The Toronto Declaration consists of 59 articles, broken into six sections, concerning
international human rights law, duties of
states, responsibilities of
private sector actors, and the
right to an effective remedy.
Preamble The document begins by asking the question, "In a world of machine learning systems, who will bear accountability for harming
human rights?" Governments should proactively create binding measures, and private entities should create internal policies, to protect against discrimination. Measures may include protections for sensitive data, especially for vulnerable populations. Systems should be designed in collaboration with a diverse community in order to prevent discrimination in design.
Duties of states: human rights obligations Governments today are deploying machine learning systems, often in collaboration with private entities. Even when development is contracted to such third parties, governments retain their obligation to protect human rights. Before implementation, and on an ongoing basis thereafter, they should identify risks and conduct regular audits, then take all necessary measures to mitigate these risks. They should be transparent about how machine learning is implemented and used, avoiding
black box systems whose logic cannot be easily explained. Systems should be subject to strict oversight from diverse internal committees and independent judicial authorities. Governments must also protect citizens from discrimination by private entities. In addition to oversight, they should pass binding laws against discrimination, as well as for data protection and privacy, and they should provide effective means to remedy for affected individuals. It is important for national and regional governments to expand on and contextualize international law.
Responsibilities of private sector actors: human rights due diligence Private entities are responsible for conducting "human rights
due diligence." Just like governments, private entities should identify risks before development by considering common risks and consulting stakeholders, "including affected groups, organizations that work on human rights, equality and discrimination, as well as independent human rights and machine learning experts." They should design systems that mitigate risks, subject systems to regular audits, and forego projects that carry too high of risks. They should be transparent about assumed risks, including details of the technical implementation where necessary, and should provide a mechanism for affected individuals to dispute any decisions that affect them.
The right to an effective remedy "The right to justice is a vital element of international human rights law." Private entities should create processes for affected individuals to seek remedy, and they should designate roles for who will oversee these processes. Governments must be especially cautious when deploying machine learning systems in the justice sector. Transparency, accountability, and remedy can help. == References ==