What You Need to Know
Sensitive material is anything that an organisation should not give to an outside service without a decision. It may be personal data, or it may be material that is sensitive because of who owns it or what was promised about it. A safeguard changes what leaves the organisation. This lesson teaches four safeguards, tried in a fixed order. In this lesson, a placebo is a step that looks like a safeguard but acts only after the material leaves the organisation. Deleting the conversation afterwards is one.
This lesson uses one made-up example throughout: an insurance company. A manager in its billing team wants Claude to explain why late payments rose this quarter. Her spreadsheet holds each customer's name, policy type, date of birth, address, health notes, bank account number and payments. If she uploads the spreadsheet unchanged, everything in it leaves the company. If she refuses, the rise in late payments stays unexplained. The answer is between these two extremes: a safeguard that lets the work proceed within policy.
Recognising Sensitive Material
The obvious list is names, addresses, dates of birth, reference and account numbers, health information, financial details, and anything about children. People miss the categories about ownership and promises:
- information held under a duty of confidence (a duty to keep private what someone gave you in private)
- commercially confidential figures
- third-party material governed by a contract that says how it may be handled
A supplier's pricing schedule has no personal data. It can still be the most restricted document on the desk, because a contract may govern it. The price list from the insurance company's car repair supplier is an example.
Anthropic's Usage Policy counts confidential or proprietary data as private information. It also counts non-public contact details and health data as private information. Under the policy, private information is not to be misused or accessed without permission. The policy also prohibits violating privacy rights as defined by applicable privacy laws, such as sharing personal information without consent.
Anthropic's protections article describes encryption in transit and at rest. The data is encrypted while it moves over a network and while it is stored. The article also says that Anthropic bars employee access by default. These measures protect the data that Anthropic holds. They do not replace your organisation's own decision about what may leave the organisation.
The Four Safeguards, in Order
Try the safeguards in this order. Stop at the first one that lets the work go ahead.
- Remove. Delete the details that the question does not need. Most analysis does not need the names. The late-payment question needs payment patterns. So the manager deletes each customer's name, date of birth, address, health notes and bank account number. She keeps the payments, and she can start today.
- Anonymise or aggregate. Use cohorts (groups of people) instead of individuals. Use ranges instead of exact figures. The manager groups customers by policy type and shows each group's late payments as a range.
- Use an authorised route. Move the work to a sanctioned workspace. This is the Claude workspace that your organisation provides, under an agreement that covers this class of material.
- Seek authorisation. Ask whoever owns the data or the policy. Use this step only when the work cannot go ahead without the sensitive material. Seeking authorisation is an action, not a refusal.
Why this order? The manager can do steps 1 and 2 herself, today. Steps 3 and 4 wait for a different workspace or for another person's answer, so they cost time. When removal already lets the work go ahead, steps 3 and 4 add that time and give no extra protection. The mistake is to stop and wait for an answer that the work does not need.
Abandonment means dropping the analysis. It is not one of the four safeguards. Anonymisation lets the analysis go ahead, so abandonment gains nothing.
Why the Placebo Fails
The policy control is about what leaves the organisation. The material leaves at upload, and a later step cannot undo the upload. Every placebo fails for this reason. The placebos are telling Claude to forget the material, declaring the upload confidential, adding a confidentiality line to the prompt, or deleting the conversation. If the manager uploads the full spreadsheet and then asks Claude not to retain it, the details have still left the company. Nothing that is typed into the chat can retroactively govern what was uploaded. The safeguard happens before the upload or not at all.
Which Product Are You On?
Two current Anthropic privacy articles share the exact title "Is my data used for model training?". One article covers consumer products, and the other covers commercial products. The two articles give different answers. By default, Anthropic does not use inputs or outputs from its commercial products, such as Claude for Work, to train its models. Data handling depends on the plan, and the details have changed over the history of the documents. So ask three questions rather than memorise today's answer. Which product am I on? What does its documentation say today? What does my organisation's agreement add? The manager asks these questions first. Objective 6.3 covers what follows from the answers.
In the exam
In the guide's sample question for this domain, a project manager wants to upload a spreadsheet of customer names and account numbers. The right answer neither uploads the spreadsheet as it is nor abandons the analysis. It anonymises the spreadsheet so the work can proceed within policy. This pattern shows the whole objective in one example: a safeguard that enables the work, between two wrong extremes. Abandonment is almost never the right answer, because the guide's rationale says anonymisation enables the analysis. It also says instructing the model not to retain data does not satisfy the policy control.
What to notice in the exam
Any option whose safeguard acts after the upload is wrong. No placebo is ever the right answer. The right answer changes what is uploaded, not what happens afterwards.
Deletion and retention details change often. The Domain 6 quick reference has the current, dated figures.
Exam Traps
Treating a confidentiality instruction or later deletion as the safeguard
The exam guide says instructing the model not to retain data does not satisfy the policy control. The disclosure happens at upload, so a later instruction or deletion cannot undo it. Remove or anonymise the identifiers before uploading.
Moving to a later safeguard when removal is enough
Moving the work to a sanctioned workspace or seeking authorisation looks diligent. But when removal of the identifiers lets the work go ahead today, steps 3 and 4 cost time and add no protection. Use the first safeguard in the order that works.
Approving a document for upload because it contains no names
Sensitivity is not only personal identifiers. Duty of confidence, commercial confidentiality and third-party contractual restrictions apply even when no one is named. Check whose material it is and what the agreement allows, then leave it out or ask the owner.
Dropping the analysis because personal data is involved
Abandoning the task is unnecessary when anonymisation enables it. Removing the identifiers lets the analysis proceed within policy, so abandoning the task loses a legitimate insight and gains nothing. Strip or anonymise first, then run it.
Practice Scenario
The head office of a retail chain runs a training course for its delivery drivers. The training coordinator wants to use Claude to understand why pass rates dropped this year. The trainee records contain names, licence numbers, dates of birth and health declarations. The data policy of the company allows personal or health details into an AI tool only where the task cannot be done without them. What is the most appropriate way to proceed?
Build Exercise
Apply the Four Safeguards to One Real Dataset
What you'll learn
- How to inventory the sensitive elements in a real document
- How to separate what the question needs from what the document contains
- How to apply removal and aggregation before using escalation
- How to name the authorised route or the owner when the work needs an element
- Pick one real document or dataset you would genuinely like help with, and inventory every sensitive element: identifiers, health, financial, children, duty of confidence, commercial confidence, third-party material under contract
Why: Recognition comes before technique, and the under-recognised categories only become visible when you look for them deliberately.
You should see: A list of sensitive elements, each tagged with why it is sensitive: who it identifies, who owns it, or what was promised about it.
- Write down the actual question the work must answer, then mark which sensitive elements that question truly needs
Why: The gap between what the document contains and what the question needs is the material that the safeguards act on.
You should see: A question in one sentence, and a much shorter list of elements it genuinely depends on; expect fewer than you assumed.
- Apply the first two safeguards: remove what the question does not need, and aggregate or anonymise what it needs only in outline
Why: The first safeguard in the order that works is the professional answer, and it usually lets the work proceed today.
You should see: A prepared version of the material with identifiers stripped and remaining sensitive elements reduced to cohorts or ranges.
- If the work genuinely needs a sensitive element, name the authorised route if one exists, or the person who owns the data or the policy, and draft the short request you would send them
Why: Escalation is an action, not a refusal, and naming the actual route or owner is what separates it from vague hesitation.
You should see: Either the name of a sanctioned workspace that covers this class of material, or a two-sentence request to a named owner.
- Write down the placebo you were most tempted by, and say in one sentence why it is not a control
Why: Naming your own placebo protects you from choosing it when it appears as a wrong option under exam pressure.
You should see: One sentence, in your own words, explaining that the placebo operates after the disclosure.
Sources
- How long do you store my data? — Anthropic
- Is my data used for model training? (consumer products) — Anthropic
- Is my data used for model training? (commercial products) — Anthropic
- How does Anthropic protect the personal data of Claude users? — Anthropic
- Usage Policy — Anthropic
- Claude Certified Associate – Foundations Exam Guide — Anthropic. The guide's Sample 3 rationale gives the first safeguard and the placebo principle. Steps 3 and 4, and the order of all four, are this lesson's reading of the guide. The Domain 6 quick reference has the current, dated retention and deletion details.