← Back to resources
Engagement & Feedback

What an AI rewrite can and can't hide in survey feedback

An AI rewrite can mask how a survey comment sounds. It can't change what the comment says, how the platform stores it, or who reads it at small team sizes.

An AI rewrite can change how a survey comment sounds. It can’t change what the comment says, how the survey platform stores it, or how many people end up reading it. That makes a rewrite useful, but it never makes a comment anonymous on its own.

HR teams often see employees hold back in open-text questions. Some employees now paste a comment into an AI assistant first and ask it to “make this anonymous.” Here’s what that rewrite protects, what it leaves exposed, and what to look for in a tool built for the job.

Key takeaways

  • Rewriting with a large language model weakens the writing-style clues that authorship tools rely on, but results vary a lot from one writer to the next.
  • Identifying content survives a rewrite. Researchers found LLMs could infer traits like location, income, and sex from real Reddit posts with up to 85% top-1 accuracy.
  • Comment thresholds, platform data links, and metadata sit outside the text, so no rewrite reaches them.
  • A useful rewrite tool flags risky details for the writer to change and is honest that the result is harder to trace, not anonymous.

What a rewrite can hide: writing style

Everyone writes with habits: pet phrases, punctuation, sentence length. Authorship tools use those habits to match unsigned text to a known writer. That’s the layer a rewrite disturbs.

Researchers have tested this directly. One study found that a single round of LLM paraphrasing sharply lowered the accuracy of authorship classifiers, with smaller drops in later rounds, and the drop tracked how far the text moved from the writer’s original style. Another found that, in its strongest setting, LLM-based rewriting could defeat an authorship verification model up to 92% of the time.

Two cautions apply:

  • Results vary by writer. A 2025 study found LLM obfuscation worked very differently across individual authors. Some writers became hard to identify, while others stayed easy to spot. A good average can hide the one person it didn’t protect.
  • None of this was tested on survey comments. These studies use texts such as news articles and Reddit posts, not 40-word comments read by a manager who knows the team. Whether a rewrite holds up in that setting is an open question.

What a rewrite can’t hide by itself: what the comment says

A rewrite is designed to keep the meaning. That’s the point of it. So every fact in the comment survives: the project, the meeting, the shift pattern, the “since I moved over from the Denver office.”

Content can reveal more than people expect. In one study, large language models inferred personal attributes such as location, income, and sex from real Reddit posts with up to 85% top-1 accuracy. The same study found that common text anonymization didn’t stop it.

In a workplace, the reader doesn’t need a model. A manager who knows who handles payroll, who joined in March, and who was in Tuesday’s meeting can often place a comment from one detail. A rewrite that keeps every detail and only changes the tone leaves that clue in place.

The step that helps is flagging. A tool that points out identifying details lets the writer decide whether to keep, generalize, or remove them. “Someone in a small regional team” protects more than “the only analyst in Denver,” and it still says something useful.

What no rewrite can change: the survey itself

Some of the biggest risks have nothing to do with wording.

Comment thresholds. Survey platforms hold back results for small groups, but comments often follow different rules from scores. Microsoft Viva Glint’s standard settings show scores at five respondents and comments at 10. It counts respondents, not comments, so a team with 10 responses and one comment shows that one comment.

Qualtrics controls open text through a separate minimum-responses setting and says its confidentiality features can’t guarantee protection in every situation.

Confidential isn’t anonymous. Most engagement surveys are confidential, which means the platform can link a response to a person. Glint, for example, connects scores to individuals in a restricted raw data file that admins can view. Two engagement consultants told SHRM in 2018 that most surveys are “confidential” but not “anonymous.”

Metadata. A survey vendor’s guide to anonymous surveys says true anonymity means suppressing IP addresses, cookies, device identifiers, and link parameters. Those are survey settings. A rewrite can’t touch them.

A rewrite can help

Writing style

  • Pet phrases, punctuation, and slang
  • Sentence length and tone
  • Results vary by writer

Helps only if the writer acts on a flag

Identifying content

  • Roles held by one or two people
  • Dates, places, and incidents only a few people saw
  • Tenure, life events, and specific numbers

Out of reach for any rewrite

The survey and the reader

  • Comment thresholds and how small groups roll up
  • Platform links between responses and people
  • IP addresses, timestamps, and other metadata
  • A reader who already knows who was there

A small-team example

Picture a hypothetical eight-person support team in a 60-person company. One person writes a careful comment about a schedule change, then has an AI assistant rewrite it into flat, neutral language. The style clues are gone.

But the comment still mentions “covering the weekend queue since the reorg,” and only two people do that.

The team doesn’t meet the comment threshold, so the comment rolls up to the department. In a company this size, the department head knows exactly who covers weekends. The rewrite did its job, and the comment is still easy to place. Changing that one phrase to “covering extra shifts” would have done more than the whole rewrite.

Where a rewrite tool fits

A rewrite tool earns its place when it’s honest about all three layers. When you evaluate one, or decide whether to suggest one to employees, look for these:

  • It rebuilds the voice rather than lightly editing the original.
  • It flags identifying details and leaves the writer in control of every change.
  • It says plainly that it can’t affect survey thresholds, platform data, or metadata.
  • It promises “harder to trace,” not “anonymous.”

The Pulse Feedback Survey Anonymizer, a Chrome extension by FeedbackPulse, is built around that job. It rewrites 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. The free version allows 10 rewrites a week, up to 500 words each. Text is sent to the FeedbackPulse server and rewritten by Anthropic’s Claude, and the server doesn’t store or log it. See the Pulse Feedback Survey Anonymizer listing.

If your question is about designing the survey itself, see the FeedbackPulse eNPS Survey Builder, which helps design an eNPS or pulse survey in ChatGPT. To see what any tool has to show before we list it, read how listings are reviewed.

FAQ

Does rewriting a survey comment with AI make it anonymous?

No. It can weaken writing-style clues, but the content, the survey platform’s records, and the reporting thresholds stay the same.

Can a manager still tell who wrote a rewritten comment?

Sometimes, yes. If the comment mentions a detail only a few people share, or it reaches a reader who knows the team well, the wording doesn’t matter much.

What is a comment threshold?

It’s the minimum number of respondents a group needs before its comments are shown. In Viva Glint’s standard settings, that’s 10 respondents, compared with five for scores.

Should HR suggest rewrite tools to employees?

Only with an honest description. Tell people what a rewrite changes, what it doesn’t, and how your own survey handles comments and thresholds.

What to do next

A rewrite is one layer of protection, not the whole job. If you run surveys, check your own comment threshold and tell employees what it is. Say whether your survey is confidential or anonymous, and use the words accurately. If you suggest a rewrite tool, choose one that flags identifying details and says “harder to trace” rather than “anonymous.”

For more tools and guides on running feedback well, browse the AI HR Skills library and the resources section.

Sources