At a glance
Explaining AI-assisted decisions has benefits for your organisation. It can help you comply with the law, build trust with your customers and improve your internal governance. Society also benefits by being more informed, experiencing better outcomes and being able to engage meaningfully in the decision-making process. If your organisation does not explain AI-assisted decisions, it could face regulatory action, reputational damage and disengagement by the public.
In more detail
- What are the benefits to your organisation?
- What are the benefits to individuals and society?
- What are the risks of explaining AI decisions?
- What are the risks of not explaining AI decisions?
As set out in the legal framework section of this guidance, a number of laws (both sectoral and cross-sector) have something relevant to say on this topic. Some explicitly require explanations of AI-assisted decisions in certain circumstances, others have broader requirements around the fair treatment of citizens. But whatever sector or business you are in, explaining your AI-assisted decisions to those affected will help to give you (and your board) better assurance of legal compliance, mitigating the risks associated with non-compliance.
Explaining AI-assisted decisions to affected individuals makes good business sense. This will help to empower them to better understand the process and allow them to challenge and seek recourse where necessary. Handing a degree of control back to individuals in this way may help to foster trust in your use of AI decisions and give you an edge over other organisations and competitors that are not as progressive and respectful in their interactions with customers.
Explaining AI-assisted decisions to affected individuals requires those within your organisation to understand the models, choices and processes associated with the AI decisions you make. So, by making ‘explainability’ a key requirement, you will also have better oversight of what these systems do and why. This will help to ensure your AI systems continue to meet your objectives and support you in refining them to increase precision.
As more organisations incorporate explanations to individuals as a core element of their AI-assisted decision-making systems, the general public will gain an increasing awareness of when and where such decisions are made. In turn, this may help the public to have a meaningful involvement in the ongoing conversation about the deployment of AI and its associated risks and benefits. This could help address concerns about AI and support a more constructive and mutually beneficial debate for business and society.
Organisations are required to identify and mitigate discriminatory outcomes, which may already be present in human decision-making, or may be exacerbated or introduced by the use of AI. Providing explanations to affected individuals can help you to do this, and highlight issues that may be more difficult to spot. Explanations should therefore support more consistency and fairness in the outcomes for different groups across society.
Giving individuals explanations of AI-assisted decisions helps to ensure that your use of AI is human-centric. The interests of your customers are paramount. As long as you have well-designed processes to contest decisions and continuously improve AI systems based on customer feedback, people will have the confidence to express their point of view.
Industry engagement activities we carried out highlighted a number of elements that may have a limiting effect on the information that can be provided to individuals when explaining AI-assisted decisions. The explanations set out in this guidance have largely been designed to take these issues into account and mitigate the associated risks, as explained below.
It could be argued that providing too much information about AI-assisted decisions may lead to increased distrust due to the complex, and sometimes opaque, nature of the process.
While AI-assisted decisions are often undeniably complex, the explanation types and explanation extraction methods offered in this guidance are designed to help you, where possible, to simplify and transform this complexity into understandable reasoning. In cases where fairness and physical wellbeing are a central issue, focusing on relevant explanation types will help you to build trust and reassure individuals about the safety and equity of an AI model without having to dive deeply into the complexity of the system’s rationale. This is particularly the case with the safety and performance explanation and fairness explanation. These show how you have addressed these issues, even if the rationale of a decision is particularly complex and difficult to convey.
You may have concerns that your explanations of AI-assisted decisions disclose commercially sensitive material about how your AI model and system works.
We don’t think the explanations we set out here will normally risk such disclosures. Neither the rationale nor the safety and performance explanations require you to provide information so in-depth that they reveal your source code or any algorithmic trade secrets. However, you will have to form a view based on your specific situation.
Where you do think it’s necessary to limit detail (eg feature weightings or importance), you should justify and document your reasons for this.
Third-party personal data
Due to the way you train your AI model, or input data for particular decisions, you may be concerned about the inappropriate disclosure of personal data of someone other than the individual the decision is about.
For some of the explanations we identify in this guidance this is not a problem. However, there are potential risks with the rationale, fairness and data explanation types – information on how others similar to the individual were treated and detail on the input data for a particular decision (which is about more than one person).
You should assess this risk as part of a data protection impact assessment (DPIA), and make justified and documented choices about the level of detail it is safe to provide for these explanations.
Depending on what you make AI-assisted decisions about, you may need to protect against the risk that people may game or exploit your AI model if they know too much about the reasons underlying its decisions.
Where you are using AI-assisted decisions to identify wrongdoing or misconduct (eg fraud detection), the need to limit the information you provide to individuals will be stronger, particularly about the rationale explanation. But you should still provide as much information on reasoning and logic as you can.
However, in other settings, there will be relatively few risks associated with giving people more detail on the reasons for decisions. In fact, it will often help individuals to legitimately adjust their behaviour or the choices they make in order to achieve a desirable decision outcome for both parties.
You should consider this as part of your initial risk or impact assessment for your AI model. It may form part of your DPIA. Start with the assumption that you will be as open and transparent as possible about the rationale of your AI-assisted decisions, and work back from there to limit what you tell people if you decide this is necessary. You should justify and document your reasons for this.
While we cannot speak for other regulators, a failure to meet legal requirements around explaining AI-assisted decisions and treating people fairly may lead a regulator to take regulatory intervention or action. The ICO uses education and engagement to promote compliance by the organisations we regulate. But if the rules are broken, organisations risk formal action, including mandatory audits, orders to cease processing personal data, and fines.
Public and media interest in AI is increasing, and often the spotlight falls on organisations that get things wrong. If you don’t provide people with explanations of AI-assisted decisions you make about them, you risk being left behind by organisations that do, and getting singled out as unethical and uncaring towards your customers and citizens.
Not explaining AI-assisted decisions to individuals may leave them wary and distrustful of how and why AI systems work. If you choose not to do this, you risk a disengaged public that is slower to embrace, or even reject AI more generally.