Why use local AI?

Why use local AI?

A local model gives us more control over where prompts and documents are processed. That matters for internal policies, unpublished work and confidential information.

Know where the data goes

A cloud service processes requests on someone else's infrastructure. Retention, training use and access depend on the product, account terms and configuration. A consumer chat account and an approved enterprise service should not be treated as the same arrangement.

Local inference can keep model requests on hardware we control. Search providers, plugins, remote embeddings, backups and logs still need their own review. Installing a local model alone does not establish privacy or compliance.

Health information needs an approved workflow

Ontario's PHIPA and other applicable privacy requirements govern how organizations handle personal information. They should not be summarized as a blanket rule that health information can never leave a building. Whether a service is appropriate depends on legal authority, safeguards, agreements and institutional approval.

The Ontario privacy commissioner's cloud-computing guidance describes risk assessment, provider due diligence, contracts and accountability for public institutions. It also flags legislative changes. Use the current guidance and the organization's privacy review for an actual deployment.

Removing names is not the whole answer

Dates, unusual circumstances and combinations of details can still identify someone. Do not assume a passage is anonymous simply because names have been removed.

The trade-off

Local hosting still needs access controls, updates, monitoring and enough hardware. Its costs include equipment, electricity and maintenance. The benefit is control over the processing environment, with the responsibility that comes with it.

The original slide is historical. These notes correct its blanket legal claim and remove the unsupported fixed breach-cost estimate.