The Risk of Losing the Child in Artificial Intelligence

Posted by NAYJ on Jul 29, 2026

Joe Allen-Fellows, a youth justice practitioner and teaching fellow at the University of Portsmouth, considers challenges for practitioners engaging with AI tools to support their work with children. 

There is a particular moment in youth justice practice that no algorithm can fully capture, such as when a child finally decides to tell you what is really going on.

Not the version written in assessment documents. Not the police summary. Not the carefully sanitised language of ‘criminogenic factors’ or ‘adverse childhood experiences’. I mean the real story. The one that emerges twenty minutes into a conversation about trainers, music, or why they hate school lunches. The moment where a child stops performing toughness long enough to reveal fear, grief, shame, exploitation, or exhaustion.

That moment matters because youth justice, at its best, is relational work. It depends upon trust, interpretation, and human connection. Yet increasingly, practitioners are being asked to operate in systems shaped by data extraction, spreadsheets, and growing institutional enthusiasm for artificial intelligence.

Across criminal justice agencies, including the UK Ministry of Justice and His Majesty’s Prison and Probation Service, there is increasing interest in AI-supported tools designed to improve efficiency, assess risk, support decision-making, and manage information (Ministry of Justice, 2024). On one level, this makes sense. Most practitioners I know are drowning in admin and paperwork. If AI could genuinely reduce paperwork and free us to spend more time with children, many practitioners would welcome it enthusiastically. Few youth justice workers entered the profession dreaming about autocomplete functions in case management systems.

But beneath the optimism sits a deeper tension. Youth justice has spent the last decade attempting to become more child-centred, trauma-informed, and participatory. The emergence of Child First principles represented an important shift away from seeing children primarily through the lens of offending behaviour and toward recognising their strengths, identities, and developmental needs (Youth Justice Board, 2021). But AI systems often operate through categorisation, prediction, and standardisation. These are not neutral processes. They shape how children are understood and as a practitioner and researcher, I find myself increasingly asking whether AI risks amplifying some of the very problems youth justice has been trying to move away from.

Children in conflict with the law are already heavily documented populations. By the time many enter youth justice services, they may have accumulated years of educational records, police intelligence, social care involvement, exclusion histories, mental health referrals, and risk assessments. AI systems trained on these forms of institutional data may appear objective, but they are built upon histories of professional judgement, systemic inequality, and selective attention (Eubanks, 2018).

This matters because youth justice data is not simply a record of children’s behaviour. It is a record of how institutions have responded to certain children, if a child has been repeatedly excluded from school, over-policed in their community, or disproportionately viewed as “risky”, those patterns may become embedded within predictive systems. AI does not arrive free from human bias and, often, inherits and reproduces it at scale (O’Neil, 2016). A computer may not consciously stereotype a child, but systems trained on unequal social realities can still produce discriminatory outcomes.

In theory, professional judgement remains central. In practice, institutional cultures shaped by risk aversion can make technological outputs feel difficult to challenge. Anyone who has worked in youth justice knows the subtle power of assessment tools. Once a child acquires a particular label, it can follow them everywhere like an unwanted social media tag. ‘High risk’ can become less an assessment and more an identity and there is a danger that AI could intensify this process.

The erosion or amplification of a distorted narrative is likely to have significant consequences. Children communicate through stories, contradictions, humour, silence, and emotion. Often, the most important information emerges relationally and gradually. A child may initially present as aggressive or disengaged, only for practitioners to later discover exploitation, bereavement, neurodiversity, or unmet mental health needs. Human interaction allows for reinterpretation. Many good practitioners revise their understanding constantly.

AI systems, however, tend to prefer fixed inputs and measurable variables. They are significantly less comfortable with ambiguity. Unfortunately, ambiguity is where much of childhood lives, as do, if we are honest, most of our adult lives.

This is where practitioners become critically important. Increasingly, retaining the child’s voice may depend upon practitioners actively resisting the reduction of children into data profiles. That does not necessarily mean rejecting technology altogether, it means refusing to allow technological systems to become the dominant narrative about a child.

Many practitioners are already navigating this tension carefully. Some use technology pragmatically while consciously protecting relational space. Others challenge overreliance on automated processes by foregrounding a child’s own account within assessments and interventions. In practice, this can involve something deceptively simple, such as writing records and assessments differently.

The language practitioners use matters enormously. There is a world of difference between writing that a child ‘demonstrates manipulative behaviours’ and writing that a child ‘has learned survival strategies within unstable environments’. One path views the child as a problem. The other situates behaviour within context.

Children notice these distinctions too. I have lost count of the number of times young people have told me they feel like reports describe someone they do not recognise. One young person once looked at an assessment and asked me, ‘Have they actually met me?’ Brutal, but fair.

Safeguarding children’s rights within AI-influenced systems therefore requires more than ethical guidance documents. It requires practitioners willing to maintain curiosity, critical reflection, and advocacy even when systems encourage speed and certainty.

This includes protecting children’s participatory rights. Article 12 of the United Nations Convention on the Rights of the Child establishes children’s right to express their views in matters affecting them (United Nations, 1989). If AI-supported assessments begin shaping decisions about intervention, risk, sentencing, or supervision, children must still have meaningful opportunities to challenge and contribute to those processes. Otherwise, participation risks becoming symbolic rather than substantive.

Practitioners may increasingly act as translators between children and technological systems.

There is also a need to protect practitioner identity itself. Youth justice workers are not simply information processors. Relational practice involves empathy, creativity, emotional labour, ethical judgement and occasionally vicarious trauma. If practitioners become overly dependent on AI-generated summaries or risk classifications, there is a danger that professional curiosity narrows. The question shifts from ‘What happened to this child?’ to ‘What category, or dataset, does this child belong to?’

To be clear, AI may bring genuine benefits. Used carefully, it could reduce administrative pressures and improve access to information, therefore increasing actual time with children. But efficiency cannot become the primary measure of good youth justice practice. Children are not administrative problems to optimise.

The uncomfortable truth is that youth justice already struggles with balancing care and control. AI risks tipping that balance further toward surveillance, prediction, and managerialism unless practitioners actively resist it. Retaining the child’s voice therefore becomes both a professional responsibility and a form of ethical resistance.

Because ultimately, children do not need systems that know more data about them. They need adults who remain willing to listen carefully enough to understand what the data cannot say.

References

Eubanks, V. (2018). Automating inequality: How high-tech tools profile, police, and punish the poor. St. Martin’s Press.

Ministry of Justice. (2024). Responsible artificial intelligence in the justice system. UK Government. UK Ministry of Justice

O’Neil, C. (2016). Weapons of math destruction: How big data increases inequality and threatens democracy. Crown.

United Nations. (1989). Convention on the Rights of the Child. United Nations. United Nations Convention on the Rights of the Child

Youth Justice Board. (2021). Child First framework. Youth Justice Board for England and Wales