Algorithm Charter for Aotearoa New Zealand

What it is

The Algorithm Charter for Aotearoa New Zealand is a commitment by government agencies to manage carefully how they use algorithms. It aims to strike the right balance between privacy and transparency, prevent unintended bias and reflect the principles of the Treaty of Waitangi. Stats NZ, through the Government Chief Data Steward, published it in July 2020.

The charter deliberately avoids a technical definition of an algorithm. It notes that the range runs from regression models and decision trees to neural networks and Bayesian models, and that a very simple algorithm can cause as much benefit or harm as a complex one. Instead, it commits signatories to focus on algorithms with a high risk of unintended consequences, or a significant impact if things go wrong, particularly for vulnerable communities.

Who it applies to

New Zealand government agencies that choose to sign it. Signing is voluntary. Each signatory’s chief executive, chief privacy officer and the senior manager responsible for algorithms sign. Stats NZ publishes the current list of signatories, which has changed since the charter launched.

What it requires

By signing, an agency commits to assessing the impact of decisions informed by its algorithms, and to applying the charter’s commitments as the risk rating guides.

Transparency. Clearly explain how decisions are informed by algorithms. That may include plain English documentation of the algorithm, information about the data and processes (unless a lawful restriction prevents it) and how data is collected, secured and stored.

Partnership. Deliver clear public benefit through Treaty commitments by embedding a Te Ao Māori perspective in developing and using algorithms, consistent with the principles of the Treaty of Waitangi.

People. Identify and actively engage with people, communities and groups who have an interest in algorithms, and consult those affected by their use.

Data. Make sure data is fit for purpose by understanding its limitations and identifying and managing bias.

Privacy, ethics and human rights. Regularly peer review algorithms to assess for unintended consequences, and act on what is found.

Human oversight. Nominate a point of contact for public enquiries about algorithms, provide a channel for challenging or appealing decisions informed by algorithms and clearly explain the role of humans in those decisions.

The risk matrix

Applying the charter to every business rule would be impossible and would not achieve its aims. Signatories therefore assess their algorithmic decisions with a risk matrix that sets the likelihood of an unintended adverse outcome against its relative impact.

• Likelihood. Improbable (unlikely but possible in standard operations), occasional (likely to occur some time) or probable (likely to occur often).
• Impact. Low (isolated or not serious), moderate (reaching a moderate number of people or of moderate severity) or high (widespread or serious).

The overall rating decides how the charter applies. At low risk the charter could be applied, at moderate risk it should be and at high risk it must be. The intention is for signatories to focus first on decisions with a high risk of unintended harm and exclude most routine business rules that give effect to legislation.

Putting it into practice

For a signatory, the commitments translate into a short list of things it should be able to show for each algorithm rated moderate or high. That means a plain English description of what the algorithm does and how it informs decisions, a record of the data it uses and its known limitations and biases, evidence of engagement with affected people and of a Te Ao Māori perspective in its design, a peer review and what was done about its findings, and a named public contact with a route to challenge or appeal. The rating itself, and the reasoning behind it, are worth keeping too. They are what shows why the charter was or was not applied.

Key dates

The charter was published in July 2020 with a review planned after twelve months. An independent review was completed in December 2021. It judged the risk matrix a relatively simple tool and suggested a more detailed triage, and Stats NZ began work on its findings. The 2020 charter text and its risk matrix are what signatories commit to.

How it relates to other frameworks

The charter sits alongside the Public Service AI Framework, which sets the wider principles for AI in the public service. Where algorithms use personal information, the Information Privacy Principles under the Privacy Act 2020 apply, including the duty to check accuracy before use. The charter’s risk-based approach is similar in spirit to Canada’s mandatory Directive on Automated Decision-Making, and its transparency commitment parallels the UK’s Algorithmic Transparency Recording Standard. In Australia, the closest public-sector equivalents are the DTA policy and the NSW AI Assessment Framework.