Privacy and preparation
Dataset metadata privacy review: what files say that bodies do not
By SourceX Editorial · Reviewed by Noah Loul ·
9 sections
Short answer
A dataset metadata privacy review checks what a body-text scan never reads: author and editor fields, comment authors, company names, local file paths and timestamps. In 563 readable Office files from Apache POI's public test corpus, 435 carried a person-like name in metadata, and 424 of those held at least one name the body never mentions. Treat metadata as its own layer, keep only fields the task needs and record the scan.
Redaction
Dataset metadata privacy
Key takeaways
- Body-text redaction cannot see author, editor, comment-author or company fields, and in the scanned corpus those fields held names in most files.
- Of 435 public test files with a person-like name in metadata, 424 (97.5%) carried at least one name that never appears in the body text.
- Timestamps were nearly universal, with created dates in 94.7% of readable files, and often carry task value, so transform them rather than strip them by reflex.
- Local paths naming a user folder appeared in 51 files (9.1%), mostly spreadsheets, where no review of the visible text would find them.
- Decide every metadata field by necessity: keep what the task reads, transform what it needs in another form, strip the rest, and scan the final files.
What should a dataset metadata privacy review decide?
It decides, field by field, which metadata travels with a dataset, in what form, and how you will show that nothing else does. The scope is every layer around the content: file and folder names, document properties such as author and last editor, comments and tracked changes, embedded images, and the audit columns that system exports attach to each record.
Those layers are written automatically, which is why reviews miss them. NIST's 2022 scientific foundation review of digital investigation techniques explains that metadata such as the full file name, the owning account, access permissions and modify, access and create times comes into being together with a file (paraphrased). Nobody types it, so nobody remembers to read it.
A sound answer rests on a scan of the final files, a list of the fields the buyer's task actually reads, and a recorded decision for every other field. This page gives a measurement of how often metadata names people the body never mentions, a privacy-utility review matrix and a final-delivery checklist. It is general information, not legal advice, and it names the jurisdiction of each piece of guidance it cites.
Why can a file's metadata say more about people than its body?
Because the body records the work and the metadata records who did it. A pricing workbook's cells hold products and numbers; its properties hold whoever created it and whoever saved it last, sometimes a manager and a company, and now and then a folder path containing a username. None of that has to appear on the page to travel with the file.
Regulators and investigators have both noticed. The ICO's freedom of information guidance warns that metadata "might include the name of the author, or details of earlier draft versions" (UK). The NIST review recounts an investigation in which a document's metadata yielded a first name and an organization, which investigators treated as a useful lead even though, the report notes, it established nothing definitive on its own (paraphrased).
For a dataset prepared for licensing the consequence is practical. A review that scans only the text a reader sees can approve a file that still names several people, and the measurement below suggests that is the normal case rather than the edge case.
How often does metadata name someone the body never mentions? A scan of 563 public files
Often, in the one public corpus measured here. Apache POI, an open-source Java library for Office formats, publishes the files its automated tests read. The scan covered the 607 Word, Excel and PowerPoint packages in its document, spreadsheet and slideshow test folders, of which 563 open as valid packages.
The script compared two things without printing either. On one side were person-like values in metadata: creator, last modified by and manager properties, comment and tracked-change authors, and the people parts newer Office versions add. On the other was the body text a reader sees: Word text, headers, footers, footnotes and endnotes; Excel cell text and values; PowerPoint slide text. Values that looked like tool defaults, such as anything containing Microsoft, Apache, user or author as a word, were set aside.
435 of the 563 files (77.3%) carried at least one person-like value. In 424 of them (97.5%) at least one value never appears in the body, and in 414 (95.2%) none does; only 11 files carried names that all also appear in the body, and 128 carried none. Empty files do not explain the gap: of the 378 files with names and some body text, 367 (97.1%) still held a name the body never mentions. By format:
| Format | Readable files | Person-like name in metadata | Name absent from body | Username in a local path |
|---|---|---|---|---|
| Word | 122 | 96 | 94 | 2 |
| Excel | 351 | 262 | 261 | 44 |
| PowerPoint | 90 | 77 | 69 | 5 |
| All formats | 563 | 435 (77.3%) | 424 (75.3% of readable) | 51 (9.1%) |
Which metadata does the task need, and in what form?
Usually far less than the files carry, and rarely in its original form. Start from the buyer's task: ordering and duration often matter, roles sometimes matter, and names, companies and paths almost never do. The corpus shows how much there is to decide: created dates in 533 files (94.7%), modified dates in 531 (94.3%), a nonzero editing-time counter in 154 (27.4%), a company name in 155 (27.5%), a printer settings part in 115 (20.4%), custom properties in 42 (7.5%) and comment author entries in 38 (6.7%). JPEG images with EXIF data were rare, embedded in 8 files (1.4%), and none carried a GPS tag.
The matrix records, for each field, when the task needs it, the form that keeps that value without the exposure, and the proof a reviewer opens on the final files. Defaults assume one recipient under contract:
| Metadata field | Needed by the task when | Form that keeps the value | Exposure if shipped as is | Proof on the final files |
|---|---|---|---|---|
| Author and last-editor names | Authorship roles matter, as in review or approval workflows | A role token from the same scheme as the body, such as Reviewer_07 | Staff names the body never mentions | Property read-back shows tokens or nothing |
| Comment and tracked-change authors | Revision history is the product, as in contract drafting | A separate, tokenized revision record; comments resolved in the delivered file | Reviewer names beside internal opinions and deleted wording | Author count is zero, or every author is a token |
| Company, manager and custom properties | Almost never | None: strip | Client names, matter numbers and internal codes | Property parts empty on every file |
| Local, template and linked-file paths | Never | None: strip and break external links | Usernames, server names and client folder names | A path search across every XML and relationship part returns nothing |
| Created, modified and editing-time values | Sequence or duration is the signal | Offsets within a case, a per-case shift or coarser periods | When a named person worked, which rosters or calendars can match | Order holds on sampled cases and removed clock times are absent |
| File and folder names | Rarely; the document type can matter | A neutral identifier plus a document-type field | Clients, people and dates written into names | The name list passes a pattern review before delivery |
| Audit columns in system exports | Assignee and status changes are the workflow signal | Tokens under the dataset's pseudonym scope, with IP addresses dropped | Requester and agent identities and network locations | Column inventory signed against the task's field list |
Should timestamps be stripped or kept?
Usually kept, in a different form. Workflow datasets earn their value from order and duration, such as how long a ticket waited or which revision followed which, so deleting every time field can destroy the reason the records were worth preparing.
The risk sits in exact clock times tied to a person. For U.S. health data, the HIPAA Safe Harbor list in 45 CFR 164.514(b)(2) includes elements of dates other than the year that are directly related to an individual, such as admission and discharge dates (paraphrased), which is why health datasets often keep only years or shifted dates. Outside health data no rule names timestamps so plainly, but the logic carries over: an exact time can be matched against a roster, a calendar or a badge log.
Three transformations keep most of the value. Offsets from the start of each case preserve order and duration; a per-case date shift keeps weekday patterns when the shift is a whole number of weeks; and coarsening to day, week or month suits tasks that need only rough timing. Document properties are a special case: created and modified values describe the file rather than the work and rarely belong in a training set at all.
If the body is clean, isn't the file clean? The strongest objection
No, and the measurement shows why: in this corpus a flawless body-text redaction would have left at least one name behind in 424 files. Cleaning the body and moving files over an encrypted channel answer the questions they were built for. They do not touch the layers around the content, and neither creates a right to license the records.
UK guidance now treats the check as routine. In July 2025 the ICO announced guidance on disclosing documents that covers finding hidden personal information, including hidden rows, columns and worksheets, metadata and active filters, and that recommends software checks such as Microsoft's Document Inspector. The principle carries over to licensing: check every layer, then check again after the last edit.
The main exception is a dataset rebuilt from structured exports, where records are written fresh from database fields. Document properties disappear there, but the audit columns in the matrix take their place, so the review still needs a metadata step.
Final-delivery checklist for metadata
Run it on the files that will actually be delivered, after redaction, conversion and every other edit, and keep the output with the dataset's preparation record:
At a glance
Final-delivery checklist for metadata
Checklist · 10 items
- Inventory: file types and counts, including files that fail to open, which need a decision rather than a skip.
- Field list: the metadata fields the task reads, each tied to a matrix row and approved by the data owner.
- Property scan: creator, last editor, manager, company and custom properties read back on every file, not a sample.
- Author scan: comment, tracked-change and people parts counted, with any remaining author a token from the dataset's scheme.
- Body comparison: every surviving metadata value compared with the body text, so a name left in a property is a decision, not an accident.
- Path scan: drive, network and file-URL paths searched across every XML and relationship part, including template and external-link targets.
- Time fields: offsets or coarsening applied as the field list says, with order checked on a sample of cases.
- Embedded media: EXIF read back on every embedded image after stripping.
- File names: renamed to neutral identifiers where names carry clients, people or dates.
- Sign-off: reviewer, date, scan script version and open exceptions.
0 of 10 ticked. Tick each item as you go.
Metadata decisions in a SourceX deal
They belong to scoping and preparation. SourceX's trust page puts the supplier in charge of the boundary, choosing the systems, records and date ranges that go in, and it documents source, rights, preparation and permitted use with each dataset. After the go-ahead, the data is "extracted, transferred and de-identified through approved processing routes."
The legal framework page says personal details such as names, emails, phone numbers and account numbers are removed or replaced, that the method is recorded for each dataset, and that no method is perfect. A supplier that brings its metadata field list and scan output to scoping gives that record a concrete starting point instead of a promise.
Before any of that, a short assessment that asks for no data upload shows whether a company's records are a fit at all; an assessment is not an offer.
Methods note: corpus, rules and limits
Corpus: Apache POI test-data at commit 07f1631 (github.com/apache/poi), read on 2026-10-28: the 607 files directly inside test-data/document, test-data/spreadsheet and test-data/slideshow whose extension, in any case, is docx, docm, dotx, xlsx, xlsm, xltx, pptx, pptm, potx, ppsx or ppsm. 44 were excluded: 37 begin with a zip signature but do not open as zip archives, 6 are OLE2 containers, and 1 has no content types part. xlsb files are binary and out of scope. The script read parts as text with regular expressions and saved only counts and per-file flags keyed by git blob hash; it never printed or stored a metadata value or a file name.
Rules: person-like values come from core creator and last-modified-by, the app Manager property, Word comment authors, Word tracked-change authors (the author attribute on ins, del, moveFrom, moveTo, cellIns, cellDel, cellMerge and the property-change elements rPrChange, pPrChange, sectPrChange, tblPrChange, tblGridChange, trPrChange, tcPrChange and numberingChange, in document, header, footer, footnote and endnote parts), Word people parts, Excel comment authors and persons parts, and the PowerPoint comment-author and authors parts. Values are unescaped, trimmed and casefolded, and empty values are dropped. A value containing apache, poi, microsoft, office, openoffice, libreoffice, aspose, user, author, admin, administrator, unknown, owner, default or test as a whole word counts as a tool default; 97 value occurrences did. The company count applies the same exclusion to the app Company property. A value is absent from the body when it is not a substring of the body text joined either with or without spaces between text runs. A username path is a drive, network or file-URL path, found in any XML or relationship part, that contains a users, documents and settings or home folder followed by another segment. Layer counts: created and modified are non-empty core properties; editing time is the app TotalTime value read as an integer above zero; a printer settings part is any part whose name contains printerSettings; custom properties means at least one property in the custom properties part; comment author entries means at least one author value, even an empty one, in the comment parts above; EXIF means a JPEG in a word, xl or ppt media folder whose APP1 segment begins with the Exif marker, and GPS means that image's first directory holds tag 0x8825. Each share is rounded a single time to one decimal place, with halves rounded up.
Limits: this is a software test corpus, not a business archive, so the shares show what the test finds rather than how common each layer is in any company's files. Substring matching misses names written differently in the body, such as initials, and can count a short value as present by chance; the default list may keep some tool names and drop a real name containing a listed word. PDFs, standalone images, email and system exports were not scanned. The script, ooxml_metadata_vs_body.py, and its aggregate output are kept for recount.
Frequently asked questions
Which metadata should an AI training dataset keep?
Only the fields the buyer's task reads, in the form it needs them. Sequence and duration usually matter, so time fields are often kept as offsets within a case. Author, company and path fields rarely matter and are usually stripped or replaced with role tokens.
Are file names part of a metadata review?
Yes. File and folder names often carry client names, people and dates, and they travel with every copy. Rename files to neutral identifiers during preparation and keep any mapping with the supplier, not in the dataset.
Do CSV exports from a help desk or CRM carry metadata too?
Not document properties, but they carry audit columns such as requester and assignee IDs, created and updated times and sometimes IP addresses. Review those columns against the task's field list in the same way, and give person IDs the same pseudonym treatment as names in the body.
Should the metadata scan cover a sample or every file?
Every file, because the scan is cheap and failures are file-specific: one workbook with a username in a path or a forgotten comment is enough to expose a name. Use human review on a sample to confirm the scan's rules catch what they should.
Who should sign off on the metadata review?
The data owner approves the field list, a privacy reviewer checks the scan output and exceptions, and counsel reviews anything that touches sensitive records or contract limits. Record the names and the date alongside the scan output.
Sources
- The ICO's July 2025 guidance on disclosing documents covers finding hidden personal information such as hidden rows, columns and worksheets, metadata and active filters, and recommends software checks such as Microsoft Document Inspector. Source
- The ICO's FOI guidance notes that metadata might include the name of the author, or details of earlier draft versions. Source
- NIST IR 8354 (2022) explains that metadata such as full file name, owner account, access permissions and modify, access and create times is created with a file, and recounts a case in which document metadata gave a first name and an organization used as a lead (paraphrased). Source
- The HIPAA Safe Harbor identifier list at 45 CFR 164.514(b)(2) includes elements of dates, except year, directly related to an individual, such as admission and discharge dates (paraphrased). Source
- Apache POI test-data at commit 07f1631 holds the 607 Office packages scanned for metadata and body text. Source
- SourceX's trust page says the supplier chooses the systems, records and date ranges in scope, documents source, rights, preparation and permitted use with each dataset, and extracts, transfers and de-identifies data through approved processing routes after the go-ahead. Source
- SourceX's legal framework page says personal details such as names, emails, phone numbers and account numbers are removed or replaced, the method is recorded for each dataset, and no method is perfect. Source
See if your company qualifies
A short company assessment. No data uploads are needed.