want their employees to leverage AI in creative and productive ways, they need to create an environment where AI proficiency is something worth developing, rather than something at risk of being ignored or stigmatized.

“The more serious cost is psychological,” Artufo added. “If your employees are producing more output but feel less ownership of their work, that’s not a victory. It’s a loss of competency disguised as efficiency.”