AI and human behaviour governance risk in community organisations
Where the risk actually lives
They asked for judgement. What came back was an opinion.
“I have sat in conversations, across years and across different organisations, where senior people described putting sensitive strategic documents into ChatGPT for review. Material that should never have left the building.
What came back, every time, was the AI’s opinion of the document. Not the judgement of the colleague they thought they were asking. Two things leave the room in that single action: information that should never have gone, and the human judgement that was the entire reason for asking.”
Adrian Hubble, Founder, Not For Humans
Nothing in those moments felt like a governance failure. No policy was consulted, because nobody thought they were doing anything that needed one.
AI does not change what people value. It changes the conditions in which they act on those values.
Nobody in those rooms was careless or indifferent. They were under pressure. The tool was to hand. It had been useful before. Nothing in the moment said stop.
Photo: Vitaly Gariev / Unsplash
Not For Humans maps the documented behavioural patterns that cause AI and human behaviour governance to fail in community organisations. Grounded in peer-reviewed research across six themes, the Not For Humans framework gives community organisations the language and tools to govern AI at the human level, before incidents occur.
The Not For Humans AI Governance Model
Four layers. Most governance reaches one of them.
Most organisations that take AI seriously produce a policy. Some produce a register. A few run an audit. All useful. All at the same layer.
Not For Humans, 2026. Read outward in: most governance frameworks stop at Layer 4. The risk lives at Layer 1.
Organisational Commitment.
Policy, register, agreed parameters, roles, review cycles. Necessary, and not sufficient alone.
Operational Practice and Visibility.
What is actually in use, including informal and embedded tools. Whether outputs are verified before being acted on.
Individual Behaviour and Judgement.
How people respond to AI outputs. Over-trust, deferred accountability, the quiet substitution of AI output for human judgement. Whether anyone feels safe raising it.
The Human/AI Interface Moment.
One person, one tool, one output, one decision about what to do next. No policy reaches this moment.
The gap between Layer 4 and Layer 1 is not a theory. It is measurable, and the numbers are not close.
Australian office professionals who used AI at work despite believing it was against company policy
Employees who say leadership expectations around AI use are clear, against 87% already using it daily
Australian and New Zealand employees using AI at work in 2023, against organisations with a formal policy in place at the time
The four layers sit on a foundation that comes before all of them: human and community-centred values. This sector has no shared profit motive, no shared structure, no shared size. The one thing every for-purpose organisation holds in common is a mission to serve a community. That is why the foundation of this model is values rather than compliance. It is the only variable the whole sector shares.
Where risk actually lives
Fifty-five documented theories. Forty-five named risks. Six themes. All of them human.
Not For Humans maps 55 documented human behaviour theories across six themes, grounded in peer-reviewed research from psychology, sociology, and organisational behaviour. From those foundations we identify 45 named behavioural risk patterns that emerge when AI enters community organisations.
Each one arises from the conditions humans operate in, not from technology failure. Hover or tap each theme to explore.
Trust and Cognition
AI outputs are received as relief, not scrutinised as risk.
Time pressure and fluency combine to replace verification with acceptance. Automation bias, the fluency-accuracy illusion, and coherence effects shape what gets questioned and what gets waved through.
Explore this theme →Governance and Accountability
Policies exist. Accountability does not.
Responsibility becomes assumed rather than assigned. Diffusion of responsibility, governance theatre, and temporal discounting mean that when something goes wrong, no one is clearly accountable, and no one planned it that way.
Explore this theme →Autonomy and Judgement
Human decision-making authority erodes gradually.
What begins as AI-assisted becomes AI-led, without anyone deciding that is what should happen. Cognitive offloading, learned helplessness, and skill erosion are the mechanisms. The loss is quiet and cumulative.
Explore this theme →Community and Equity
AI trained on majority data fails the communities your organisation exists to serve.
Marginalisation is compounded rather than corrected. Encoded bias, epistemic injustice, and assumed data representation mean the people most dependent on your organisation may be the least visible to the tools you use to serve them.
Explore this theme →Culture and Psychological Safety
Unspoken norms shape how staff behave, regardless of what any policy says.
Stigma around AI use, status threat, and pressure dynamics create the real governance environment. What people feel safe to say, question, or refuse matters more than what the policy permits.
Explore this theme →Data and Vendor Dynamics
What goes into AI systems, and what those systems do with it, shifts over time.
For organisations holding sensitive community data, this is non-negotiable. Consent drift, vendor lock-in, and unintended IP transfer are patterns already emerging in community sector AI adoption.
Explore this theme →AI and human behaviour governance risk in community organisations
What good looks like
Seven conditions under which human judgement survives contact with AI.
These are not specific to any organisation, tool, or framework. They are the conditions under which human judgement is preserved in any AI-assisted environment, regardless of what is in use or how mature the governance is.
Human judgement is non-negotiable.
AI assists decisions. It does not make them. A named person is accountable for every significant AI-assisted output. That accountability cannot be delegated to a system.
Governance must reach the interface.
A framework that does not address individual behaviour at the point of AI use is incomplete. The outer layers matter. They are not sufficient alone.
Transparency is a practice, not a declaration.
Transparency exists when staff habitually disclose AI involvement and when that disclosure is safe and normal. Publishing a policy does not make an organisation transparent.
Fluency is not accuracy.
Well-written, confident AI output is not more likely to be correct. Presentation quality must never substitute for verification of the underlying claims.
Community accountability comes before compliance.
Governance exists to protect the relationship between an organisation and the communities it serves. Regulatory compliance is a floor. Community safety is the standard.
Psychological safety enables real governance.
Frameworks only work if people feel safe enough to use them honestly. Where staff hide their AI use or fear raising concerns, governance exists in name only.
Adaptation is a governance requirement.
AI changes faster than any fixed governance cycle. The ability to notice what is changing, assess it, and respond deliberately is itself a governance capability.
These principles are drawn from the Not For Humans AI Governance Model, which sets out the four layers, the foundation, and what good practice looks like at each layer.
Not For Humans maps the documented behavioural patterns that cause AI and human behaviour governance to fail in community organisations. Grounded in peer-reviewed research across six themes, the Not For Humans framework gives community organisations the language and tools to govern AI at the human level, before incidents occur.
Why this cannot wait
By the time a behavioural shift is visible as an incident, the pattern is already embedded.
Prevention has passed by then. The organisations that manage this well are not the ones with the longest policy. They are the ones whose people already know what to do in the moment the paperwork cannot see.
The Response
Capability at the individual and team level is what makes governance work. This is the order we build in.
Our Approach →Not For Humans: Because AI’s greatest risk isn’t technical, it’s Human.
- PagerDuty Shadow AI Survey (2026). Wakefield Research, 1,250 office professionals, Australia, Japan, UK, US.
- ELMO Software, Employee Sentiment Index (2026). YouGov, 1,055 Australian employees.
- ISACA (2023). Australia and New Zealand shadow AI research.
- Employment Hero, AI Paradox at Work (2026). Focaldata, 1,634 Australian workers and 1,008 business leaders.