Community and equity AI governance documents nine research-backed concepts that explain how AI adoption affects the communities organisations exist to serve. From encoded bias and epistemic justice to proxy discrimination, digital redlining, and participatory exclusion, these entries show how AI systems inherit and amplify existing inequity and what not-for-profits can do about it.
Community and equity AI governance for community organisations and not-for-profits
Community and Equity
Whether AI adoption compounds or corrects existing disadvantage, and the mechanisms that determine which outcome occurs.
How community and equity risks shape AI governance in nonprofits and community sector organisations
Community and equity AI governance documents nine research-backed concepts that explain how AI adoption affects the communities organisations exist to serve. From encoded bias and epistemic justice to proxy discrimination, digital redlining, and participatory exclusion, these entries show how AI systems inherit and amplify existing inequity and what not-for-profits can do about it.
About this theme
AI does not affect everyone equally
The communities most likely to be harmed by poor AI governance are the same communities that not-for-profit organisations exist to serve. That is not a coincidence. It is a structural consequence of how AI systems are built and on whose data they are trained.
The research behind this theme comes from philosophy, organisational sociology, science and technology studies, and the empirical literature on algorithmic systems. None of it was written to describe not-for-profit organisations specifically. All of it describes, with precision, the conditions that make AI adoption a community equity risk for any organisation accountable to people who are already underserved by mainstream systems.
What the research documents consistently is that AI does not introduce inequity into neutral systems. It inherits and amplifies inequity already present in data, institutions, and assumptions. Sampling bias provides the underlying mechanism: when the people whose data trains an AI system are not representative of the people the system will be used to assess, the system produces less accurate outputs for those it knows least. The communities already marginalised by mainstream data collection get less useful AI, less accurate AI, and less recourse when AI outputs are wrong. Negative space operates alongside this: what AI excludes shapes the meaning of what it includes, and absence is rarely labelled as such.
Research note: Encoded Bias
A 2019 study published in Science found that a widely used US hospital algorithm for allocating healthcare resources systematically assigned lower risk scores to Black patients than to white patients with equivalent clinical conditions. The algorithm used healthcare cost as a proxy for health need. Because Black patients had historically had less spent on their care due to systemic inequity, the algorithm learned that they were healthier. The bias was structural, invisible within the system, and dressed in the language of objectivity.
For not-for-profits and community organisations, this theme sits at the most direct intersection of mission and risk. An organisation that adopts AI without examining these patterns is not being neutral. It is, however unintentionally, placing operational efficiency above the communities it exists to serve. Framing Theory adds a further dimension: the frame embedded in an AI system shapes what questions can be asked and what answers are possible, and the frame is almost never visible to the people using the output.
The distinction between absence of evidence and evidence of absence matters here too. When an AI system produces no output about a community, or produces thin, generic output, that is not the same as evidence that the community has no relevant characteristics. It is evidence of a gap in the training data. Organisations that treat AI silence as a form of completeness risk making decisions on the basis of what AI does not know, without recognising that is what they are doing.
Research note: Algorithmic Bias
MIT Media Lab researcher Joy Buolamwini found that commercial facial recognition systems from major technology companies had error rates of up to 34% for darker-skinned women, compared to under 1% for lighter-skinned men. The systems had been trained predominantly on lighter-skinned faces. Buolamwini discovered the disparity when she had to wear a white mask for a system to detect her face at all. The people most misidentified were the people least represented in the training data. The system was not malfunctioning. It was performing exactly as its training had prepared it to.
Further theoretical grounding
These concepts inform this theme without warranting standalone entries. They are part of the research foundation this theme draws on.
When training data does not represent the people a system will be applied to, the system produces less accurate outputs for those it knows least. The underrepresentation is often invisible inside the system itself.
What AI excludes shapes the meaning of what it includes. Absence in AI output is rarely labelled as absence. It arrives looking like comprehensiveness.
When AI produces no output about a community, that is not evidence the community has no relevant characteristics. It is a data gap. Treating AI silence as comprehensiveness produces decisions grounded in what AI does not know.
The frame embedded in an AI system shapes what questions can be asked and what answers are possible. The frame belongs to whoever built the system, and it is almost never visible to the people using the output.
Multiple valid knowledge frameworks exist simultaneously. AI systems trained on one framework treat others as gaps or noise. The knowledge of communities not centred in mainstream data collection is structurally excluded before any individual decision is made.
You may also be interested in
Algorithmic Justice League
Documents the human impact of algorithmic bias in AI systems across healthcare, criminal justice, and community services
Australian Human Rights Commission: Technology and Human Rights
The Australian framework for understanding human rights obligations when adopting AI and automated decision-making systems
NIAA: Framework for Governance of Indigenous Data
The Australian government framework for Indigenous data sovereignty and governance, directly relevant to AI adoption in community organisations serving First Nations communities
Community and equity AI governance resources for the Australian community sector
Community and equity AI governance documents nine research-backed concepts that explain how AI adoption affects the communities organisations exist to serve. From encoded bias and epistemic justice to proxy discrimination, digital redlining, and participatory exclusion, these entries show how AI systems inherit and amplify existing inequity and what not-for-profits can do about it.
Nine entries
The research behind the theme
Each entry takes an established concept and connects it directly to AI adoption in not-for-profits and community organisations. What the research shows, how it appears in practice, and what to watch for.
Social License to Operate
Organisations need more than legal permission to operate. They need ongoing trust from the communities they serve. AI adoption can erode that trust quietly and quickly, in ways that do not show up in any metric until it is already gone.
Read entry Miranda Fricker, 2007Epistemic Justice
Whose knowledge counts as valid and whose gets systematically discounted. AI systems trained on mainstream data inherit existing patterns of epistemic injustice, producing less accurate and less useful outputs for communities already marginalised.
Read entry Daston and Galison, 2007Mechanical Objectivity
The belief that removing human judgement produces unbiased results. In practice, AI embeds the values and assumptions of its designers invisibly, and computational outputs feel more authoritative than human ones, making them harder to challenge.
Read entry Obermeyer, Buolamwini and othersEncoded Bias
Training data encodes existing inequities. AI systems do not inherit neutral data. They inherit human history. The communities least represented in training data receive the least accurate outputs, and the bias is invisible inside systems that present themselves as objective.
Read entry Barocas and Selbst, 2016Proxy Discrimination
AI systems can discriminate against protected groups without using protected characteristics directly. Postcode, device type, browsing history, and service use patterns all function as proxies for race, class, and disability. The discrimination is real and its mechanism is invisible.
Read entry Coates, 2014 and Eubanks, 2018Digital Redlining
The practice of systematically providing inferior digital services, access, or outcomes to communities based on geography, demographics, or socioeconomic status. AI systems that replicate or deepen service access inequities perform digital redlining whether or not that was the intent.
Read entry Chouldechova, 2017 and Corbett-Davies et al, 2017Algorithmic Fairness Illusion
Mathematical definitions of fairness are mutually incompatible. A system optimised for one fairness criterion will violate others. When AI vendors claim their systems are fair, the claim describes one mathematical definition, not an absence of harm to specific communities.
Read entry Lyon, 2001 and Zuboff, 2019Surveillance Creep
Data collected for one purpose gradually becomes available for others. AI tools adopted for service delivery create data trails that can be repurposed for performance management, funder reporting, or third-party access in ways that were not anticipated when the tools were adopted.
Read entry Arnstein, 1969 and Eubanks, 2018Participatory Exclusion
The communities most affected by AI-assisted decisions are the least likely to be involved in designing or governing those systems. Participation frameworks that do not actively address power and access reproduce the exclusions they claim to remedy.
Read entryAll themes
Continue exploring the lexicon
From theory to practice
The lexicon is the foundation. Our resources are what you build with it.
Understanding why these behavioural patterns occur is the first step. Not For Humans resources give your organisation the practical infrastructure to identify which patterns are active, name them with shared language, and build governance commitments that hold under pressure.
Not For Humans: Because AI’s greatest risk isn’t technical, it’s Human.