Is DeepSeek Safe? Privacy and Government Bans

No, we would not approve DeepSeek’s public chatbot for identifiable client data. Its privacy policy permits model training and says personal data is processed and stored in China.

Is DeepSeek Safe for Client Data?

Keep identifiable client material out of DeepSeek’s public app and website. For a tax practice, that means using fictional examples for experimentation instead of uploading client documents or copying an email into a prompt.

The Office of the Australian Information Commissioner (OAIC) recommends keeping personal information out of public generative AI tools. That recommendation particularly concerns sensitive information. A tool’s answer quality does not establish that its data handling suits client work.

DeepSeek’s published pricing lists usage-based developer access, without a separate business chatbot plan. Paying for an application programming interface (API) does not itself give the practice a confidential workspace or negotiated client-data terms.

DeepSeek does have Open Platform terms for enterprise developers. These place responsibilities on developers for downstream applications and personal information. Assess a proposed API arrangement separately from the public chat service.

For a graduate checking a calculation, a fictional amount and generic question can avoid disclosing a client’s affairs. Copying the client’s workpaper into the chatbot changes that assessment. Review the calculation against the underlying tax rules before using it.

Why the Australian Government Banned DeepSeek

Home Affairs assessed DeepSeek as an unacceptable security risk to government systems. The Protective Security Policy Framework direction applies to non-corporate Commonwealth entities.

Direction 001-2025 requires those entities to remove existing DeepSeek products and prevent access, use or installation. It covers government systems and devices, including authorised non-government devices within its definitions. Blocking installation alone does not address browser access.

Home Affairs identifies extensive data collection and exposure to foreign-government directions that conflict with Australian law. The direction also provides a narrow exemption route for national security and regulatory functions, subject to required mitigations.

This direction does not impose a general DeepSeek ban on every private accounting practice. Private practices must make their own decision under applicable obligations and client contracts. Work involving government systems or contractual security conditions needs a separate assessment.

The direction’s full text excludes certain open models deployed locally, where the entire codebase is inspectable and appropriate mitigations apply. That conditional exclusion does not approve every installation labelled local.

What DeepSeek’s Privacy Policy Says

DeepSeek’s privacy policy permits collection of prompts, uploaded files and chat history. It also covers account details and device information, including IP addresses and usage logs. DeepSeek says personal data is collected, processed and stored in China.

The policy lists model training among its purposes and provides a right to opt out of training or technology optimisation. An opt-out does not establish local storage or eliminate other permitted processing.

DeepSeek also tells users not to provide sensitive personal data.

Sharing includes service providers and corporate-group entities. Search providers receive input keywords. Disclosures can also occur for legal requests or corporate transactions. Retention depends on service needs and other stated grounds, without one universal deletion deadline.

DeepSeek’s Terms of Use also explain that anyone with a shared conversation link can view its contents. A link created to show a colleague an answer can expose the prompt too. Keep client information out of shared examples.

Changing the training preference therefore answers only one part of the privacy decision. A practice also needs to assess disclosure, access and retention before approving any service for client records.

Running DeepSeek Models Locally

Open DeepSeek models can run on hardware the practice manages without sending prompts to DeepSeek. This requires downloaded model files and software configured to perform the calculation locally, with cloud services and external connections excluded from the task.

DeepSeek’s R1 model repository supplies open model weights and smaller distilled models. Those smaller models require fewer computing resources than the full model. The chosen model and its licence determine what the practice can deploy.

Hardware needs depend on model size and the amount of text processed at once. A model download’s file size is not the complete memory requirement.

Have the person managing the computers select a model that fits the available memory and workload.

Ollama’s local-processing documentation distinguishes local models from its cloud features. It states that locally run prompts stay out of its service and describes disabling cloud models and web search. The surrounding chat application still needs its own assessment.

A local setup gives the practice responsibility for access and saved chats. Include backups and connected document-search tools in that assessment.

Downloading an open model does not establish that every component around it keeps client information on the same computer.

Use the local LLM guide for the setup choices. If the practice chooses a hosted service instead, assess its business terms and settings, including the distinctions in ChatGPT privacy, before authorising client use.