The Toronto Declaration: Protecting the rights to equality and non-discrimination in machine learning systems

The Toronto Declaration: Protecting the rights to equality and non-discrimination in machine learning systems is the 16 May 2018 joint civil-society manifesto drafted by Amnesty International and Access Now and endorsed at launch by Human Rights Watch and the Wikimedia Foundation, which proposes that the existing framework of international human rights law — and specifically the right to equality and non-discrimination — be applied as the load-bearing register for naming, preventing, and remedying the discriminatory harms produced by machine-learning systems in deployment domains "from policing, to welfare systems, to healthcare provision, to platforms for online discourse". The Declaration was finalised and announced at the RightsCon Toronto 2018 conference (16-18 May 2018, Beanfield Centre at Exhibition Place) following a full-day live-drafting convening of more than thirty civil-society, public-sector, and academic experts on 15 May 2018. Within the make-AI-good corpus it is the foundational publication-side artefact of the human-rights wing of algorithmic-accountability advocacy: the document under which a substantial slice of the international digital-rights and human-rights movement subsequently anchored its position that international human rights law, not voluntary industry ethics codes, is the binding normative framework for machine-learning governance.

What the Declaration asks for

The Declaration is structured as fifty-nine numbered articles across six sections — a preamble; a section titled Using the framework of international human rights law; Duties of states: human rights obligations; Responsibilities of private sector actors: human rights due diligence; The right to an effective remedy; and a references section — and organises its substantive demands into a double-track architecture mirroring the United Nations Guiding Principles on Business and Human Rights's state-duty-to-protect and corporate-responsibility-to-respect division.

On the states-side, the Declaration directs governments to identify risks and conduct regular audits of machine-learning systems before and after deployment, to avoid the procurement and use of "black box" systems whose decisional logic cannot be meaningfully explained, to maintain transparent oversight through diverse review committees and independent judicial authorities, to pass binding anti-discrimination and data-protection legislation that explicitly covers machine-learning decision-making, and to provide effective remedies for individuals and groups whose right to equality and non-discrimination is violated by an algorithmic decision.

On the private-sector side, the Declaration directs companies developing or deploying machine-learning systems to conduct human-rights due diligence — including meaningful consultation with stakeholders, "affected groups, organizations that work on human rights, equality and discrimination" — to design systems that actively mitigate identified discrimination risks rather than treating bias as a downstream artefact to be patched, to subject systems to regular independent audits, to decline projects whose discrimination risks cannot be adequately managed, and to create operational-level grievance mechanisms through which individuals affected by an automated decision can seek redress without first having to litigate access to the system's internals.

A separate section on The right to an effective remedy is the load-bearing innovation of the Declaration relative to the contemporary AI-ethics-principles literature: it names remedy as a substantive entitlement of the affected individual under article 8 of the Universal Declaration of Human Rights and article 2(3) of the International Covenant on Civil and Political Rights, not as a discretionary policy lever available to the developer or regulator. This framing is the Declaration's principal contribution to the legal-instrument scaffolding under which subsequent civil-society AI-rights advocacy — including the EU AI Act civil-society coalition's right-to-an-explanation and effective-redress demands — was built.

The coalition behind the Declaration

The Declaration's drafting was led by a five-person team of staff technologists and human-rights lawyers from the two convening organisations: Anna Bacciarelli (Technology and Human Rights Advisor) and Joe Westby at Amnesty International, and Fanny Hidvegi, Estelle Massé, and Drew Mitnick at Access Now. A seven-member steering committee of machine-learning and equality-rights experts from civil society and academia shaped the substantive direction: Boye Adegoke of Paradigm Initiative Nigeria; Frederike Kaltheuner of Privacy International; Malavika Jayaram of the Digital Asia Hub; Yasodara Córdova of the Berkman Klein Center for Internet & Society; Sherif Elsayed-Ali of Amnesty International; Solon Barocas of Cornell University; and William Isaac of the Human Rights Data Analysis Group.

The Declaration was live-drafted at a full-day expert convening on 15 May 2018 — the day before the RightsCon Toronto opening — bringing together more than thirty representatives from civil society, the public sector, and academia. The named institutional participants in that drafting meeting included Facebook, Harvard University, the United Nations, UNESCO, Human Rights Watch, Carnegie Mellon University, and digital-rights organisations and academic institutions across North America and Europe.

The launch on 16 May 2018 at 12:00 Eastern Time, at the RightsCon Toronto conference at the Beanfield Centre at Exhibition Place, was carried by a four-speaker panel: Estelle Massé from Access Now, Anna Bacciarelli from Amnesty International, Jan Gerlach (Public Policy Manager) from the Wikimedia Foundation, and Dinah PoKempner (General Counsel) from Human Rights Watch. The Wikimedia Foundation and Human Rights Watch are the two organisations that joined as endorsers at launch, alongside the two drafting organisations — a four-organisation founding bloc spanning international human-rights advocacy (Amnesty International, Human Rights Watch), digital-rights advocacy (Access Now), and the free-knowledge / platform-governance wing (Wikimedia Foundation).

At launch, Bacciarelli framed the Declaration as unique in "setting out tangible and actionable standards for states and private sectors" — positioning it explicitly against the proliferating set of voluntary AI-ethics principles documents being produced by industry and inter-governmental bodies in 2017-2018, none of which were binding and most of which collapsed remedy into a procedural goodwill statement. Massé summarised the Access Now position as a call urging "the public and private sectors to promote and respect human rights in the digital age, including the right to non-discrimination". The endorsement from the Wikimedia Foundation explicitly extended the framing beyond equality to "protecting not only equality but freedom of expression and access to knowledge". The endorsement from Human Rights Watch was framed by PoKempner with the line that "this declaration — that universal, international human rights law applies also to AI — is critically needed as the debate on ethics and bias in machine learning proceeds".

Reception and downstream adoption

Human Rights Watch republished the Declaration text on its own news channel on 3 July 2018, six weeks after launch, framing it as a document "drafted by rights groups, technologists, and researchers" and routing it into HRW's standing AI-and-rights advocacy track. The Toronto Declaration site has since collected open-endorsement signatures from a broader range of civil-society and research organisations beyond the four launch organisations, and the framework has been cited as a normative reference point in subsequent civil-society interventions on AI governance — including the human-rights-due-diligence and effective-remedy language in the EDRi-coordinated EU AI Act civil-society coalition's 30 November 2021 statement and the international-human-rights-law register that Access Now and Human Rights Watch have anchored their machine-learning-and-rights advocacy on through subsequent reporting and UN submissions.

Position within the corpus

Within the corpus the Toronto Declaration is the foundational publication-side artefact on the algorithmic-accountability ↔ international-human-rights-law coalition framework — the document that named the binding human-rights register as the load-bearing scaffolding for machine-learning governance and set the substantive ceiling against which the subsequent generation of voluntary AI-ethics principles, industry self-governance codes, and inter-governmental "responsible AI" frameworks were measured. It sits at the head of a publication-side lineage that the corpus also tracks through the EDRi-coordinated EU AI Act civil-society coalition's 30 November 2021 statement (regulatory-instrument-side application of the same human-rights-due-diligence and effective-remedy framing to EU horizontal AI legislation) and the broader international-human-rights-law register that Access Now and Human Rights Watch have carried into UN special-procedures submissions, OHCHR engagement, and the recurring civil-society statements on biometric surveillance, welfare-system automation, and AI in conflict that have appeared in the years since launch.

As a publication type it fills the manifesto slot, structurally distinct from the corpus's other coalition-letter and joint-statement entries by its standing as a substantive normative framework (fifty-nine articles, six sections) rather than a single-issue political position or open letter; the EDRi-coordinated EU AI Act statement is its closest corpus neighbour on the policy-instrument side, while the Feminist Principles of the Internet is its closest neighbour on the standalone-framework-manifesto side — the FPI's 17-principle five-cluster substantive architecture parallels the Toronto Declaration's 59-article six-section architecture as the two corpus-resident foundational multi-org civil-society normative frameworks for the make-AI-good movement, originating from adjacent generations of the field (the FPI from the 2014-2016 feminist-internet wave, the Toronto Declaration from the 2018 algorithmic-accountability turn) and converging on closely overlapping concerns about discrimination, remedy, due-diligence, and the political question of who the underlying systems are being built for and against. The Toronto Declaration's distinct contribution within that pairing is its direct anchoring in binding international human rights law (the Universal Declaration of Human Rights; the International Covenant on Civil and Political Rights; the International Convention on the Elimination of All Forms of Racial Discrimination; the UN Guiding Principles on Business and Human Rights) — a register that gave the document the specific evidentiary purchase to be cited in court filings, UN submissions, and regulatory consultations in a way that the parallel AI-ethics-principles documents of the same period did not enable.