Honesty About Workplace AI Still Draws a Career Penalty

A workplace AI experiment found that U.S. knowledge workers rated a peer who disclosed AI help as 10 times lazier, even when the finished work was identical. They were also 24 percentage points less likely to recommend that person for a high-visibility project.

Companies keep getting told to write clearer rules so coworkers can trust each other. The harder finding is that honesty is what gets punished, while HR already reads company prompt logs and model makers have started leaving machine receipts in the text.

Disclosing AI Use Gets You Labeled Lazy

Between March 30 and April 7, 2026, Atlassian’s Teamwork Lab ran a controlled experiment of 961 workers, all full-time U.S. knowledge staff. Everyone judged the same email summarizing a business proposal from the same imaginary colleague. The only change was a short note about how the work was done: no mention of AI, or an AI disclosure that varied in detail and motive.

When AI use was disclosed, raters scored the creator as lazier, less principled, and less worth sending to a senior leader. A companion pulse survey of 1,006 U.S. knowledge workers, taken April 24 to May 6, 2026, found 94% had used AI at work in the prior month, and two-thirds had done so regularly. Roughly 3 in 4 said they were open about that use with managers and peers.

THE AI PENALTY IN ONE TEST

• The laziness hit: Identical output, plus an AI disclosure, produced a 10 times lazier rating versus silence.
• The career hit: Disclosure cut the chance of a high-visibility recommendation by 24 percentage points.
• The softer pitch: Framing the tool as help for the team and client lifted effort scores from 45% to 56% and recommendations from 43% to 51%.
• The floor: Both disclosure styles still ranked well below people who never mentioned AI.

Molly Sands, head of the Teamwork Lab, has been watching companies tell staff to use the tools and then watching those same staff punish the people who admit it.

This isn’t a technology problem; it’s a social and cultural one. Companies are telling the workforce to use AI, but employees are penalizing each other for being honest about it. Until leaders change the culture, you don’t have an AI strategy; you have an AI contradiction.

Molly Sands, Head of the Teamwork Lab, Atlassian

In teams that actively celebrate AI use, with leaders who model it and wins that get named, the laziness stigma nearly disappears, and disclosed users can even be rated more efficient than silent peers. Sands says that pattern is still concentrated in tech-forward pockets. In her own industry, the person who never uses AI is the one who has to explain themselves. For most workers, the same openness is a liability.

More Than Half of Workers Still Hide the Assist

The Atlassian pulse is a U.S. knowledge-work snapshot from spring 2026. A wider baseline sits in the University of Melbourne and KPMG global study of 48,340 people across 47 countries, fielded from November 2024 to mid-January 2025.

In that sample, 58% of employees regularly used AI for work, and 31% did so daily or weekly. Free public tools were the main route (70%) rather than employer-provided systems (42%). Only 34% said their organization had a policy or guidance on generative AI. Forty-four percent said they had used AI in ways that broke those rules anyway. Sixty-six percent used model output without checking accuracy, and 56% said they had made work mistakes because of AI. Fifty-six percent had used the tools without knowing whether they were allowed. About 48% to 49% had uploaded sensitive company material, including financial, sales, or customer information, into public systems.

Over half, 57%, admitted they had hidden AI use or presented AI-generated content as their own. That is a different population and a different question from Atlassian’s “are you vocal with your manager” item, and both can be true at once: some workplaces now talk about the tools in the open, while a large global share still treats the assist as something to bury.

FOUR SURVEYS, FOUR VERSIONS OF THE GAP

Founder Reports, which surveyed 2,078 employed U.S. adults in April 2026, also split the trust drop: 11% said they trusted the work much less, 32% a little less, 20% trusted it more, and 37% said it made no difference. Workers under 40 were more skeptical (48%) than those 50 and over (34%).

The Extra Review Load Falls on the Careful Ones

Once a teammate knows AI was in the mix, the file does not travel the same path. In the Founder Reports sample, 77% reviewed a coworker’s work more carefully, and 36% said they reviewed it much more carefully. Daily AI users were not kinder on this point: 80% of them still tightened the review. Forty-five percent had already had to fix or redo a coworker’s work that leaned too hard on a model, including 57% of managers. Where employers required AI use, that fix rate rose to 73%.

That is the quiet tax on the people who still read the work. Speed shows up in the sender’s calendar. Rechecking shows up in someone else’s.

WHAT DISCLOSURE CURRENTLY BUYS A TEAMMATE

• A lazier label: Atlassian’s raters punished disclosed AI use even when the email was the same.
• A smaller stage: They were 24 percentage points less willing to send that person to a senior leader.
• A slower inbox: 77% of Founder Reports respondents scrutinized AI-assisted work more than fully human work.
• A redo queue: 45% had already cleaned up a coworker’s AI-heavy output, and managers carried more of that load.

On engineering teams, the same split is already a daily workflow. One group prompts a ticket, pastes the answer, and stays barely on the machine. The other group inherits 15 pull requests, then a 20,000-line diff, and answers Slack threads that were themselves drafted by a model. Pedro Domingos, a University of Washington computer science professor, boiled the incentive problem down to one line: “Simple: reward the lazy less and the craftsmen more.”

Microsoft’s 2026 Work Trend Index, covering 20,000 workers in 10 countries, shows why people keep dancing. Sixty-five percent of AI users fear falling behind if they do not adopt quickly, yet only 13% say they are rewarded for using and experimenting with AI. Only 26% say leadership is clearly and consistently aligned on AI. Microsoft’s modelling attributed 67% of the variance in AI impact to culture, manager behaviour, and talent practices, against 32% for individual mindset.

HR Already Reads Company AI Prompt Histories

While coworkers argue about etiquette, the official tools are already leaving a trail. BambooHR’s Redesigning Work survey, fielded June 26 to July 15, 2026 among 1,608 full-time U.S. desk workers, including 520 HR professionals, and published on September 1, 2026, found that HR professionals review prompt histories in 89% of cases.

Half of workers (50%) did not realize their employer could see the questions they had asked company-provided AI. Forty-four percent were trying to cover personal-account use by deleting prompts right after they typed them. Among HR staff who had reviewed those histories, 63% had found people using paid company tools to job-search, run side businesses, or chase other income, and 49% had found general personal tasks.

The boundary between work and home accounts is already gone. Fifty-nine percent of workers used personal AI accounts for work. Among that group, 71% had entered client data, proprietary strategy, or other sensitive company information into tools their employer cannot see. Sixty-two percent used company AI for personal purposes, including 17% for a side business or freelance work and 27% for a job hunt. Fifty-four percent of organizations still lacked a clear, documented AI usage policy that was consistently communicated.

Time spent with the tools is not the same as time saved. Workers averaged 87 minutes a day with AI, or 22,526 minutes a year, about 47 eight-hour workdays. Forty-two percent of that time went to troubleshooting and prompt iteration, against 35% that furthered the actual workload. That troubleshooting slice was 36 minutes a day, 9,461 minutes a year, about 20 workdays. VPs and C-suite staff spent 101 minutes a day with AI, against 54 minutes for individual contributors. Thirty-five percent said knowledge transfer in their company now ran primarily through AI rather than people, and 75% said AI had changed how they actually do the job.

Claude Now Leaves a Fingerprint in the Text

Peer suspicion is a guess. Model makers are building a receipt. Anthropic’s August 14, 2026 note on how Claude’s text watermark works says future Claude models generate text with a statistical watermark, a version of Google DeepMind’s SynthID-Text method from a 2024 Nature paper. Nothing visible is added to the page. The model still picks among likely next words, but the source of that randomness is a key, so someone with the key can score the chance that Claude was involved.

The watermark does not name a user, a company, or a chat. It does not prove a human wrote the rest. It is weak on short samples, on tightly factual passages where the next word is not a free choice, and on code that has to be exact. Light editing probably will not wipe it; a full rewrite will. Anthropic is rolling this out to comply with the EU AI Act, alongside other major providers and about 190 signatories of the EU Code of Practice on Transparency of AI-Generated Content, signed in July 2026. The firm applied the mark worldwide at launch because it did not yet have a durable way to limit it by region.

For files such as PNG, JPG, and SVG, Claude attaches a C2PA content credential, a signed metadata note that the file was made or processed with Claude. That is a different mechanism from a text watermark: the pixels do not change. C2PA’s implementation guide for Content Credentials spells out how those manifests can flag synthetic content, including an AI disclosure assertion and, where useful, which region of a file was touched.

THE MARKING CALENDAR

1. July 2026: Major model providers and about 190 other signatories back the EU Code of Practice on marking AI-generated content.
2. August 2, 2026: The EU marking duty applies, and new Claude models begin carrying a text watermark, with C2PA credentials on supported files.
3. August 14, 2026: Anthropic publishes the public explainer on the method, its limits, and the difference from after-the-fact AI detectors.
4. September 1, 2026: Anthropic updates the note on a detection API in private preview for regulators, law enforcement, media, researchers, schools, EU civil-society groups, and enterprises that must verify marks for their own compliance.

That API is the missing piece for the coworker fight. A manager in a random U.S. firm cannot yet drop a Slack paragraph into a public Claude checker and get a clean yes. Claude text watermarks and C2PA file signing still settle a likelihood, not authorship, and they do not say who typed the prompt. Google’s watermarking under EU transparency rules follows the same dual stack, SynthID in the content and C2PA on the wrapper, which is built for platforms and regulators first, not for a stand-up argument about who drafted the brief.

Why Written AI Rules Keep Failing

A policy memo can still be worth writing. It cannot do the job the etiquette debate assigns it. In the Founder Reports sample, workers at companies that allowed AI with restrictions trusted AI-assisted output less (52%) than workers at companies that allowed it without restrictions (28%). Required use produced the highest redo rate on coworker output (73%). Rules, in other words, can raise the volume of suspicion rather than lower it.

Microsoft’s manager data points the other way. In a separate study of 1,800 workers, employees reported a 17-point lift in the value they got from AI, a 22-point lift in critical thinking about their AI use, and a 30-point lift in trust in agentic AI when managers visibly used the tools themselves. Frontier professionals were much more likely to say their manager openly used AI (85% versus 64%) and set quality standards for AI work (83% versus 57%). Atlassian’s culture finding matches that picture: the laziness penalty fades where leaders treat the tools as shared infrastructure, not a private cheat.

Until that happens, the rational move for a lot of staff is the one the surveys already show. Stay quiet, because disclosure costs status. Use a personal account, because the company log is readable. Paste the output into a doc, because a watermark detector is not in the intern’s hands. The people who follow the official stack get monitored. The people who tell the truth get rated lazy. The people who keep the secret still pass work to teammates who then spend their 20 troubleshooting days cleaning it up.

Sands’s line still holds: saying you used AI for the team can soften the stigma, but it does not erase it. The detection API Anthropic is previewing for obligated enterprises will, over time, make some of those secrets checkable. It will not decide who gets credit, who does the review, or whether a 10-times-lazier label was ever a fair read of the same email.