AI Usage Policy

1. Where machine processing is used

One place is approved: extracting the text printed on a worker's identity document so the worker does not have to type it. That is described in the OCR Usage Policy.

The approved provider is Google Cloud Vision, called server-side only. As at this version it is not yet integrated: no document has been sent to it, and identity fields are entered by hand. The Subprocessor Register records it as approved and not yet in use.

Turning it on is a material change under section 11 of the governance framework. It requires a new version of this policy, the OCR Usage Policy and the Privacy Policy, an updated Subprocessor Register, a governance decision, and renewed acceptance from affected workers before their documents are processed by it.

Warsha uses no other machine learning. There is no automated ranking model, no automated pricing model, no automated content moderation model, and no generative feature.

2. What machine processing may never decide

These are absolute limits, enforced in the platform rather than stated as an intention:

A low-confidence or ambiguous extraction never produces a rejection. Confidence values are internal, never shown to the person, never shown to a customer, and never a reason for a decision.

• Whether an identity document is genuine.
• Whether a document belongs to the person who submitted it.
• Whether a document has been altered.
• Whether a criminal record makes someone eligible to work.
• Whether an account is suspended, restricted or closed.
• The outcome of an appeal.

3. Human decision-making

Every adverse decision about a person on Warsha is made by a person, recorded with the evidence they actually saw, and open to appeal to a different person.

Machine output may be shown to a reviewer as information. It is never presented as a recommendation, a score, or a conclusion, because a reviewer shown a score decides the score rather than the case.

4. Training

Identity documents, criminal-record certificates and the fields extracted from them are not used to train, fine-tune, evaluate or improve any machine-learning model. Not by Warsha, and not by a provider on Warsha's behalf.

This is not a current preference that a future team could quietly reverse. Changing it requires all of the following, and the platform is built so that none can be skipped:

• A recorded governance decision, taken by people with the authority to take it.
• An updated version of this policy, the Privacy Policy and the Worker Verification Policy.
• An updated Subprocessor Register and Data Processing Register.
• Notice to every affected person, in advance, stating what would be used and for what.
• Explicit, separately recorded consent from each person whose documents would be included, where consent is the basis relied on.
• A versioned rollout, so that it is always known which documents were in scope from which date.

5. Provider terms

Where a provider is engaged, Warsha will only engage one whose terms prohibit using Warsha's data to improve the provider's own models, and the register records that this was checked.

A provider that requires such use is not eligible to be a Warsha subprocessor for identity data.

6. Future features

Warsha may add machine-assisted features later — better search, fraud signals, help-centre assistance.

Each would require its own entry in this policy, its own governance decision, and its own assessment of what it may and may not decide. None may cross the limits in section 2, and none may take a decision about a person without a human confirming it.