Artificial intelligence in Australian prison psychiatry: Promise,pitfalls,and ethical imperatives

Artificial intelligence (AI) is increasingly discussed as a tool to enhance healthcare and risk management in correctional settings. Potential applications in the custodial setting may include automated risk prediction, natural language processing of clinical records, behavioural monitoring, and personalised treatment planning.

In Queensland, as in most states, correctional services operate under substantial workforce pressures. High rates of mental illness, substance misuse, and behavioural disturbance place significant demands on stretched clinical resources. Risk assessment is already deeply embedded within Australian correctional practice, where structured tools are routinely used to inform decisions around management, rehabilitation, and release. However, these instruments remain largely actuarial in nature, rather than genuinely AI-driven, and there is as yet no coordinated or sustained move toward the adoption of fully AI-based predictive models. Nonetheless, given the broader trajectory of technological development, this position may not remain static. Internationally, there are early indications of a shift in this direction; the United Kingdom’s Ministry of Justice has outlined an AI Action Plan exploring the use of algorithmic tools to assess the risk of violence within prisons. ¹

These factors create challenges for early identification of deterioration and timely intervention. AI-driven systems capable of analysing large volumes of clinical and behavioural data may support clinicians in prioritising care and identifying emerging risks.^2,3

However, deploying AI in prison psychiatry raises profound ethical, clinical, and legal concerns. Correctional environments differ substantially from community settings: autonomy is restricted, surveillance is pervasive, and clinical decisions may intersect with security objectives. AI applications must therefore be carefully governed to ensure that they augment rather than undermine humane, rights-based care,^4,5 and that prison psychiatry services are not taking on the role and objectives of the custodial system within which it operates.

AI-based tools may assist clinicians in several domains relevant to prison psychiatry. Risk-prediction models can estimate the likelihood of self-harm, suicide, aggression, or recidivism based on historical clinical and behavioural data.^3,6 Ideally, such models would identify individuals at elevated risk more quickly than traditional screening methods, potentially allowing for timely intervention and targeted support. Yet there remains limited evidence to suggest that risk-prediction models are effective.^7,8

Natural language processing (NLP) tools can extract relevant information from potentially voluminous electronic health records or incident reports, helping clinicians detect symptom trends or changes in behaviour or speech.^2,3 However, caution is warranted. Systems trained on general psychiatric datasets may lack the contextual sensitivity required to accurately interpret the linguistic and emotional nuances of custodial environments. In such settings, expressions of frustration, anger, or trauma may be normative or adaptive, yet risk being misclassified as markers of heightened risk. This raises the possibility of systematic overestimation, with downstream clinical and relational consequences—including the potential to inhibit openness in therapeutic encounters if individuals perceive their communications to be subject to algorithmic scrutiny.

Predictive models may also guide personalised treatment planning, potentially improve rehabilitation outcomes and facilitate targeted psychosocial interventions. ³ By providing timely decision support, AI could enhance efficiency without replacing clinical judgement. ³

In addition to violence and self-harm prediction, some correctional agencies are exploring biometric behavioural profiling combined with AI to prevent in-custody deaths and medical emergencies. For example, the Maricopa County Sheriff’s Office in Arizona has proposed wearable devices to monitor heart rate, body temperature, and other physiological indicators, while jails in Colorado, Alabama, and other parts of Arizona have already implemented similar systems. ⁹ These technologies aim to detect early signs of medical deterioration, including overdoses, cardiac events, or suicidal crises, potentially enabling staff to intervene more rapidly than traditional observation and even preventing in-custody deaths. While promising, these approaches raise serious ethical considerations that feel almost dystopian in nature: including privacy, consent, and the potential misuse of sensitive health data for the purposes of controlling an already vulnerable population. Indeed, the premise of the 2002 film Minority Report, ¹⁰ starring Tom Cruise, centres on a futuristic “Precrime” unit that seeks to prevent violent acts before they occur through predictive insight derived from three psychic individuals—a concept that bears a striking resemblance to emerging discussions around the use of artificial intelligence to anticipate and prevent violence.

These challenges intersect with the longstanding issue of dual loyalty in prison healthcare, where clinicians balance patient welfare against institutional security mandates. If clinical data are used for disciplinary or security purposes, the boundary between healthcare and custodial surveillance may blur.^4,11 Strong governance frameworks must ensure that AI supports clinical care without compromising autonomy, privacy, or trust.

Effectiveness of such tools depends on local validation. Models trained on datasets from other jurisdictions or populations may not generalise to the specific demographic and cultural characteristics of prisoners in Queensland.^2,3 Pilot implementations with rigorous evaluation and cultural oversight are therefore essential before broad adoption. AI-driven CCTV systems are beginning to be used in some prisons around the world to detect and analyse “suspicious” patterns of behaviour to purportedly predict violence. There is a lack of high-quality evidence showing that AI-enabled biometric monitoring in prisons definitively improves outcomes such as mortality, self-harm, or violence reduction. Most evidence is vendor or prison reported, rather than based on robust independent trials.

Algorithmic risk-prediction models carry significant risks of bias. Such tools rely on historical datasets, which frequently reflect and perpetuate systemic inequalities in policing, judicial, and correctional settings. When these datasets are used to train AI models, racial and other social disparities risk being reproduced or amplified.^5,6 Aboriginal and Torres Strait Islander people are significantly overrepresented in custody. Conventional risk assessment instruments in Australian prisons have been found to over-classify Aboriginal and Torres Strait Islander men as high risk despite not necessarily reoffending. ¹² AI systems trained on similar data would very likely perpetuate these inequities, leading to inflated risk scores for Indigenous prisoners and potentially affecting parole decisions, classification levels, or access to rehabilitation programs,^5,6,12 further increasing rates of incarceration. Similarly, predictive models for self-harm may misclassify individuals who exhibit culturally or linguistically diverse expressions of distress if these patterns are underrepresented in the training data. ⁶

Mitigating these biases requires demographic auditing, culturally informed model design, and ongoing evaluation. Risk-prediction AI should not replace professional judgement; it must be interpreted by clinicians who understand the social, cultural, and environmental context of the population they serve.^5,11 There is a danger that AI could create an illusion of scientific precision particularly in legal or correctional settings where algorithmic outputs may appear objective and authoritative. Training AI models would therefore require human oversight to minimise bias; and ideally should be done with caution, over time. Ethical and practical concerns may then be able to be addressed as they arise and mitigated prior to more widespread adoption of such models.

AI models are dependent on high-quality data. Correctional health information is often fragmented across multiple systems, such as electronic health records, incident management databases, and paper-based records. Incomplete or inconsistent datasets can compromise model accuracy.^3,6 Integration across legacy platforms poses technical challenges, particularly when internet access is restricted for security reasons. ⁶

Infrastructure limitations, staff training, and resource constraints further complicate deployment. Queensland Corrective Services would need to invest in secure digital systems, workforce training, and data governance protocols to realise the potential of AI in practice. ⁶

At a policy level, Australia is developing national guidelines for high-risk AI applications, emphasising transparency, accountability, and public trust. ⁵ Correctional AI systems fall directly within this category given their potential impact on liberty, safety, and clinical care. Governance frameworks should mandate independent audits, explainable models, and ongoing human oversight.^5,11

To ensure AI improves rather than undermines care, implementation must be guided by rigorous validation, transparency, cultural competence, and sustained human oversight. Clinicians should interpret AI-generated risk scores, informed consent or assent should be sought where feasible, and mechanisms for challenge or appeal must exist. ¹¹ The potential contribution of artificial intelligence in prison psychiatry in Queensland and across Australian forensic services is unlikely to lie in the definitive prediction of adverse events. Its more plausible role may be in supporting clinicians to work more effectively with uncertainty—enhancing dynamic, preventative, and person-centred approaches to care. Rather than positioning AI as a technological solution to the inherent indeterminacy of forensic practice, it may be better understood as a limited but potentially valuable adjunct: one that can assist psychiatrists in interpreting complex patterns, while leaving clinical judgement, context, and ethical responsibility firmly in human hands.^6,11