Using AI responsibly

Cross-administrative systems and the cross-use of big data continue to make situations more challenging. Citizens’ ability to see what their data is used for, how they can manage said data and whether they can have it deleted becomes obscured.

– Publication of the Prime Minister’s Office: Artificial intelligence in authority use (in Finnish)Opens in a new window..

If an AI system has direct or indirect impacts on citizens’ everyday lives and rights, or on companies’ everyday operations and rights, users have to be informed of the system’s operations, objectives, outputs and impacts in a way that is as comprehensible and plain-language as possible.

For commercial applications, this falls under consumer protection, and for public sector and administrative applications, it falls under the legal and democratic rights of citizens and businesses. If systems are described and explained in a way that is difficult to understand or incomprehensible to their targets, these rights are compromised.

Read more about the responsibility explanations of decisions made with AI in this publication by the Alan Turing Institute (PDF)Opens in a new window..

The responsibility for having an understandable description of an AI service rests as much with the system supplier as the party procuring the system. Neither can avoid their part in the matter. The information explaining a system is always context-dependent, which is why the author of the description has to understand the target and user groups.

In fact, it is best if the information is produced in collaboration by the system supplier, the procuring unit and the procuring organisation’s communications. This ensures the accuracy of the technical information and the clarity of the message.

Clear communication also benefits the brand internally, especially in larger organisations where some members of staff may not have even known that an AI system was being procured, much less its objectives or features.

Explaining is not the same as being understood. Even if we could transparently present the logical function of an AI to a person subject to a decision made by the AI, the person might feel like they got an explanation but still remain as confused as before.

Since the logic behind the decisions of an AI system is difficult to explain, it is important that the transparency of a service is promoted at least with clear information on

• who is or which parties are responsible for the service
• what purpose the service was made for
• what information the user has to give the service
• why the user must give this information specifically.

Focus on the clarity and comprehensibility of the communication as well as its style and tone. The aim is to be as inclusive as possible. Transparency also includes clear instructions to users on what to do in case they want to correct errors or submit a claim for rectification.

Different contexts require different levels of openness. For example, you cannot let ordinary service users inspect your algorithm because

• for the majority of citizens, the algorithm would be incomprehensible, probably confusing and would not promote user understanding and trust
• even in the public sector, an AI system is likely provided by a company, and the algorithm is part of a trade secret
• an algorithm that is publicly available for anyone would likely be threat to the security of the service.

Although algorithms cannot be fully opened to users, they still need to be reviewed. Accountable AI implementation includes system audits by an independent operator.

Keep the details of the audit strictly confidential between the system supplier, owner and auditor, but publish the main findings and any subsequent measures. Do so at a level that does not compromise the security of the system but gives users enough information to maintain trust.

Read more about auditing algorithms in this article on the German Marshall Fund websiteOpens in a new window..