What AI must not be asked to decide

Most organisational AI policies are written after the first deployment, which means they are written to accommodate it. The line ends up drawn around whatever was convenient to build. We would rather draw it before, when it costs nothing to draw it in the right place.

Where AI is genuinely useful to us

There is real work here and no reason to be precious about it.

Transcription of recorded teaching, so that everything published has a text alternative without waiting for a person to type it. Tagging content against a controlled taxonomy. First-draft translation, for a human translator to work from rather than start from nothing. Search. Administrative triage — routing an enquiry to the right queue faster than a person reading a backlog.

Each of those is a labour problem, and in each case a mistake is recoverable and visible. That is the pattern: AI is well suited to work where being wrong is cheap and being wrong is obvious.

Where it must not go

Five decisions, in our case, are permanently off the table. None of them may be made, scored, ranked or recommended by a model.

Whether someone is saved. This should not need saying, and it does, because the machinery for it is trivial to build and someone will build it. A model can classify a form response. It cannot know what happened between a person and God, and a system that assigns a confidence score to that is not a tool, it is a category error with a database behind it.

Doctrinal fitness. Whether a person is suitable to teach, mentor or be commissioned is a judgement about character, conviction and competence, made by people who have met them and who can be held responsible for getting it wrong.

Eligibility for pastoral support. Someone asking for help in a moment of exhaustion is not a classification problem. A model trained on previous decisions will reproduce whatever bias was in them, and it will do so at a speed that makes the bias invisible.

Safeguarding outcomes. A concern is assessed by the designated safeguarding lead. Triage may flag urgency; it does not assess, and it does not close anything.

Funding and admissions. Both affect people materially and both must have a named human being answerable for the outcome.

The principle underneath all five: where a decision has an irreversible effect on a person's life, faith or safety, somebody must be answerable for it who can be asked why, and who can be wrong in a way that means something.

No AI wearing a person's face

There will be no AI on this site presented as a pastor, a counsellor, a spiritual adviser, or as our founder.

The technology to build a convincing one exists and is cheap. The temptation is obvious: it would answer at three in the morning, in any language, without getting tired — precisely when a person is least able to tell the difference and most likely to be affected by it.

It would also be a lie about the most important thing on the site. If a person asks to talk to someone about God, the entire value of what happens next is that there is someone there. Simulating that is not scaling pastoral care; it is removing it while leaving the label on.

Disclosure, and the boring parts

Where AI has assisted in producing something published — a translation, a transcript, a summary — that will be disclosed. Human review is required before publication in every case, and the reviewer is named in our editorial workflow.

Two further constraints that matter more than they sound. AI tooling does not get access to confidential or special-category records: prayer requests, Gospel follow-up, applications, pastoral-support enquiries and safeguarding reports are outside its reach entirely, not merely excluded by policy. And no personal data is sent to a model provider for training.

The honest caveat

We are an organisation whose founder has technology interests, building a site with technology suppliers who may be connected to him. That is exactly the situation in which an AI policy gets quietly relaxed to fit what a related party is selling.

So the policy is published, the related-party governance is published alongside it, and both are things we can be held to. If we ever move one of the five lines, it will be visible, it will be dated, and someone will have to explain why.