Does the framework require every dataset to be anonymous?

No. The overall arrangement can support controlled access to detailed data, depending on the applicable assessment and rules.

Are the five dimensions independent?

No. They interact and should be considered together.

Does following Five Safes replace other obligations?

No. It structures data-access decisions; the organization must still establish the applicable permissions, responsibilities and requirements.

AI Glossary / Data engineering and architecture / Intermediate

Five safes

Definition

The Five Safes is a framework for assessing and managing data access through five connected dimensions: safe projects, safe people, safe settings, safe data and safe outputs. It considers the whole use of data rather than relying on a single protective measure.

What it means

The framework is associated with controlled research access to sensitive or confidential data. It asks whether the proposed purpose is appropriate, the users are suitable, the working environment is controlled, the data is prepared appropriately and the released results avoid unacceptable disclosure.

Safe does not mean risk-free. The dimensions guide a contextual assessment, and their implementation depends on the data, purpose and access arrangement. They are not independent certificates guaranteeing acceptable use.

The framework also differs from simply removing names. A dataset can remain revealing through combinations of attributes, while a controlled environment and reviewed outputs can support uses that would be unsuitable for unrestricted distribution.

How it works

Safe projects concerns the proposed use and whether it meets the data service's approval criteria. Safe people concerns the users' suitability, training, responsibilities and authorization. Neither is established merely because a person has a technically valid account.

Safe settings concerns the environment and controls through which analysis occurs. These may limit copying, external connections or the movement of files. Safe data concerns the information available to the project and the treatment needed to manage its disclosure risks while supporting the approved purpose.

Safe outputs concerns what leaves the controlled environment. Tables, figures, text and other derived material can require assessment before release. An aggregate is not automatically harmless if it reveals information about a very small or otherwise identifiable group.

More detailed data may require stronger controls elsewhere, but one strength does not excuse an unrelated weakness without assessment. Services apply their own procedures and decision rules, so the framework should not be mistaken for a universal numerical scoring formula.

Three practical examples

A research service reviews a project's purpose and researcher credentials before allowing controlled access to detailed records.
A data owner provides an approved subset inside a restricted analysis environment instead of distributing the full dataset as an ordinary download.
An output-review team checks a proposed table and accompanying narrative before either is released from a secure workspace.

Think of it this way

Think of a supervised archive reading room. The institution considers why someone needs access, who they are, where they work, which materials they receive and what copies they can take away.

The analogy is limited because digital disclosure can arise through combinations, calculations and models rather than a direct photocopy. The Five Safes requires attention to the actual data use and outputs, not just physical supervision.

Worked example

Checking an output that appears to contain only totals

An invented research project analyses three small business groups. The proposed table reports a combined total of ninety for all groups and separate totals of forty and thirty for two of them. It suppresses the third group's value because that group contains only one business.

A reader can nevertheless calculate the hidden value as 90 − 40 − 30 = 20. Suppressing one visible cell has not removed the information when other released totals reconstruct it.

An output review therefore considers the table as a whole and related material already available. Possible changes depend on the service's rules and the disclosure context; this example does not prescribe a universal suppression threshold or approved solution.

The researchers may be trained, the project approved and the analysis environment controlled, yet the proposed release still needs attention. The example shows why safe outputs is a dimension and why reviewing cells in isolation can miss a problem.

Illustrative case study

A fictional research partnership studies how organizations adopt automation. Its team plans to work with detailed survey records inside a controlled environment. The data owner reviews the research purpose, approves named researchers and provides training on the environment's restrictions.

During analysis, a researcher proposes using an external AI assistant to summarize selected rows. The team recognizes that this would move information beyond the approved setting and pauses that proposed workflow for assessment. Existing access approval does not automatically cover a new processing service.

The final report is reviewed for disclosure in both tables and narrative examples. This includes checking whether contextual descriptions could reveal a respondent even without a name. The case illustrates that the five dimensions remain relevant as tools and outputs change throughout a project.

Common misconceptions and mistakes

Removing names completes the framework. Other attributes, environments, purposes and outputs still matter.
Approved researchers can use any external tool. Their authorization is tied to the applicable project and access conditions.
Aggregated results never disclose information. Small groups and combinations of released values can reveal protected details.
Five Safes is a guarantee of zero risk. It is a framework for contextual risk management and controlled use.

Business use

The framework gives data owners, analysts and service operators a common structure for discussing access. It clarifies responsibilities and changes requiring review.

Document the approved purpose, participants, environment, data scope and output process. Revisit these when a project changes tools, adds collaborators or proposes new forms of release. A decision about one workflow should not silently become approval for a different one.

For AI work, examine both input movement and derived artifacts. A model, embedding or generated narrative may need assessment under the applicable output process; it should not be treated as automatically unrestricted merely because it differs from the source table.

Check your understanding

A suppressed cell can be reconstructed by subtracting two published group totals from a published overall total. Has hiding that cell alone prevented disclosure?

Answer: No. The combined output still reveals the value. Review must consider relationships among released information.

Related terms

Statistical Disclosure Control (SDC)
Data governance
Access Control
De-identification
Trusted research environment
statistical disclosure control
output checking
data stewardship

Learn next

Read statistical disclosure control and trusted research environments. Map a proposed analysis to the five dimensions and identify who decides whether a change of tool, participant or output remains within the approved arrangement.

Statistical disclosure

Sources

Office for National Statistics, About the Secure Research Service: https://www.ons.gov.uk/aboutus/whatwedo/statistics/requestingstatistics/secureresearchservice/aboutthesecureresearchservice. Primary service description of the Five Safes in practice.

Five Safes, Using the Five Safes: https://fivesafes.org/using-the-five-safes/. Primary framework guidance on interactions among controls and access arrangements. The arithmetic and research-partnership case are original teaching examples, not an assessment of a real release.

Previous: Fitness function Next: Fixed Effects Model