When feedback reveals more than the writer chose to share
Open-text comments can reveal sexual orientation, religion, caste, or home language. Here's what that means for HR teams who read and share feedback.
AI HR Skills Editorial TeamOpen-text feedback can reveal a writer’s sexual orientation, gender history, religion, caste, or migration status even when it never reveals their name. A partner’s name, a mention of a transition, a prayer time, a surname, or a visa status can tell readers something the writer never meant to share at work.
HR teams usually treat open-text privacy as a question of authorship. But a comment can be “anonymous” and still out someone, or reveal a faith they keep private at work. Here’s which details carry that risk, how to read comments without guessing, and what to check before feedback gets shared.
Key takeaways
- In a survey of LGBTQ US workers (data collected in 2023), 46% weren’t out to their current supervisor, so a detail in a comment can disclose something a manager doesn’t know.
- In 2022 the EU’s top court ruled that data “liable indirectly to reveal” sensitive traits, such as sexual orientation, falls under GDPR’s special categories.
- A study across 33 countries found that differences in why employees stay silent didn’t neatly match cultural stereotypes, so read severity from the facts, not from tone or presumed background.
- Before sharing comments, check for identity details, keep the substance of any allegation, and never label a writer based on what they wrote.
Which details reveal identity
Some details point to a person. Others reveal something about them. In open-text feedback, these come up most often:
- Partners and family. “My wife and I” from a woman, or a partner’s name.
- Gender history. “Since I transitioned,” a former name, or pronouns at issue.
- Religious practice. Prayer times, fasting periods, religious holidays, dress, or diet.
- Caste or community. Surnames, community references, or “people like me.”
- Migration status. “The only one here on a work permit.”
- Language. Writing in, or mixing in, a home language few colleagues share.
Often these details are the point of the comment. “I was treated differently because of my religion” is the feedback. So the question is less whether the detail is there and more who ends up reading it.
Why the stakes are uneven
Not everyone at work has shared these things. In a Williams Institute survey of 1,902 LGBTQ US workers, 46% weren’t out to their current supervisor and 21% weren’t out to any coworker. Those who were out to at least a few coworkers or their supervisor were three times as likely as those out to no one at work to report experiencing discrimination at some point in their lives (39% vs 12%).
Reporting is already fragile. In a 2018 Human Rights Campaign survey, the top reason LGBTQ workers gave for not reporting negative comments about LGBTQ people to a supervisor or HR was that they didn’t think anything would be done. A comment that gets shared carelessly confirms that fear.
For a global workforce, the risk can go beyond the office. ILGA World data reported by GCN shows the number of countries that criminalize consensual same-sex activity rose to 65 in 2025, the first increase in almost a decade. A comment that’s harmless to share in one office may not be in another.
How regulators look at indirect clues
Data protection law already treats indirect details as revealing. GDPR Article 9 lists racial or ethnic origin, religious or philosophical beliefs, and sex life or sexual orientation among its special categories of personal data. In a 2022 ruling summarized by Covington, the EU’s top court said data “liable indirectly to reveal” sensitive information falls under that article, and noted that a partner’s name can reveal sexual orientation.
Protections vary by place and characteristic. Caste is one example. JURIST reported that California’s governor vetoed a 2023 bill adding caste to state anti-discrimination law, saying existing categories already cover it. Seattle and Fresno had passed city-level bans that same year.
This article isn’t legal advice. If a comment raises a legal question, talk to your own counsel.
Read what’s written, not who you think wrote it
Open-text comments invite guessing. Three habits keep that guessing from turning into harm.
Don’t infer identity. A comment that mentions a same-sex partner doesn’t tell you how the writer identifies or who knows. Summaries and reports should describe what was said, not label the person who said it.
Don’t read severity from tone or background. Some people understate everything. The tempting shortcut is to adjust for culture, but a study of 8,222 employees in 33 countries found that differences in why people stay silent “did not necessarily support cultural stereotypes”.
“The brake is maybe a little problem, it’s been three weeks” still describes a safety issue. Read the facts.
Don’t treat non-native English as suspect. Second-language writing can itself point to a person. Researchers have built systems that can often identify a writer’s first language from their English. Stanford researchers also found that several widely used AI-text detectors consistently misclassified non-native English writing as AI-generated.
If your team uses AI tools to screen or summarize comments, check that they don’t single out people writing in a second language.
A hypothetical example
Imagine a 70-person company running its annual engagement survey. One comment reads: “The new leave policy doesn’t account for Eid. I’ve had to use vacation days two years running, and I’m the only one on my team who has to.”
An HR partner copies it word for word into a summary for the team’s manager. The comment didn’t name anyone, but the manager knows exactly one person on the team observes Eid.
The writer’s faith, and their complaint, now sit with the person they report to. The writer didn’t choose that.
A better route keeps the substance and changes the exposure. “The leave policy doesn’t account for some religious holidays, so some employees use vacation days” carries the same issue to the policy owner, while the full comment stays with HR.
What HR teams can do before sharing comments
The goal isn’t to hide feedback. Decide who needs which version.
- Review comments before they leave HR. Look for partners, gender history, faith practice, caste or community, visa status, and home language.
- Keep the substance of any allegation. If someone reports discrimination, the ground of it (“because of my religion”) is the point. Generalize the detail, not the complaint.
- Route by need. A policy owner usually needs the issue and a manager the theme. Anyone investigating a complaint needs the facts.
- Tell employees how comments are handled before they write, so they can decide what to include.
- Check your AI tools. Anything that summarizes or screens comments should describe content without inferring the writer’s identity.
Where a tool fits
A writer-side tool can help people see these risks before they hit submit. When you evaluate one, these capabilities matter more than polish:
- It flags details that could reveal something about the writer.
- The writer stays in control. They decide what to keep, generalize, or remove, especially in a report of discrimination.
- It doesn’t label anyone. Flags say “this could reveal,” never “you are.”
- It’s honest about limits. “Harder to trace” isn’t the same as anonymous.
The Pulse Feedback Survey Anonymizer is a free Chrome extension by FeedbackPulse, which also makes the AI HR Skills library. It rewrites survey answers and feedback in a neutral style and flags details that could still point to the writer, with options to keep, generalize, or remove each one. It lowers the chance of being identified but can’t guarantee anonymity, especially when the content itself is unique to the writer. Text is sent to the FeedbackPulse server and rewritten by Anthropic’s Claude, and the server doesn’t store or log it. The free version covers 10 rewrites a week, up to 500 words each. See the Pulse Feedback Survey Anonymizer listing.
For designing the survey itself, see the FeedbackPulse eNPS Survey Builder, which helps design an eNPS or pulse survey in ChatGPT. If you’ve built an identity-aware feedback tool, send us a skill for review, and see what a skill has to show before it is listed.
FAQ
Can an anonymous survey comment still out someone?
Yes. A comment can protect who wrote it and still reveal something about them, such as a same-sex partner or a religious practice, to readers who didn’t know.
Should HR remove identity details from comments?
Not automatically. Often the detail is the substance of the feedback. Keep the full comment within HR, and share a generalized version with people who need only the issue.
Is it safe to use AI to summarize employee comments?
It depends on the tool. Check that summaries describe content without labeling writers, and that the tool doesn’t treat non-native English as suspect.
What to do next
Open-text feedback carries more than opinions. It can carry a person’s relationships, faith, background, and language, and the writer may have shared those with very few people at work.
Start with your next survey. Before any comment leaves HR, ask who will read it and whether a detail in it could expose the writer.
For more on running feedback well, see other guides in the resources section.
Sources
- Source: LGBTQ employees who are out at work are three times more likely to have experienced discrimination, Williams Institute (2024)
- Source: Startling Data Reveals Half of LGBTQ Employees in the U.S. Remain Closeted at Work, Human Rights Campaign (2018)
- Source: Number of countries criminalising same-sex acts grew for first time in almost a decade in 2025, GCN (2026)
- Source: Art. 9 GDPR, Processing of special categories of personal data
- Source: Special Category Data by Inference: CJEU significantly expands the scope of Article 9 GDPR, Covington Inside Privacy (2022)
- Source: California governor vetoes anti-caste discrimination bill, JURIST (2023)
- Source: International differences in employee silence motives across 33 countries, Knoll et al., Journal of Organizational Behavior (2021)
- Source: A Report on the 2017 Native Language Identification Shared Task, Malmasi et al., ACL Anthology
- Source: GPT detectors are biased against non-native English writers, Liang et al., arXiv (2023)