Using AI responsibly

• Tekoälyn etiikka (Helsingin yliopisto, 2021, in Finnish)Opens in a new window.
• Ethics in the AI Lifecycle. Theoretical Perspectives, Practical Resources and Recommendations. (Otto Sahlgren, Tampereen yliopisto, 2023)Opens in a new window.
• Digital Technology, social welfare and Human Rights (UN Special Rapporteur for Human Rights, 2019)Opens in a new window. 
• Getting the future right: Artificial intelligence and fundamental rights (EU Agency for Fundamental Rights, 2020)Opens in a new window. 
• Using artificial intelligence for social impact (Garcia & Richardson / Nesta Foundation, 2022)Opens in a new window. 

Systems based on deep neural networks have the capacity to produce speeds and volumes that overshadow conventional systems. Therefore, their possible errors and negative impacts are also scaled up in their extent and effect.

At least for the time being, training AI systems almost always requires significant masses of data. There are critical questions related to topics such as:

• the suitability of collected data to the objective that has been set
• data collection procedures
• the transparency and acceptability of data collection and use
• quality criteria for training data, and how to define and monitor said criteria.

AI systems are inevitably designed, implemented and maintained by many actors. Determining the liability for loss and damages may be extremely difficult. This is because the root of an issue can just as well be in the biases of the training data, miscalculations in the data selection or in the parameters of the algorithm.

One general solution has been to consider that the party responsible for the implementation of a system bears the main accountability, but ethical evaluation finds that this policy can often neglect important background factors.

When autonomous systems can produce seemingly credible material and are capable of making formally valid decisions on matters concerning individuals, they can reduce the power held by responsible human actors and their willingness to intervene in the machine’s outputs.

In ethical research, this risk is called the “rubber stamp problem”: people often have a natural desire to do less and transfer accountability to someone else, even if that someone is a machine.

Effective algorithmic systems can now target influencing and steering measures on individuals and population groups. This may be benevolent or harmless (e.g. apps that recommend entertainment to users), but public discourse and research have also revealed many cases of deliberate efforts to influence people negatively.

It is often difficult to get a clear picture of actors and their real motives, which makes it difficult to recognise the nature of a situation.

We are delegating more and more decisions to intelligent systems, from how to get to work to what to eat for dinner. This can improve our lives, by freeing up time for activities we find more important. But it is also forcing us to rethink our understandings of human dignity and agency, as algorithms are increasingly sophisticated at manipulating our choices.

– Report of the UN Secretary-General’s High-level Panel on Digital Cooperation, 2019

New technologies cause disadvantages that are unpredictable and often societal. No-one specifically planned to prepare Facebook algorithms for hostile political manipulation and spreading disinformation. Twitter was never meant to become an arena that aggravates and polarises opinions, and Instagram was not meant to become a platform that causes body image issues and self-esteem crises, especially for girls and young women.

No longer do people generally believe that the internet is an impartial information superhighway for everyone, here to help us end wars and promote mutual understanding, as Bill Clinton envisaged during his presidency. We have to assess the promises of new technologies with a particularly critical eye, and we have to make an effort to see the possible obstacles and reverse effects of their realisation alongside potential benefits and progress.

Studies have shown that AI systems have a tendency to produce decisions or analyses that systematically favour certain groups at the expense of other groups.

Individual cases of algorithmic discrimination have also been observed in Finland. Most cases are based on gender, age, ethnicity, language, socio-economic status or other similar protected personal characteristics. This is often due to training data that inevitably reflects past circumstances and may include old discriminatory structures. For example, we know that face detection algorithms have difficulties in recognising people with dark skin, as there are considerably more white people in the masses of online images that are used as training data.

Sometimes discrimination can also occur if a characteristic of a person that can be grounds for discrimination (like age or gender) is used as a critical variable in an algorithm. This might cause an application to treat people unfairly, for example with loan applications or in job recruitment.

The descriptive metadata generated for images is also often discriminatory in itself. If you ask these machines, a woman in a white coat is usually a “nurse” while a man in a white coat is usually a “doctor”. When these kinds of biases are introduced to high-speed and mass-scale systems, their effects spread rapidly and globally and continue to impact the training of new AI systems, unless we consciously break this vicious cycle.

By nature, the personal data of an individual and other data related to a person is almost always relational and social: in a deep sense, they do not belong to any single individual. Most of the data describing an individual and their situation is generated in interaction with other people or services. Similarly, the decisions, services, benefits, recommendations or limitations generated for individuals by algorithmic systems are often reflected on their social environment: family members, work community and friends.

Due to this same mechanism, people consenting to or refusing cookies and access to user information in online services is ultimately not an individual matter. When the overwhelming majority of individuals gives their consent in any case, the right of an individual who denies that consent to protect their privacy while receiving the same service begins to lose its meaning.

The era of machine learning effectively renders individual denial of consent meaningless.

– Martin Tisné, Luminate