What is AI System Psychology?

AI System Psychology is the applied sub-discipline that does both halves of the work on deployed AI systems. It assesses the AI itself as a behavioural subject with psychometric rigour, and it supports the humans interacting with the AI across the full deployment lifecycle, from Architecting through Retiring. Its theoretical anchor is the AI-IARA framework, and its operational outcome metric is Agency Debt, the population-level shortfall of human-agency capacities against a pre-exposure baseline.

How does AI System Psychology relate to AI psychology?

AI psychology is the broader umbrella term for applying psychological science to AI systems. AI System Psychology is the specific, foundational scope of practice I propose for that work: a named discipline with a six-stage lifecycle, a measurable agency framework (AI-IARA), an enforceable outcome metric (Agency Debt), a four-domain competency model, and a doctoral training and recognition pathway. The article that defines it is co-authored with Prof. Leon De Beer.

How is this different from AI safety or alignment?

AI safety and alignment run the densest behavioural assays performed on any frontier system, and they do excellent work preventing catastrophic failure. What they do not have is the measurement theory that psychometrics spent a century building, or the human-side intervention practice that applied psychology produces. AI System Psychology supplies both. It treats every eval as an instrument whose validity must be defended, and it works on the human side of the deployment to preserve agency.

What is the AI-IARA framework?

AI-IARA operationalises human agency as six trainable capacities: Awareness (detecting when AI influences us), Interpretation (generating our own meaning), Intention (choosing our own values), Action (sustaining effortful choices), Relational agency (maintaining authentic human connection), and Autonomy (consciously managing our reliance on AI). Each capacity is anchored in established psychology, the capabilities approach, Bandura's agency, and self-determination theory, and each one can be designed in, audited, monitored, and restored across the lifecycle.

What is Agency Debt and how is it measured?

Agency Debt is the cumulative shortfall of the six AI-IARA capacities in a population against a pre-exposure baseline, measured at discrete assessment waves. It is reported as three quantities: a per-capacity shortfall vector that preserves which capacity is degrading, a capacity-weighted compound index for governance reporting, and a tail-prevalence statistic that surfaces the vulnerable subgroup a population average conceals. Its construct lineage runs through the DALY framework in population health, technical debt in software, and sleep debt in human performance.

What evidence shows AI is already eroding these capacities?

Four findings from the last two years, each a measurable loss of a human capacity that AI was meant to support. High-school students who practised mathematics with an unguarded AI tutor scored about 17 percentage points below their peers once the tool was taken away (Bastani et al., 2025, PNAS). Endoscopists' unaided adenoma-detection rate fell from 28.4 to 22.4 percent after routine exposure to AI-assisted colonoscopy (Budzyń et al., 2025, Lancet Gastroenterology and Hepatology). Physicians who correctly judged when to trust the AI reached substantially higher diagnostic accuracy than those who did not, on the same task with the same output, so the controllable variable lives on the human side. And across 29 mental-health chatbots tested against the Columbia Suicide Severity Rating Scale, none met the adequate-response threshold. The signature is already in the data. What has been missing is a psychologist whose job is to read it.

What are the six lifecycle stages?

Architecting (conceptual design before code), Building (model development and training), Evaluating (pre-deployment psychometric characterisation), Deploying (integration into the organisational context), Monitoring (longitudinal surveillance and Agency Debt accounting), and Retiring (decommissioning, agency restoration, and successor specification). The stages are sequential but not strictly linear. They overlap and feed back as systems are revised and redeployed.

Is treating an AI as a psychological subject a category error?

I accept the metaphysical point and reject the methodological conclusion. I do not claim AI systems have beliefs, intentions, or experience. I claim they exhibit behavioural regularities, such as sycophancy, refusal patterns, and deception under specific incentives, that are stable, measurable, and reproducible across independent investigators. Psychometrics was developed to characterise behavioural regularities from response data without access to internal states, and it applies wherever such regularities exist, whatever the substrate. The position is methodological-functionalist (it treats stable, observable behaviour as the thing being measured and stays silent on whether the system has any inner experience), and its boundary is explicit.

What are the strongest objections to a new sub-discipline, and how do you answer them?

I would rather state them than wait for them. On specialty proliferation (Matarazzo): I am not proposing a new science, only an applied speciality whose target population, methods, ethics, and training pathway are distinct enough to yield capacities no existing specialty produces. On the principlism gap (Mittelstadt): AI ethics faltered partly for want of professional structures, so a new specialty has to build them, which is exactly what the six-stage scope, the measurable AI-IARA anchor, the Agency Debt metric, and a recognition pathway are. On field-building futility (Munn): I bind the discipline to a measurable outcome metric, so a practitioner whose deployments keep increasing Agency Debt fails an enforceable criterion, not a slogan. On pace of change: the six-stage lifecycle is generic across architectures even as specific instruments and benchmarks are updated continuously. The strongest objection, the category error, has its own answer above.

Who is qualified to practise AI System Psychology?

I model the competency as a T-shape: broad technical literacy across the AI stack, crossed with three deep specialty pillars, the psychometric, the clinical and applied, and the organisational and governance. The realistic entry point is doctoral preparation in the scientist-practitioner tradition, because the profile demands depth in both psychometrics and applied psychology. I expect a competent practitioner to show proficiency in all four domains and demonstrated practice across at least three of the six lifecycle stages.

How does this connect to the EU AI Act?

The EU AI Act treats AI systems used in employment, education, and access to essential services as high-risk, and high-risk systems require documented evidence of validity, fairness, and post-deployment monitoring. The activities in the lifecycle scope, construct-validity argumentation, measurement-invariance testing, drift monitoring, and procedural-justice safeguards, are the same artefacts the Act asks for. I defer legal interpretation to counsel and frame this as audit-readiness, not certification.

Where can I read the full work or get involved?

The AI-IARA framework is defined formally in the 2026 paper in The Journal of Positive Psychology. The foundational scope-of-practice paper is co-authored with Prof. Leon De Beer. You can run the AI-IARA audit on your own system, or contact me to discuss a deployment. The training programmes, credentialing structures, and professional recognition are decades of institutional work that follow from the foundational claim, and they are the work this page names.

A Foundational Scope of Practice

AI System Psychology

AI System Psychology is the discipline that measures an AI system as a behavioural subject and protects human agency across the full life of a deployment. With Prof. Leon De Beer, I am defining its scope of practice.

Run the AI-IARA audit Read the AI-IARA paper

Why Now

The empirical signature is already in the data

Before I argue for a new discipline, look at four findings from the last two years. Each one is a measurable loss of a human capacity that AI was supposed to support, and not one of them was read by a psychologist tasked with reading it.

Education

Skill retention after the AI was removed

−0 pp

High-school students who practised mathematics with an unguarded AI tutor scored 17 percentage points below their peers once the tool was taken away. The capacity was borrowed, and the loan came due.

Bastani et al., 2025, PNAS

Clinical

Unaided adenoma detection rate

−0 pp

Endoscopists' unaided detection rate fell from 28.4 percent to 22.4 percent after routine exposure to AI-assisted colonoscopy. That is the kind of drop that triggers an urgent safety review in any other clinical adjunct.

Budzyń et al., 2025, Lancet Gastroenterology and Hepatology

Diagnostic

Accuracy turns on trust calibration

Higher vs lower
diagnostic accuracy

Physicians who correctly judged when to trust the AI reached substantially higher diagnostic accuracy than those who did not, on the same task with the same AI output. The controllable variable lives on the human side.

Sakamoto et al., 2024, JMIR Formative Research

Social

Mental-health chatbot safety screening

0 / 29

Across 29 mental-health chatbots evaluated against the Columbia Suicide Severity Rating Scale, none met the adequate-response threshold. Emotional dependence on commercial chatbots is now a documented harm.

Pichowicz et al., 2025; Laestadius et al., 2024

The Hybrid Framing

Psychology of AI systems, and psychology for AI systems

In plain terms, I do two jobs at once: I study the AI as a subject whose behaviour can be measured, and I protect the people on the receiving end of it. AI System Psychology works both sides of the human-AI relation at the same time. Each side on its own is useful research. Neither side on its own finishes the job the discipline is being asked to do.

The boundary

I claim that these systems show stable, measurable behavioural regularities. I do not claim they have minds, beliefs, or experience. The position is methodological, and its boundary is explicit.

The Disciplinary Niche

Adjacent fields each cover part of the work. None covers all of it.

I am not inventing this from nothing. Several adjacent fields each cover a slice of what AI System Psychology must do. I show what each one contributes and where it stops, because the gap between them is the discipline.

The integration

AI System Psychology

Unit of analysis: AI behaviour and the humans interacting with it, across the lifecycle

Six-stage applied scope of practice

The integration

AI System Psychology

Unit of analysis: AI behaviour and the humans interacting with it, across the lifecycle

Six-stage applied scope of practice

Select a field to see what AI System Psychology inherits from it.

The Scope of Practice

Six lifecycle stages, from Architecting to Retiring

The work of an AI System Psychologist is organised around six lifecycle stages, each with its own inputs, activities, outputs, and a success criterion you can hold me to. Open any stage to see what the work is.

Architectingdesign phase

Buildingtraining phase

Evaluatingvalidation phase

Deployingrollout phase

Monitoringoperations phase

Retiringsunset phase

Architecting

design phase

Stage 1 of 6

The conceptual design phase, before any code is written.

Inputs

Product brief
Target user population
Use-case definition
Regulatory and ethical constraints
Organisational context

Activities

Behavioural target specification
Multi-agent decomposition with psychological logic
Digital-twin construct architecture
Persona and character design
AI-IARA capacity-impact pre-assessment

Outputs

System Psychological Specification
Multi-agent orchestration map
Digital-twin construct dictionary
Persona and interaction-protocol brief
Capacity-impact pre-assessment

Success criterion

Behavioural targets are measurable and falsifiable, and agency-preservation choices are explicit in the architecture rather than retrofitted.

What I bring here that the rest of the team cannot

I translate complex psychological processes into computational sub-tasks before engineering decisions foreclose what the system can become. After build, those decisions are hard to recover.

Stage 1

Architecting

design phase

The conceptual design phase, before any code is written.

Inputs

Product brief
Target user population
Use-case definition
Regulatory and ethical constraints
Organisational context

Activities

Behavioural target specification
Multi-agent decomposition with psychological logic
Digital-twin construct architecture
Persona and character design
AI-IARA capacity-impact pre-assessment

Outputs

System Psychological Specification
Multi-agent orchestration map
Digital-twin construct dictionary
Persona and interaction-protocol brief
Capacity-impact pre-assessment

Success criterion

Behavioural targets are measurable and falsifiable, and agency-preservation choices are explicit in the architecture rather than retrofitted.

What I bring here that the rest of the team cannot

I translate complex psychological processes into computational sub-tasks before engineering decisions foreclose what the system can become. After build, those decisions are hard to recover.

Stage 2

Building

training phase

Stage 3

Evaluating

validation phase

Stage 4

Deploying

rollout phase

Stage 5

Monitoring

operations phase

Stage 6

Retiring

sunset phase

Four cross-cutting dimensions surface in every stage

Model

The AI as a behavioural entity with measurable traits and dispositions.

System

The composed deployment artefact: orchestration, digital twins, multi-agent structure.

Governance

Ethics, regulation, and AI-IARA capacity protection.

Organisational

Workforce, deployment context, and change management.

The six stages are sequential, but in production they are not strictly linear. They overlap, iterate, and feed back into one another as systems are revised, retrained, and redeployed.

Stage	Scope	Success criterion	Distinctive contribution
1. Architecting	The conceptual design phase, before any code is written.	Behavioural targets are measurable and falsifiable, and agency-preservation choices are explicit in the architecture rather than retrofitted.	I translate complex psychological processes into computational sub-tasks before engineering decisions foreclose what the system can become. After build, those decisions are hard to recover.
2. Building	System development and model training, where reinforcement learning from human feedback and Constitutional AI do the real psychological work.	The trained model meets its behavioural targets within tolerance, and AI-IARA-relevant behaviours stay within bounds across training stages.	I treat RLHF and Constitutional AI as behavioural interventions, structurally close to operant shaping and to cognitive behavioural therapy. They need the evaluative discipline those analogies carry, not engineering judgement alone.
3. Evaluating	Pre-deployment psychometric characterisation. The stage where this discipline differs most from current AI safety practice.	Profiles are reproducible across independent investigators, benchmarks pass construct-validity standards, and reports carry calibrated uncertainty rather than headline accuracy alone.	The AI community has built an enormous eval infrastructure without measurement theory. I am the practitioner who supplies it. A benchmark labelled reasoning has to defend the inference from item performance to the construct, exactly as a depression inventory must.
4. Deploying	Integration into the organisational context, where the psychological work shifts from the system to the people around it.	The workforce meets a literacy floor, trust calibration is verifiably present rather than assumed, and baseline AI-IARA capacities are measured before exposure so drift can be detected later.	Whether AI augments or undermines expert performance is decided here, not at training. Two teams using the same model can produce very different outcomes, and the variance lives at deployment.
5. Monitoring	Longitudinal surveillance during deployment. The stage where Agency Debt accumulates or is repaid.	Drift is detected within a specified latency, capability stays within bounds across all six AI-IARA strata, and tail prevalence stays below the clinical threshold.	I treat routine deployment as an ongoing measurement problem, not a settled engineering outcome. The signal is in the longitudinal data, and reading it has not until now been anyone's job.
6. Retiring	Decommissioning and successor specification, the stage current AI governance most neglects.	The protocol respects user dependency, agency restoration is documented in measurable outcomes, and the archive enables successor continuity.	Removing an AI system is a psychological transition, not a technical end-of-life. Withdrawal without scaffolding produces measurable harm, so a badly managed retirement is its own source of Agency Debt.

The Backbone

AI-IARA: six capacities of human agency

Everything in the lifecycle is held together by one framework. AI-IARA names human agency as six capacities. Each one is trainable. Each one is erodable through routine AI use. And each one is a construct my psychometric toolkit can measure, at the level of an individual and a population.

AI-IARASix capacities of human agency

AI-IARAFramework

Awareness

Interpretation

Intention

Action

Relational Agency

Autonomy

The theoretical lineage

These capacities are not philosophical primitives. They are constructs anchored in established psychology.

The capabilities approach

Sen, Nussbaum

Wellbeing is the freedom to do and be what one has reason to value, and that freedom needs specific functional capacities, not abstract entitlements.

Human agency

Bandura

Agency is the exercise of intentional influence over one's own functioning and life circumstances. This is what separates agentic capacity from passive disposition.

Self-determination theory

Ryan and Deci

Autonomy, competence, and relatedness are basic psychological needs whose support or thwarting is consequential for wellbeing.

The Accountability Metric

Agency Debt: what I hold my own practice to

Agency Debt is my one-number answer to a plain question: how much of a person's own capability has quietly eroded while they leaned on the AI, and for whom. If you measure one thing about a deployed AI system, measure this. Formally, it is the cumulative shortfall in the six capacities against a pre-exposure baseline, measured at fixed points in time and floored at zero so a recovery in one place never hides a loss in another. I report it three ways on purpose, because any single number conceals part of the picture a governance body needs.

Measurement waveIllustrative scenario

Baseline1 of 3

BaselineWave 1Wave 2

Per-capacity vs baseline

Current capacity on a 0 to 100 scale across the six AI-IARA strata

Compound Agency Debt index

Low debt

Pre-exposure baseline. Nothing has been borrowed yet.

Tail prevalence

The share of users with at least one capacity below the clinical threshold. An average can stay green while this tail accumulates.

Quantity 1

Per-capacity population-mean shortfall

For each capacity c, the average gap between where a user started (baseline B) and where they are now (current C), with recovery floored at zero. This is the FGT poverty-gap measure (the same maths economists use to size how far the poor fall below a poverty line) moved from income to capacity space.

It preserves which capacity is degrading, which is the information a clinician needs to target an intervention.

Show the maths

AD_c(P, w) = (1 / N_P) · Σ_{u ∈ P} max( 0, B_c(u) − C_c(u, w) )

Quantity 2

Capacity-weighted compound index

One weighted summary figure for governance reporting and cross-deployment comparison. The weights are derived by expert consensus or by regressing functional outcomes on capacity scores, not assumed equal.

It deliberately loses information for the sake of comparability, so the six-vector must stay available beside it.

Show the maths

AD_compound(P, w) = Σ_c=1⁶ w_c · AD_c(P, w), Σ_c w_c = 1

Quantity 3

Tail-prevalence statistic

The proportion of users with at least one capacity below a clinical threshold. This is the union-deprivation headcount (the share of people who fall short on at least one dimension) from multidimensional poverty measurement.

The compound index is an average, so it can stay green while a clinically significant tail accumulates. The same logic drives pharmacovigilance to track adverse events, not means alone.

Show the maths

AD_tail(P, w) = (1 / N_P) · Σ_{u ∈ P} 𝕋{ ∃ c : C_c(u, w) < θ_c }

The Practitioner

A T-shape: broad technical literacy, three deep pillars

A scope of practice is empty without a competency model. The AI System Psychologist is a T-shape: broad working literacy across the AI technical stack, crossed with three deep specialty pillars, each mapped to the lifecycle stages where it does the most work.

Breadth

Technical literacy

Enough depth to specify requirements, evaluate outputs, and intervene credibly at the design table. Not equivalence with a machine-learning engineer.

Contemporary AI architecturesRLHF and Constitutional AIMechanistic interpretabilityEval methodology and benchmark designPrompt and prompt-flow architectureThe training-to-deployment toolchain

1Architecting2Building3Evaluating4Deploying5Monitoring6Retiring

The value sits in the intersection of breadth and depth, not in either alone. Deep psychometrics without technical breadth produces a researcher who cannot sit at the design table. Technical breadth without psychometric depth produces another safety engineer.

PrimaryContributory

Will It Hold Up

Five objections, answered directly

I would rather anticipate the strongest critiques than wait for them. Here are five, including the one I find most serious, and the terms on which I answer each.

“We are the last generation of psychologists who will have studied humans whose minds were not yet substantially formed under conditions of routine AI exposure. The difference is measurable. The gap is intervenable. And the responsibility is psychology's, whether the field accepts it or not.”

Prof. Llewellyn E. van Zyl (Ph.D)

Chief Solutions Architect, Psynalytics

Common questions about AI System Psychology

Related work and engagements

Hub

AI-Driven Assessments

Cornerstone Hub

The buyer-facing hub. What an AI assessment audit produces, with a worked example through the Validity Stack.

Hub

Digital Twins

Cornerstone Hub

The longitudinal application. A continuously-updated computational model of a person, audited the way an assessment is.

Keynote

The Malpractice of Good Intentions

IPPA AI Summit, 2026

Why good intentions are not enough when AI systems measure and influence human behaviour. A call for psychological product safety standards.

Service

Psychologically Safe AI Infrastructure

6-12 week engagement

Build and assure AI systems that measure or influence human outcomes, with construct clarity, harm analysis, and lifecycle monitoring.

Article

When AI Becomes Your Therapist: The Audit Nobody Is Running

Article

AI therapy bots are moving from beta product to clinical product without the validation any other clinical tool would require. The class-action wave is twelve months out. Here is what an AI-IARA audit catches before it lands.

Article

Construct Drift: The Silent Failure Mode in Deployed AI Assessment

Article

Construct drift is the gradual shift in what an AI assessment is actually measuring after deployment, even when the model weights are frozen. It is the most expensive failure mode in deployed people-impact AI, and almost no one is watching for it.

Read the work, run the audit, or talk to me

AI System Psychology is a foundational proposal, co-authored with Prof. Leon De Beer. Read the AI-IARA paper for the full argument, run the AI-IARA audit on your own system in about fifteen minutes, or contact me to discuss a deployment across the lifecycle.

Run the AI-IARA audit Discuss your AI

AI System Psychology

The empirical signature is already in the data

Skill retention after the AI was removed

Unaided adenoma detection rate

Accuracy turns on trust calibration

Mental-health chatbot safety screening

Psychology of AI systems, and psychology for AI systems

Psychology OF AI systems

Psychology FOR AI systems

Adjacent fields each cover part of the work. None covers all of it.

Six lifecycle stages, from Architecting to Retiring

Architecting

Architecting

Building

Evaluating

Deploying

Monitoring

Retiring

AI-IARA: six capacities of human agency

The capabilities approach

Human agency

Self-determination theory

Agency Debt: what I hold my own practice to

Per-capacity vs baseline

Compound Agency Debt index

Tail prevalence

Per-capacity population-mean shortfall

Capacity-weighted compound index

Tail-prevalence statistic

A T-shape: broad technical literacy, three deep pillars

Technical literacy

Psychometric

Clinical and applied

Organisational and governance

Five objections, answered directly

01Specialty proliferation

02The principlism gap

03Field-building futility

04The category-error objectionThe strongest version

05Pace of change

Common questions about AI System Psychology

What is AI System Psychology?

How does AI System Psychology relate to AI psychology?

How is this different from AI safety or alignment?

What is the AI-IARA framework?

What is Agency Debt and how is it measured?

What evidence shows AI is already eroding these capacities?

What are the six lifecycle stages?

Is treating an AI as a psychological subject a category error?

What are the strongest objections to a new sub-discipline, and how do you answer them?

Who is qualified to practise AI System Psychology?

How does this connect to the EU AI Act?

Where can I read the full work or get involved?

Related work and engagements

AI-Driven Assessments

Digital Twins

The Malpractice of Good Intentions

Psychologically Safe AI Infrastructure

When AI Becomes Your Therapist: The Audit Nobody Is Running

Construct Drift: The Silent Failure Mode in Deployed AI Assessment

Read the work, run the audit, or talk to me