Safeguarding with AI

Module 6 taught you what AI does to language: it inflates, it smooths, it drifts toward the average it was trained on. Now apply that to safeguarding, where the inflation lives next to a statutory threshold and someone has to decide whether to cross it.

Two thresholds, one principle.

Different statutes, same shape, and the threshold call belongs in both cases to the practitioner and the people above them in the chain. The decision is not something AI is positioned to make.

How AI inflates risk

Module 6 showed you the mechanism. Here it is again, with the safeguarding stakes attached.

In an AOP module these matter because of bias. In safeguarding they matter because the inflated language sits next to a threshold word. Words like significant and unable to protect are doing legal work in s47 and s42, not adjective work, and AI uses them as adjectives anyway. The threshold word lands in the document with no awareness of what it has just triggered.

A short worked example. A practitioner uploads a four-line referral and AI returns a "professional summary" that uses significant harm in the second sentence. The threshold language is now in the document, attributed implicitly to a human reader, and authored by AI. By the time the document reaches the panel, nobody has flagged who wrote the phrase or whether it was warranted.

Agreeing with the machine when we don't actually agree

In Section 1 we named this as automation bias. It deserves its own pillar because in safeguarding it is the most insidious failure mode.

This is where the practitioner-supervisor-manager chain becomes load-bearing. You are not the only safeguard. Saying "the AI summary was wrong, so I made changes" out loud, in supervision, is the system functioning as it should. That sentence in a supervision room is the bit that catches the drift the document hides.

Iteration as formulation aid (the good use)

The opposite of agreeing with the machine is arguing with it. Use AI iteratively to think a case through, not to be told what the answer is.

Three prompts worth keeping on a sticky note next to your screen:

These are formulation prompts. They produce questions for you, rather than answers for the file. The output is your sharper thinking, not the record itself. What you put in the document is still yours.

The decision chain (and where AI is not in it)

A safeguarding decision is never the practitioner's alone. Every local authority is different, but your chain might look something like this:

AI is not part of this chain. It is a tool that any practitioner may reach for, but it has no seat at strategy, no signature on a court report, and no role in the moments where someone has to take responsibility for a decision. The chain is human throughout, and that is by design.