Begin with the actual material and task
“Is this AI safe?” is too broad to answer usefully. Identify the information, the proposed operation, the account and service, the people with access, and where the output will be kept. A public training exercise and an active client record raise different questions.
The Law Society of Ontario's technology resource centre directs licensees to its generative-AI guidance. Consult current applicable professional guidance, firm policy, client arrangements, and the actual service terms before using client material. This lesson provides a technical review structure, not a conclusion that a particular setup satisfies those obligations. [source]
Draw the data journey
For a PDF analysis workflow, list: original document location; upload or local reading; OCR; embeddings; retrieval storage; model inference; logs; output; backups; and sharing. For each stage, record the operator, location, retention, access, and onward transfers.
A local model can sit inside an application that uses remote OCR or stores logs elsewhere. A cloud application may have different enterprise and personal arrangements. Inspect the selected configuration rather than inferring data handling from a brand or desktop icon.
Ask questions with verifiable answers
| Question | Evidence to seek |
|---|---|
| Where is processing performed? | Architecture and configuration for every stage |
| Who can see inputs and outputs? | Actual account, workspace, and sharing settings |
| Is material retained or used for training? | Applicable terms and current account controls |
| What do logs contain? | Logging configuration and a controlled test |
| Can access be revoked and data removed? | Documented procedures and their limits |
| Can one matter retrieve another's material? | Authorization design and boundary tests |
A vendor's general statement can begin the review. It should be matched to the account, product feature, contract, and configuration actually in use.
Test with fictional material first
Create two synthetic matters with unmistakably different names. Confirm that a user limited to one cannot retrieve the other's documents through direct search, generated answers, exports, or shared outputs. Test what happens when access is removed. Keep sensitive facts out of diagnostic logs where the task does not require them.
For offline claims, test the complete pipeline under the intended network restrictions after setup. Ollama documents local-only configuration, but a separate interface or connected component needs its own review. [source]
Try it: a private interface
The web interface requires sign-in, but the underlying model endpoint is reachable without authentication and the document store uses shared credentials. Has privacy been established?
Check your understanding
No. The review has identified additional access paths requiring investigation and controls. Securing one interface does not establish that every component enforces the same boundary. Restrict the unresolved paths and test the complete architecture before relying on its access model.
Leave a decision record
Record the permitted use, approved data categories, required settings, review owner, unresolved questions, and date for reconsideration. Make the scope explicit. Authorization to test a fictional chronology does not itself authorize uploading a client production or enabling external tools later.