SWANK AI Guidance Note 09
Children · Safeguarding · Human Accountability

Core Standard

HEIGHTENED CAUTION WHERE CONSEQUENCES ARE HIGH

AI may assist administration.

It should not silently convert complex human evidence into determinations about children, families, credibility or risk.

Where consequences are significant, source evidence and identifiable human judgment remain essential.

Purpose

Artificial intelligence may offer administrative benefits in environments involving children and safeguarding.

These environments also require heightened caution.

Information concerning children and families is frequently:

incomplete;
sensitive;
contested;
developmental;
contextual;
and consequential.

Small interpretive errors can therefore have disproportionate effects.

The relevant governance question is not simply whether AI can process the information.

It is whether the institution can use that technology without obscuring:

source evidence;
context;
uncertainty;
professional responsibility;
and the child’s own voice.

Appropriate Administrative Uses

AI may potentially assist with functions such as:

administrative organisation;
scheduling;
document indexing;
identifying missing information;
organising records;
retrieving relevant documents;
summarising clearly identified source material;
and workflow support.

These uses may improve administrative efficiency.

They are different from asking AI to reach conclusions about complex human relationships, behaviour or risk.

The distinction should remain explicit.

Administrative Assistance Is Not Professional Determination

Greater caution is required where AI-assisted systems begin to infer:

credibility;
parental capacity;
a child’s wishes or feelings;
emotional state;
motivation;
family dynamics;
safeguarding risk;
or likelihood of future behaviour.

These are not simply text-classification problems.

They may require:

context;
chronology;
direct human engagement;
professional judgment;
source evaluation;
competing evidence;
and understanding of developmental factors.

AI may help organise the information used in a professional process.

It should not become the invisible professional making the judgment.

Children’s Voices

Where children’s views are recorded, AI-assisted processing should preserve the distinction between:

what the child actually communicated

and

what another person interpreted that communication to mean.

Where material, records should preserve:

the child’s actual words;
the circumstances in which the statement was made;
whether questions were open or directed;
relevant uncertainty;
changes in the child’s account;
and professional interpretation as distinct from the child’s own expression.

An AI-generated summary should not become a substitute for the original record of what a child communicated.

Source Over Summary

A summary may make records easier to review.

It may also remove important context.

For example, a later summary may record:

“The child stated X.”

while the original material may show:

a more qualified statement;
a question immediately preceding it;
uncertainty;
several different statements;
or professional interpretation inserted alongside the child’s words.

Where the child’s views materially affect a consequential process, decision-makers should remain able to inspect the underlying source.

Avoiding Inference From Language Alone

Language patterns may be influenced by:

age;
neurodevelopment;
education;
stress;
family culture;
disability;
communication style;
context;
and the questions asked.

Systems should therefore not treat linguistic characteristics alone as reliable evidence of:

risk;
truthfulness;
emotional state;
motivation;
family relationships;
or future behaviour.

Language is evidence of communication.

It is not automatically evidence of the underlying human characteristic an institution may be trying to understand.

Developmental Context

Children’s communication may change over time.

Their:

vocabulary;
understanding;
confidence;
emotional expression;
ability to describe events;
and willingness to communicate

may vary according to age, circumstances and context.

An AI-assisted system should therefore avoid treating every difference between records as evidence of inconsistency or unreliability.

The existence of variation may itself require human interpretation.

Context Should Survive Processing

The meaning of a statement may depend upon:

who was present;
when the statement was made;
what question was asked;
what occurred immediately before it;
the child’s age;
the communication environment;
or later clarification.

AI-assisted summarisation can preserve words while losing those surrounding conditions.

In safeguarding environments, contextual loss may materially change how later readers understand the record.

Source Integrity

Where consequential decisions rely upon summaries or AI-assisted processing, decision-makers should remain able to identify:

the underlying source;
who recorded it;
when it was recorded;
whether it was disputed;
what evidence supports it;
what qualifications were present;
and what later information may contradict or modify it.

The existence of a later summary should not break the evidence pathway back to the original material.

Preserve Disagreement

Safeguarding records may contain competing accounts.

An AI system should not resolve disagreement merely because producing a single narrative is linguistically easier.

Where relevant, the record may need to preserve distinctions such as:

The child states…

The parent states…

The professional records…

The available evidence indicates…

This remains disputed…

This has not been independently verified…

Preserving disagreement can be more accurate than producing artificial consensus.

Chronology Matters

Safeguarding information often develops over time.

A later reviewer may need to distinguish:

when an event occurred;
when it was first recorded;
when professionals became aware of it;
when an interpretation was made;
when contradictory evidence emerged;
when the child’s view changed;
and when a decision was reconsidered.

AI-assisted chronologies should preserve these distinctions.

Sequence can materially alter meaning.

Human Responsibility

AI should not become the invisible author of a safeguarding conclusion.

Where AI contributes to processing information, professional responsibility should remain explicit.

The organisation should be able to identify:

who reviewed the source material;
who interpreted it;
who made the decision;
what role AI played;
whether uncertainty was recognised;
who had authority to disagree;
and who remained accountable for the outcome.

Responsibility should remain human and identifiable.

Meaningful Human Review

Human review is not meaningful if the reviewer sees only the AI-generated conclusion.

Where the issue is consequential, reviewers may need access to:

original records;
relevant chronology;
competing evidence;
disputed information;
identified uncertainty;
corrections;
and later evidence.

A reviewer should be capable of rejecting an AI-generated interpretation where the underlying evidence does not support it.

Reassessment

Safeguarding systems should be capable of reconsidering earlier conclusions when:

new evidence emerges;
circumstances materially change;
a factual error is identified;
earlier information is contradicted;
a child’s views change;
or an initial interpretation is no longer adequately supported.

Earlier classifications should remain reviewable.

AI-assisted systems should not turn past interpretations into self-reinforcing conclusions merely because they continue appearing in later records.

Correction

Where an AI-assisted process introduces an inaccurate summary, classification or description, the institution should consider:

correcting the record;
preserving the original where audit history requires it;
identifying affected downstream records;
reviewing decisions that relied upon the error;
and making the correction visible to later users.

Correcting one field is not necessarily sufficient if the same information has already entered multiple reports, summaries or decision-support processes.

High Consequence Requires Higher Evidential Discipline

Safeguarding decisions may affect:

safety;
family relationships;
education;
healthcare;
placement;
access to services;
professional intervention;
and other significant interests.

The greater the consequence of a potential error, the stronger the case for:

source traceability;
meaningful human review;
preservation of uncertainty;
explicit disagreement;
correction pathways;
reconsideration;
and identifiable decision ownership.

High consequence should produce greater care.

It should not produce greater certainty than the evidence supports.

Questions for Organisations

Where AI is used in environments involving children or safeguarding, organisations may ask:

What task is the AI actually performing?
Is it organising information or making an inference about a person?
Can decision-makers access the original source?
Are children’s actual words preserved where material?
Is professional interpretation clearly distinguished from the child’s expression?
Is the context in which a statement was made retained?
Are disputed accounts visibly preserved?
Can an AI-generated summary be challenged?
Can a factual error be corrected across downstream records?
Can new evidence trigger reassessment?
Who made the final professional judgment?
Can the institution demonstrate what role AI played?

SIAAF Relevance

This Guidance Note principally relates to:

Domain 02 — Decision Integrity & Human Oversight

Are humans genuinely governing AI-supported decisions, or merely approving them?

Domain 03 — Evidence & Traceability

Can consequential conclusions be traced back to identifiable source evidence?

Domain 05 — Escalation, Challenge & Contestability

Can inaccurate information or AI-supported interpretations be challenged and reconsidered?

Domain 06 — Risk, Harm & Operational Resilience

Does the institution understand the consequences of error in high-stakes environments?

It may also engage:

Domain 04 — Communication & Feedback Integrity

Where children’s views, corrections and contradictory information must move accurately through the institution.

SWANK AI Standard

HEIGHTENED CAUTION WHERE CONSEQUENCES ARE HIGH

AI may assist:

organisation

retrieval

administration

workflow

and

clearly bounded summarisation.

It should not silently transform complex human evidence into determinations about:

children

families

credibility

motivation

relationships

risk.

Where consequences are significant:

preserve the source

preserve the context

preserve uncertainty

preserve the child’s voice

and

keep human responsibility identifiable.

Related SWANK AI Guidance

Guidance Note 02 — AI Summarisation and Record Integrity

Guidance Note 03 — Human Oversight in AI-Assisted Decision Systems

Guidance Note 05 — Preserving Disagreement in AI-Assisted Records

Guidance Note 08 — Reassessment Alongside Escalation

Guidance Note 17 — Source Traceability by Design

Guidance Note 20 — Governing AI in High-Stakes Environments

SWANK AI

Independent AI & Institutional Assurance

We do not just review AI. We review the institutional systems responsible for governing it.