Employees are seeing artificial intelligence tools produce dangerous, unethical, or biased outputs at work, but many do not feel safe raising concerns internally, according to new research that points to a growing governance problem inside companies adopting AI at pace.
Research from Writer, the enterprise AI platform provider, found that 28% of employees had witnessed AI tools produce outputs at work that were dangerously wrong, unethical, or biased. Despite that, three in ten employees said they did not feel safe reporting harmful or unethical AI behaviour internally because they feared retaliation.
The findings point to a widening gap between executive assumptions and employee experience. HRreview reported that 90% of executives believed employees were comfortable raising concerns about AI, suggesting senior teams may be underestimating anxiety around the technology.
The survey of 2,400 enterprise executives and employees also found that risky behaviour is already spreading. More than a third of employees admitted uploading confidential or sensitive company information into public AI tools, while 16% said they were using AI systems explicitly banned by their employer.
Among employees using prohibited tools, 40% said they were simply doing “whatever it takes” to complete their work. Others cited poor-quality internal AI systems and weak enforcement policies as reasons for using unapproved tools.
Many companies also lack clear visibility over how AI is being used. More than a third of executives admitted they did not have full visibility or control over which AI systems staff were using.
That is no longer a marginal IT issue. AI tools are now embedded in everyday work across writing, coding, analysis, customer communication, research, HR, finance, legal operations, and sales support. When employees use unapproved tools or feed sensitive information into public systems, the risk moves quickly from productivity experimentation into data protection, intellectual property, confidentiality, and regulatory exposure.
The survey also highlighted concern around AI agents. More than a third of executives said they were not confident they could quickly shut down AI agents if systems started causing reputational or financial damage, while 36% said their organisation lacked a formal documented plan for supervising agents.
Security and data protection were identified as the biggest governance concerns among leaders, followed by employee training, transparency, and explainability. Ethical alignment was identified as a major concern by only a quarter of executives, despite employees reporting harmful or biased outputs.
The findings sit within a broader shift from AI experimentation towards operational control. As companies move from AI adoption to AI advantage, governance, workflow redesign, and business outcomes are becoming more important than simply deploying tools. Writer’s research adds a cultural dimension: AI adoption is unlikely to mature if employees do not trust the organisation enough to report failures.
The silence risk is especially significant because many AI errors are context-specific. A model may generate a plausible but inaccurate response, apply flawed assumptions, misinterpret sensitive material, reproduce bias, or expose confidential data through a workflow that only the user can see. If the employee involved does not report the issue, the company may never know that its controls have failed.
Written policy will not be enough on its own. A policy can prohibit certain tools, restrict data inputs, or define approval routes, but employees will still work around the rules if approved systems are too weak, too slow, poorly integrated, or unavailable for the work they are expected to complete.
The findings also reveal a deeper leadership challenge. If employees believe reporting AI problems will make them look resistant to innovation, unproductive, or responsible for the error, companies will receive a distorted picture of AI performance. Senior teams may see adoption dashboards, licence usage, and productivity claims, while missing informal workarounds, suppressed concerns, and near misses inside teams.
Similar patterns have appeared before in cybersecurity and compliance. Employees often bypass official systems when the approved route does not match operational reality. AI raises the stakes because the tool may process sensitive information, create decision-support material, generate customer-facing content, or act through agents connected to enterprise systems.
Companies now need mechanisms that treat AI issues as operational learning rather than personal failure. Safe reporting routes, visible escalation processes, clear rules on confidential data, approved alternatives that work, and practical training all become part of the same control environment.
Boards and senior teams are also likely to need better evidence. AI risk cannot be managed only through statements of intent or central procurement policies. It requires visibility into usage, exceptions, employee concerns, agent permissions, and the quality of outputs in real workflows.
The next phase of AI governance will be shaped as much by culture as by technology. Companies can buy approved tools, restrict access, and publish policies, but if staff stay silent when AI produces harmful results, the organisation is left managing a system it cannot fully see.
You must be logged in to post a comment.