2026-11-20 - Why Removing Metadata Is No Longer Enough


You removed the location from the photo and the author’s name from the document. Can the recipient still trace them back to you?

Take a hypothetical case. A document with no metadata describes an accounting error, gives the month of the audit and mentions the purchasing office. In that company, two people carried out the audit. The file is clean, but anyone reading it already knows whom to look for.

Cleaning deals with a file’s hidden fields, not with what the file shows and says. Removing metadata is still useful, but text and images can reveal information even without metadata.

I have already explained how to use mat2 in a

. This post starts where that note ends: what remains after cleaning.

Why now

Two recent developments. On 7 July 2026, the

adopted draft Guidelines on anonymisation. They state that agentic AI is expected to cut the time and cost of re-identification (para. 92). On 22 October,

published a test on 52 unpublished photos, without metadata or context: the best models placed more than half of them within 25 kilometres.

This does not mean that everything can now be identified. It means that trying costs less: in one of the studies below, roughly 1 to 4 dollars per profile.

What files contain

FileWhat it can containWhy it matters

Photos

, any camera data and, if enabled, the

It can show where you were and when

PDF

It can show who produced the file and with what tools

Word, Excel, PowerPoint

(Microsoft's list for Word)It can show who worked on the text and what they changed

What AI can infer

With the metadata gone, three sources of clues remain:

.

In my

, the context was around the message. Here it is inside the message.

StudyWhat it measuredResultLimits

Style: three anonymous posts compared with about 100,000 blogsCorrect blog ranked first in about 20% of cases, in the top twenty in about 35%Assumes an author who does not disguise their style. English-language blogs

Context: eight personal attributes inferred from 520 Reddit profilesGPT-4: about 85% correct on the first guessLabels assigned by the authors, data not public

(preprint)Context: 338 Hacker News users, after removal of direct identifiers226 identified (67%) at 90% precision, at roughly 1 to 4 dollars per profileFew cases, users probably less cautious than average: the authors flag possible overestimation

Images: 20,000 Street View test photos, model trained for the task28.7% within 1 km, 59.2% within 25 kmStreet View over-represents North America and Europe. Indoor and low-context images excluded

Images: 52 unpublished photos, without metadata, context or major landmarksSeveral models: at least 10 photos within 1 km. The best: over half within 25 kmSmall sample. Even the best can confidently name the wrong place

These figures do not say that anyone can be recognised. Each test measures a specific population and task, and they should not be added together. Staab’s study infers attributes, not identities. Lermen’s uses content clues as well as style, so it does not measure pure stylometry. Bellingcat concludes that every result must be verified independently.

In my view, errors matter too: a wrong attribution can hit someone who has nothing to do with it.

What the law says

When data is truly anonymous

Under

, personal data is any information about an identified or identifiable person, including indirectly. Recital 26 places anonymous information outside the Regulation. Whether a person is identifiable depends on the means that the controller or a third party can reasonably use, taking into account costs, time and technology, including future developments.

One simple point follows: “without metadata” does not mean “anonymous”. The first phrase describes a file. The second is an assessment of what someone can reasonably find out.

Two judgments of the Court of Justice help here. In

(C-479/22 P, 7 March 2024), a press release did not name the scientist concerned, but the Court took into account what other people could know, such as colleagues and journalists. In

(C-413/23 P, 4 September 2025), it made clear that pseudonymised data is not automatically personal for everyone and that personal opinions are closely linked to those who express them. For this judgment, I refer to my

.

Whistleblowers, sources and redacted documents

In Italy,

, which implements

, prohibits disclosure, without express consent, to anyone not competent to handle the report. The ban covers the reporter’s identity and any other information from which it can be deduced, directly or indirectly. The rule protects more than the name.

It binds, however, those who receive and handle the report. In practice, it does not stop a colleague from recognising a detail known to few. If an anonymous reporter is later identified and suffers retaliation, Article 16(4) extends the protections. The

, the Italian anti-corruption authority, also provide for a ban on tracking reporting channels. For internal channels,

regards email as inherently unsuitable for ensuring confidentiality, unless specific countermeasures are in place.

For journalists, the

,

,

, subject to exceptions, from requiring disclosure of information that identifies or could identify sources. The limit is similar: it concerns States, not what a reader infers from a published document.

The same applies to redacted documents. Covering a name does not cover a role, dates and facts. The law takes this into account:

requires, among other things, that published court decisions omit data from which the identity of minors can be

derived. I will devote a future post to the technique of redaction.

What may change

What follows is not law in force.

Through the

, the Commission proposed a more subjective notion of personal data. The Council keeps the current definition and adds a provision on pseudonymisation and identification. The text still has to go through trilogues.

The

are also a draft, with a

. They propose three criteria: the data must not isolate a person, must not be linkable to other data about them, and must not allow specific inferences about them. If one is not met, further analysis is needed. The likelihood of identification must be insignificant in reality, not zero (para. 22), and it tends to rise over time, so anonymous data can become personal again (para. 35).

In my view, an unsigned text full of details would hardly pass this test when read by someone who knows that environment.

What you can do

1. If you need to report wrongdoing, start here

Before cleaning any file, choose the channel and secure the original.

, a project I am part of. Hush Line is text-only: you have no file to clean, but your style remains.

2. Check and clean the file

Always work on a copy. The table shows the paths given by the vendors, which may change with updates. Labels may differ with your interface language.

ToolHow to do itGood to know

File → Info → Check for Issues → Inspect Document → Inspect → Remove AllIt does not detect white text on white or objects covered by others. Removed data may not be recoverable

To share: Share → Options → turn off Location. To remove the data: open the photo → More → Adjust Info → Adjust Location → No LocationFor the future: Settings → Privacy & Security → Location Services → Camera → Never

Select the photos → Image → Location → Hide LocationThe guide does not explain what remains in the exported file: check it again

Tools → Show Location InfoIf the item is grey, the image contains no location. The guide covers reading, not removal

To read: exiftool -a -u -g1 photo.jpg. To remove: exiftool -all= photo.jpgIt keeps the original with the suffix _original: that is not the one to send. Not everything can be deleted. On PDFs, changes are reversible:

(qpdf --linearize in.pdf out.pdf)

To read: mat2 -s file. To remove: mat2 file, which creates file.cleaned.extension. Details in

It can alter the file: PDF text may no longer be selectable

After cleaning, check the result with a second tool. Even then, the

project warns, a clean check does not prove that the file has no metadata.

3. Reread the content as a stranger would

Guidelines, para. 24).

How much does it really help?

Each measure covers one part of the problem. Data on effectiveness are few and partial.

CountermeasureWhat it reducesWhat remains

Removing metadataAuthor, dates, location and revisions stored in the file's fieldsContent, style, background. No tool guarantees it has found everything

Reviewing text and photosThe most obvious cluesThe studies cited do not measure its effectiveness

Manually rewriting styleIn a

with ten candidate authors, accuracy fell from about 40% to about 20%Small groups and a 2022 study, predating current models. Context remains

Model-guided rewritingIn a

, attribution among many authors fell from about 16% to about 1%Against an already suspected author, the effect is smaller. English texts only. With an online service, you hand your text to a third party

Automatic text anonymisationIn

, location inference fell from about 86% to about 55%The overall drop is modest

Dedicated channelThe risks of email and attachmentsStyle and content of the message

Legal protectionsWhat those who receive the data may doThey do not stop inference from public material, and they often come afterwards

Removing metadata is the first step, not the last. The second is deciding what to write and what to show, and no tool can do that for you.

Sources

, version 1.0 for consultation, 7 July 2026;

;

, Article 4 and Recital 26

, 7 March 2024;

, 4 September 2025

, Articles 12 and 16 (in Italian);

, resolution no. 311/2023, and

(in Italian)

;

;

, Article 52 (in Italian); Garante,

, 2 December 2010 (in Italian)

;

;

;

;

;

;

, 22 October 2026

,

,

,

,

,


New to Gemini? This capsule is best read with a Gemini client, such as Lagrange.