Text Anonymization in the AI Era
AI and search make it easier to find correlations in writing style, topics, experience timelines, expertise, past posts, images, files, and URLs.
When writing anonymously, there was a time when "not writing your real name," "not writing the company name," and "generalizing the region name" could seem somewhat sufficient.
That is no longer enough.
Search, text analysis, summarizing AI, similar-text search, and long-term social media histories make it easier to find writing habits, topic bias, timelines of personal experiences, specialized fields, and overlap with past posts.
In AI-era text anonymization, you need to check not only which words have been removed, but also "what kind of person the whole text creates."
This article organizes what clues can be seen from text in the AI era and how to prepare text before publishing.
The Basic Issue Is Still Correlation, Even in the AI Era
AI does not mean the way we think about anonymity has completely changed.
The basic issue is correlation.
The problem is that separate pieces of information connect to the same person, same workplace, same region, same past account, or same activity. However, AI and search technologies have made the work of finding that correlation faster.
| Clues in text | What they connect to | AI-era caution |
|---|---|---|
| Writing style | Past posts, other accounts | Similar writing is easier to find |
| Topics | Occupation, region, interests | A person profile is built from long-term history |
| Specialized terms | Affiliation, area of responsibility | The industry or organization can be narrowed down |
| Personal experiences | Timeline, people involved | The order of events can be checked |
| Named entities | Personal names, places, organizations | Search can surface candidates more easily |
AI-era text anonymization is not about fearing AI.
It is about checking on the assumption that correlations humans used to miss by manual review may be picked up mechanically.
Writing Habits Remain Even After Removing Real Names
Text has habits.
Words you often use, sentence length, how you connect ideas, how you write headings, how you choose examples, punctuation, emphasis, how you get angry, and the order of explanation. These remain even when the writer does not notice them.
| Habit | Example | Caution |
|---|---|---|
| Phrasing | Stock phrases you always use | Strong when it overlaps with the real-name side |
| Structure | Order of introduction, example, conclusion | Similar across the whole article |
| Punctuation | Comma placement and line breaks | Small habits accumulate |
| Example choice | Examples from the same industry or region | Experience comes through |
| Emotional expression | Anger, sarcasm, certainty | Connects to past posts |
Writing-style correlation is covered in detail in another article.
What matters here is that "replacing words alone does not change writing style." In anonymization, look not only at proper nouns, but also at text structure and example choice.
Having AI Rewrite It Does Not Make It Safe
Having AI rewrite text can change the writing style.
However, that alone does not make it safe. The input to the AI includes the original text, proper nouns, internal information, personal experiences, and information about people involved. If you enter it into an external service, you are trusting that service.
| Method | What changes | Remaining problem |
|---|---|---|
| Change sentence endings | Surface impression | Topics, timeline, and expertise remain |
| Have AI summarize it | Text length and expression | Input content is handed to an external service |
| Remove only proper nouns | Direct naming | Surrounding information narrows candidates |
| Translate and translate back | Part of the writing style | Meaning and example choice remain |
AI can be a useful aid.
However, you should think carefully before putting high-risk writing, whistleblowing, sources, unpublished materials, or personal victim information into an external AI service. The destination service, retention policy, account, and usage environment become new trusted parties.
Layers to Review in Text Anonymization
In text anonymization, review several layers in order.
Looking only at words, only at writing style, or only at the timeline is not enough.
| Layer | What to check | Example |
|---|---|---|
| Direct identifiers | Names, addresses, organization names | Real name, school name, company name |
| Quasi-identifiers | Region, occupation, age, role | Information that narrows to a small group |
| Writing style | Phrasing, structure, habits | Same writing as the real-name side |
| Content | Personal experiences, specialized field, interests | Overlap with past posts |
| Time | When events happened, posting time | Checked against real-world records |
| External elements | Images, URLs, files | Information outside the body text |
Reviewing in this order reduces missed checks.
After removing names, check occupation and region. Next, look at writing style and timeline. Finally, check images and files too. Text anonymization is a step-by-step process.
Where to Keep Specificity
What makes anonymization difficult is not making the text too thin.
If you write everything as "in some place, someone experienced some event," anonymity may improve. But the reader learns nothing.
What matters is separating specificity that readers need from specificity that moves closer to the person.
| Purpose | Specificity to keep | Specificity to remove |
|---|---|---|
| Warning | Failure pattern, review steps | Real organization names, dates |
| Consultation | Problem, needed support | School name, workplace name, names of people involved |
| Technical explanation | Mechanism, examples that cannot be reproduced | Internal URLs, real data |
| Sharing an experience | The problem felt, the structure | Detailed timeline, small-group roles |
Even in the AI era, good anonymization is not simple deletion.
It preserves meaning and lowers the precision used for matching.
Think About Trusted Parties Before Entering Text Into AI
In AI-era text anonymization, before "having AI fix it," think about what you are handing to that AI service.
If the entered text contains workplace names, names of people involved, internal circumstances, unpublished evidence, victim information, or source information, you are handing that information to an external service at that point.
| Information entered | What happens | What to check |
|---|---|---|
| Personal consultation text | Routine places and people involved may be included | Whether you can trust the destination service |
| Whistleblowing draft | Organizations and evidence may be included | Consider where to consult before entering it into external AI |
| Reporting notes | Sources and contact timing may be included | Whether it could involve information providers |
| Text planned for publication | Writing style and proper nouns may be included | Generalize at least the minimum before input |
Using an AI service itself is not bad.
However, for high-risk writing, first remove proper nouns and information about people involved locally before entering it into AI. If concern remains even after that, prioritize consulting a trusted person or specialist instead of using external AI.
Text Is Reused After Publication Too
Published text does not end at the moment of publication.
It is picked up by search, quoted, summarized, screenshotted, and redistributed elsewhere. Even if you delete it later, the original text may remain.
If you add replies and supplements after publication, new clues are added.
For text that requires anonymity, manage not only pre-publication review but also post-publication reactions. Emotional replies, additional personal experiences, rebuttals to people involved, and timeline supplements can become stronger clues than the first post.
Pre-Publication Review
In AI-era text anonymization, review in the following order.
- Remove direct identifiers such as names, organization names, and place names
- Change occupation, region, role, age, and years of experience to broader expressions
- Check whether the timeline of personal experiences is too detailed
- Look for the same writing style or topics as real-name accounts
- Check whether combining it with past posts could create a person profile
- Check images, files, URLs, and screenshots too
If any item remains uncertain, do not post as-is.
For anything you do not understand, move to one of these options: generalize it further, delay publication, do not publish it, or consult someone you trust.
Summary
In AI-era text anonymization, removing real names and company names is not enough.
Writing style, topics, expertise, personal experiences, timelines, past posts, images, files, and URLs can combine and point toward the person or people involved.
AI and search technologies make these correlations easier to find.
In anonymization, keep the meaning readers need while lowering the precision used for matching.
It is important to check not only the text, but also past information and the whole set of published materials.
Related tools
OSINT Framework
An external resource related to this article. Open it only when it fits your situation and threat model.
Why it is listed: It can help with the article topic, but it is outside Anonymity Sense and should be checked before use.