A broader frame for ethics in computer vision
The authors of Beyond Fairness in Computer Vision (Gebru and Denton 2024) observe that much of the current debate around machine learning and computer vision centers on fairness metrics: ensuring that models do not discriminate across demographic groups or perpetuate historical bias. While such efforts are important they warn that they are insufficient. Building a “fair” model does not guarantee that the broader social and systemic harms embedded in research practices are addressed. In particular they highlight three core limitations: data collection and annotation practices rarely capture the lived realities of affected communities; research is often done without consulting those communities; and downstream uses of vision systems often entail social risks (surveillance, privacy invasion, misrepresentation) which remain ungoverned.
They propose instead a holistic approach that moves beyond individual models to the entire research lifecycle: from data gathering to deployment and governance. This includes participatory design methods, community-rooted annotation processes, transparent documentation of datasets and model limitations, plus continuous reflexivity about power imbalances and social contexts. Such shift would help ensure that vision systems serve social aims defined by and with communities rather than external stakeholders alone.
Why this matters for digital sustainability
Within the framework of Digital Sustainability Hub (DSH) at IULM University the mission explicitly embraces not only technological efficiency but also social, cultural and environmental responsibility in digital infrastructures. The holistic vision proposed by Gebru and Denton aligns deeply with this mission. It reminds us that sustainable digital innovation cannot be reduced to energy efficient systems or low-impact infrastructure. True digital sustainability must include ethical and governance dimensions ensuring that technologies respect human dignity, foster inclusion, and avoid reproducing existing social inequalities.
Moreover, adopting community-centered computer vision research practices can contribute to more just, inclusive and context-sensitive digital ecosystems. Rather than imposing one-size-fits-all models trained on biased data, systems could be tailored to specific communities, with their participation and consent. This can reduce risks of misrepresentation and harm while increasing relevance and fairness in real-world applications such as social services, health, public space surveillance, or cultural heritage.
Key Recommendations
The authors suggest concrete practices to reshape the research culture of computer vision:
Use participatory data collection and annotation involving members of the communities whose images will appear in datasets.
Maintain transparency and documentation (dataset cards, model cards, clear record of assumptions, limitations, and contexts).
Foster accountability and governance mechanisms; institutions and research labs must acknowledge potential harms, allow external audits, and engage with affected stakeholders.
Prioritize social value and justice over performance metrics; rather than optimizing only for accuracy, models should be evaluated also on ethical, social and community impact.
Implications for Researchers, Practitioners and Policy-makers
For researchers this implies a paradigm shift: from viewing models as isolated technical artifacts to seeing them as socio-technical systems embedded in broader power structures. Data ethics becomes not a compliance add-on but a central component of research design. For practitioners and companies deploying vision systems this approach encourages responsible innovation, where deployment is preceded by social impact assessment, community consultation, and ongoing monitoring. For policy-makers and institutions this work underscores the need for governance frameworks that regulate not only outputs but research processes, data practices, transparency and inclusion.
Conclusion
This holistic ethic means advocating for funding schemes that require participatory methods, supporting collaborations between communities, civil society and researchers, and promoting transparency through open documentation. It also means engaging educational programmes to raise awareness among future practitioners about power, consent and social justice in data and machine learning.
The shift from fairness to holistic responsibility in computer vision is not only timely but urgent. As vision systems increasingly permeate public and private spheres, their impacts become more pervasive. The work by Gebru and Denton provides a critical roadmap to ensure such systems do not reproduce harm, but instead support community agency, transparency and justice. For digital sustainability defenders, including the DSH community, it offers an invitation to widen our focus: not only on green computing or efficient infrastructure but on the design, governance and social meaning of technology.
Full Paper:
acm.org