Can you remove personal information from a document before you give it to AI?
Shadow AI in organisations is a risky topic but unfortunately, it still happens. We focus this blog on how to limit that risk when you need to think about using external AI tools with your organisation data. Disclaimer: we do not promote the use of sensitive or private data with external tools that are not approved by your organisation. This blog identifies where and how data can be de-identified.
The question
You want help turning a privacy plan into training actions.
Can you remove the personal details before AI sees the document?
Yes: start with permission and a smaller local copy.
The problem
A name is easy to spot. Usually, you would do a Find & Replace to remove names, brands, product names, and sensitive data in a document. But there are other ways to identify an individual or an organisation.
A distinctive role, a date and a progress note can also tell a story about someone.
Staff need an approved route and practical training, with a clear owner for the decision to share.
How can you ensure and reduce risk around exposing organisation documents to external AI tools.
What's real
OAIC treats identifiability as a question of information and context. Removing direct identifiers may be insufficient when other details or available data identify someone. [1]
The OAIC/CSIRO framework connects intended use, recipients and their environment with choices about the data. Pair less information with appropriate access controls, an accountable reviewer and continuing review as circumstances change. [2, 3]
Redaction removes selected content; de-identification asks whether people remain reasonably identifiable; permission determines whether this particular use is authorised.
Practical use
Set the boundary. Confirm purpose, recipient, service, account and legal/contractual authority. If prohibited, stop; if unclear, refer it to the document owner or privacy/IT lead. [4]
Take only what is needed. Keep required originals under retention rules. For a training outline, prepare generic actions in a separate local text file, leaving responsibility and progress details behind. Keep originals out of AI, cloud redactors and online converters. [3, 4]
Check the saved result. Reopen it; review the complete text, filename and relevant hidden information. Compare it with the local control copy to confirm useful meaning remains. [5]
Own the destination decision. Before optional AI use, assess what recipients could combine with other data. Confirm access, retention, input-training and connected-service controls. Record the decision and revisit it when the purpose, recipient or environment changes. [2–4]
Test it and report back
Try the local exercise this week. Did useful content survive? What needed another look? Tell us about the process in the comments, without sharing files, screenshots or identifying details. LinkedIn post: https://www.linkedin.com/feed/update/urn:li:activity:7518117445321166849
Get the next practical exercise every fortnight. Want help with staff training or your approved AI-use process? Book a 20-minute catch-up, or email hello@25hours.ai.
Exercise block
(prepare locally, without AI)
What you need: the actual free public OAIC Privacy management plan template PDF, an approved local PDF viewer and a local plain-text editor. Keep cloud sync and AI assistance out of this exercise.
The source: this is a blank planning template, not a populated personal record. The official landing page is dated 16/05/2016. The fictional values below are additions for practice, not information found in the original.
Check permission and rights.
Confirm this local practice is allowed. Read the OAIC copyright conditions and any notice on the source before adapting text. OAIC’s general CC BY 4.0 International grant excludes branding, images and third-party material. Use licensed text only; omit excluded material. Public access does not supply privacy permission or permission for every reuse.Keep the actual source.
Download the PDF to an approved local location outside cloud sync and open it locally. Preserve it unchanged.
Locate Step 1, Embed and Step 2, Establish, and the columns Position responsible, Due and Status. Use the template’s training-related material for the task: a generic privacy-training outline.Make a fictional control copy.
In a new plain-text file, create a short practice extract from that training material, paraphrased in your own words. This is a text extract from the real template, not an edited PDF.
Add Position responsible: TRAINING ROLE ALPHA; Due: TRAINING DATE ONLY; Status: TRAINING STATUS ONLY. Head it FICTIONAL PRACTICE, NOT A REAL RECORD. Use no realistic names, contacts, identifiers or credentials.Keep the credit.
Include the original source link, licence link and this change note: “Training text paraphrased; fictional practice fields added; source layout and branding omitted.” Include the attribution below in the control copy and any prepared extract that reuses licensed text.Prepare the minimum.
Save a separate plain-text output containing only the generic training action, the fictional-practice heading and the source credit/change note. Omit the responsibility, due and status fields and their invented values. Update the change note to record that these practice fields were omitted. Keep the source and control copy locally under applicable retention rules. Do not upload either for cleaning.Reopen and inspect.
Close and reopen the saved output in the local editor. Read all of it, including beyond the visible window. Search for each invented value and check the complete selected text. Inspect the filename and available file properties for identifying details. Compare the action with the control copy: preserve its useful meaning without adding facts. Record what you checked and what the local method could not check.Finish locally.
The exercise ends here. A person owns any later recipient, purpose, necessity, authority and channel decision. Before any optional AI input, confirm the particular account’s permissions, access, retention and input-training controls, and review the prepared content in that destination’s context. Repeat the review when circumstances change.
AI tiers: no AI account or prompt is needed for this exercise. Free, personal, licensed and paid accounts all require the same permission-first decision. Agents are outside this exercise.
Why it works
A controlled fictional extract makes the information choices visible.
Starting a separate minimal output keeps the training task distinct from responsibility and progress details.
Reopening checks the saved file, not just the editing screen.
If it doesn’t work
Stop if policy forbids the task or service; refer uncertainty to the owner or privacy/IT lead. Sanitisation cannot override DLP or contracts, and paid access, enterprise protection or a no-training setting is not approval. Keep real personal, client, worker, student, health, financial and confidential/IP files outside this exercise. Do not identify people, profile them or assess employees.
Use local, format-appropriate help if text, properties or saved content cannot be checked. This route does not edit the PDF or preserve its layout or form fields. Other formats need their own checks for comments, revisions, images, layers, attachments and embedded material. Empty searches, black rectangles and token replacement do not establish anonymity or compliance; public-source blankness does not establish forensic cleanliness.
Data-handling note
Before you try this: free and personal AI tools may use what you paste to improve their models, depending on your settings.
Don’t paste personal, health, student, resident or client information.
Use a de-identified sample, or your organisation’s licensed tool if it has one.
Sources
OAIC, De-identification and the Privacy Act, 21/03/2018; foundational guidance.
OAIC/CSIRO, De-identification Decision-Making Framework, 18/09/2017; foundational, with publisher age warnings.
Report mirror: Executive Summary; sections 1.1.3, 1.2, 1.3.OAIC, APP 11 security guidance, October 2025; paragraphs 11.3–11.16, 11.29–11.46, 11.54–11.58.
OAIC, Commercially available AI guidance, updated 17/01/2025.
Microsoft, Document inspection guidance, undated; desktop Office, format-specific scope.
Disclosure
How this was made: researched and drafted with AI assistance, edited and approved by the 25 hours team.