Employers can encourage AI and still punish the people who admit using it. A credible promise about disclosure must reach the appraisal meeting.
On November 4, Hebrew University publicised research (https://phys.org/news/2026-11-chatgpt-boss.html) suggesting that employees who use AI can lose their supervisors' trust. The announcement drew attention to earlier research, rather than a new workplace experiment this week. It does not establish that workers are losing promotions. It does expose a consequential tension: an employer can welcome faster output while viewing the person producing it as less worthy of confidence. Approved AI use should be safe to disclose, and credit should follow the judgment people actually contribute.
The study by Samuel Meyers, Roman Briker and Yochanan Bigman (https://cris.iucc.ac.il/en/publications/psychological-reactions-to-subordinates-llm-use-at-work/) combines five incentivised economic games with an academic field setting. Its findings concern supervisory trust and evaluations, including perceived promotability. When an experimental organisation explicitly encouraged AI use, the reduction in trust was no longer statistically significant. That is a reason to test clearer rules, not proof that a policy announcement cures workplace prejudice.
An independent 2025 study by Duke researchers (https://scholars.duke.edu/publication/1674960), involving four preregistered experiments and 4,439 participants, also found negative judgments of AI users' competence and motivation. These studies cannot tell us how every manager will behave. Together, they make it harder to assume that providing software is enough to make adoption fair.
Imagine a junior analyst in an Indian services firm using an approved assistant to improve a client note. She checks the figures, rejects a misleading conclusion and takes responsibility for the result. If admitting that assistance makes her seem less capable, she faces a poor choice: surrender a useful tool or conceal how she works. This is a hypothetical example, but the incentive is straightforward. Demanding transparency while penalising it makes honest disclosure costly.
The loss could extend beyond a hurt reputation. Managers allocate difficult assignments, training and introductions. If disclosed assistance becomes a reason to withhold those opportunities, a worker may produce more today while gaining less experience for tomorrow. Senior colleagues with established reputations might be better placed to claim credit for directing the same tool. That unequal outcome is a risk to investigate, not a result the studies have established.
There is a real productivity opportunity to protect. Research on 5,172 customer-support agents (https://arxiv.org/abs/2304.11771), published in the Quarterly Journal of Economics in 2025, found that AI assistance improved the speed and quality of less experienced and lower-skilled workers. It also found evidence of improved English fluency, particularly among international agents. One deployment cannot establish universal benefits. It does show why dismissing assistance as laziness could obstruct useful learning and access.
For India, this matters in the difficult passage from education into credible professional work. The ILO–IHD India Employment Report 2024 (https://www.ilo.org/sites/default/files/2024-08/India%20Employment%20-%20web_8%20April.pdf) documents mismatches between education and employment and high unemployment among educated youth. That is historical labour-market context, not evidence of an AI effect. It strengthens the case for judging new entrants by demonstrated understanding, while ensuring that approved tools, accessible interfaces and training are available beyond well-resourced teams. Guidance should also be understandable to workers using different languages.
The serious objection is that supervisors have reasons to scrutinise AI-assisted work. A polished answer can conceal poor understanding, invented facts or mishandled confidential information. Protecting disclosure cannot mean immunity from scrutiny. Ask workers to explain important choices, show what they checked and demonstrate essential skills without assistance where the role requires them. Apply those expectations to senior staff too.
Three futures are plausible. In a better one, employers fund checking time and reward reliable judgment, allowing assistance to become a route to advancement. In an uneven one, established professionals gain efficiency while juniors lose credit for similar use. In a harmful one, firms demand rising output, punish admission of AI use and then blame employees when concealed practices fail. The difference lies partly in how managers are evaluated themselves.
Questions for founders and investors
Can someone disclose permitted AI use without an automatic downgrade? Do productivity calculations include licences, training, checking and rework? Are employees receiving better opportunities or merely higher targets?
Before the next appraisal cycle, employers should publish task-specific rules, explain how contribution will be assessed and offer a route to challenge inconsistent judgments. Investors should examine quality and progression alongside adoption figures. Workers and their representatives should help design those rules. A firm that wants the gains from assisted work must give credit to the person who used the tool, checked its work and made the decisions.
Sources
• Hebrew University research announcement, November 4, 2026; Meyers, Briker and Bigman, The Leadership Quarterly, institutional record.
• Reif, Larrick and Soll, PNAS, 2025, Duke institutional record.
• Brynjolfsson, Li and Raymond, Generative AI at Work, QJE 2025; accessible revised working paper.
• ILO–IHD, India Employment Report 2024.
On November 4, Hebrew University publicised research (https://phys.org/news/2026-11-chatgpt-boss.html) suggesting that employees who use AI can lose their supervisors' trust. The announcement drew attention to earlier research, rather than a new workplace experiment this week. It does not establish that workers are losing promotions. It does expose a consequential tension: an employer can welcome faster output while viewing the person producing it as less worthy of confidence. Approved AI use should be safe to disclose, and credit should follow the judgment people actually contribute.
The study by Samuel Meyers, Roman Briker and Yochanan Bigman (https://cris.iucc.ac.il/en/publications/psychological-reactions-to-subordinates-llm-use-at-work/) combines five incentivised economic games with an academic field setting. Its findings concern supervisory trust and evaluations, including perceived promotability. When an experimental organisation explicitly encouraged AI use, the reduction in trust was no longer statistically significant. That is a reason to test clearer rules, not proof that a policy announcement cures workplace prejudice.
An independent 2025 study by Duke researchers (https://scholars.duke.edu/publication/1674960), involving four preregistered experiments and 4,439 participants, also found negative judgments of AI users' competence and motivation. These studies cannot tell us how every manager will behave. Together, they make it harder to assume that providing software is enough to make adoption fair.
Imagine a junior analyst in an Indian services firm using an approved assistant to improve a client note. She checks the figures, rejects a misleading conclusion and takes responsibility for the result. If admitting that assistance makes her seem less capable, she faces a poor choice: surrender a useful tool or conceal how she works. This is a hypothetical example, but the incentive is straightforward. Demanding transparency while penalising it makes honest disclosure costly.
The loss could extend beyond a hurt reputation. Managers allocate difficult assignments, training and introductions. If disclosed assistance becomes a reason to withhold those opportunities, a worker may produce more today while gaining less experience for tomorrow. Senior colleagues with established reputations might be better placed to claim credit for directing the same tool. That unequal outcome is a risk to investigate, not a result the studies have established.
There is a real productivity opportunity to protect. Research on 5,172 customer-support agents (https://arxiv.org/abs/2304.11771), published in the Quarterly Journal of Economics in 2025, found that AI assistance improved the speed and quality of less experienced and lower-skilled workers. It also found evidence of improved English fluency, particularly among international agents. One deployment cannot establish universal benefits. It does show why dismissing assistance as laziness could obstruct useful learning and access.
For India, this matters in the difficult passage from education into credible professional work. The ILO–IHD India Employment Report 2024 (https://www.ilo.org/sites/default/files/2024-08/India%20Employment%20-%20web_8%20April.pdf) documents mismatches between education and employment and high unemployment among educated youth. That is historical labour-market context, not evidence of an AI effect. It strengthens the case for judging new entrants by demonstrated understanding, while ensuring that approved tools, accessible interfaces and training are available beyond well-resourced teams. Guidance should also be understandable to workers using different languages.
The serious objection is that supervisors have reasons to scrutinise AI-assisted work. A polished answer can conceal poor understanding, invented facts or mishandled confidential information. Protecting disclosure cannot mean immunity from scrutiny. Ask workers to explain important choices, show what they checked and demonstrate essential skills without assistance where the role requires them. Apply those expectations to senior staff too.
Three futures are plausible. In a better one, employers fund checking time and reward reliable judgment, allowing assistance to become a route to advancement. In an uneven one, established professionals gain efficiency while juniors lose credit for similar use. In a harmful one, firms demand rising output, punish admission of AI use and then blame employees when concealed practices fail. The difference lies partly in how managers are evaluated themselves.
Questions for founders and investors
Can someone disclose permitted AI use without an automatic downgrade? Do productivity calculations include licences, training, checking and rework? Are employees receiving better opportunities or merely higher targets?
Before the next appraisal cycle, employers should publish task-specific rules, explain how contribution will be assessed and offer a route to challenge inconsistent judgments. Investors should examine quality and progression alongside adoption figures. Workers and their representatives should help design those rules. A firm that wants the gains from assisted work must give credit to the person who used the tool, checked its work and made the decisions.
Sources
• Hebrew University research announcement, November 4, 2026; Meyers, Briker and Bigman, The Leadership Quarterly, institutional record.
• Reif, Larrick and Soll, PNAS, 2025, Duke institutional record.
• Brynjolfsson, Li and Raymond, Generative AI at Work, QJE 2025; accessible revised working paper.
• ILO–IHD, India Employment Report 2024.