Using AI responsibly

• Tekoälyn soveltamisen kyberturvallisuus ja riskienhallinta (National Cyber Security Centre, in Finnish)Opens in a new window.
• AI Risk Assessment Methodologies (Tri Lateral Research, 2021, PDF)Opens in a new window.
• The NIST Artificial Intelligence Risk Management Framework (US Government, 2023, PDF)Opens in a new window.
• How to Securely Design and Operate Machine Learning (Wilco van Ginkel / Gartner, 2023)Opens in a new window.

There are currently several definitions for feedback loop, but all of them revolve around the impact that feedback has on the functioning of an algorithmic system:

• A person using a system inevitably produces feedback inputs for the system, either consciously and intentionally or otherwise in connection with other actions. An AI that uses machine learning will modify its outputs based on the feedback that it collects over time.
• Interconnected machine learning systems can form a feedback loop that enables an AI model to become more accurate over time. Read more in this Ultimate.ai blog post How AI Uses Feedback Loops to Learn From Its MistakesOpens in a new window..
• A newer phenomenon is a feedback loop that arises when the amount of online material produced by generative AI (images, audio, video, text) grows exponentially and becomes training data for other AIs. This can result in a “copy of a copy of a copy” situation, as AIs train each other with materials they have learned from each other. If the materials contain errors or discrimination, they are iterated and escalated, unless these loops are sufficiently controlled.

There are two main reasons:

• The public sector has a large amount of valuable information openly available on the internet.
• In principle, all systems can be breached with the required motive, skills and resources.

Using proactive analytics requires data competence, commitment and understanding the substance of the system and broader contexts. If the outputs of an AI system are used in the work of persons who are not directly involved with the AI system, they need to have a sufficient understanding of the data used to train the system and how its outputs should be interpreted.

This is particularly important when

• a person’s work has a direct or indirect impact on people and their circumstances
• a person’s work is such that it creates an opportunity to give the system more decision-making power than it actually should have.

One essential method is proactive impact assessment. When designing and implementing new services, it is important to stop and consider whether the service can promote passiveness instead of making things streamlined, and what kind of leeway does it leave for human agency. In principle, everything can be automated, but the question is whether everything shouldbe automated.

We have to avoid creating services and structures that downright prevent encounters between people and the understanding that arises and is conveyed in these encounters. Inclusive development with civic involvement is important for gaining insight already at the design stage of how the services we are developing support people’s agency instead of taking it away from them.

We have to maintain a distance between human and machine, not the least to maintain our identity and a grasp on reality.

This does not mean that a customer service bot cannot be polite or that the main building of a large organisation should not have a lobby bot, and so on. However, no technological solution should have features that could genuinely be interpreted as human characteristics. Users should not be led to believe that a machine is capable of compassion or understanding like a human.

A person can be viewed as a machine in some philosophical frameworks, but a machine should never be viewed as a person.

Artificial intelligence entered the everyday lives of people and companies when OpenAI released ChatGPT in November 2022. In practice, ChatGPT is a product by a company. Technically speaking, it is a user interface for accessing a large language model developed and managed by the company.

Other companies’ “smart” chatbots and AI assistants, such as Microsoft Copilot, are based on a similar structure, but they each use their own language model.

All AI systems, including applications based on large language models, are powerful calculators using mathematical statistics. They are backed by huge sets of training data and algorithms with up to trillions of variables.

For example, an application could calculate that “most cats like to eat...” is more likely to be followed by “fish” than something like “deer”, because “cat”, “food” and “fish” are clearly more closely linked statistically in massive text materials than “cat”, “food” and “deer”.

The application does not actually understand the content. Even if the text outputs of ChatGPT or its competitors might seem like human thought, the selected words are merely strings from the AI’s perspective. Applications producing images and other materials generate their outputs on the same principle.

1. Generative AI does not create anything new. Instead, it repeats and statistically combines the material that it was trained on. Note that any bias and discrimination present in the training data and the algorithm will affect the output.
2. Remember that content produced by AI may contain factual errors or even “hallucinations”, such as references to non-existent media, decisions or events. Factual errors are usually caused by inadequate training material, whereas hallucinations are generated because the applications always try to produce some output, regardless of whether they have data on something or not.
3. Do not use content you produce with AI until you have verified the output personally or with the help of an expert. Blatant factual errors and hallucinations are usually easy to detect, but more subtle errors are easy to miss and disseminate further if they are not addressed.
4. Be always cautious when using material you suspect was produced by AI. Generative applications make it possible to produce an unprecedented amount of credible fake information, images, audio and video. Fake content is relatively easy to produce and difficult to distinguish from genuine content.
5. If you use a free AI application, do not enter any information about your organisation in the prompts you use in the application. Free versions of generative applications collect and store prompts as training material for the AI company. If you are using an application with an enterprise or premium licence, check your organisation’s policy of using the application.
6. If you use AI to produce, edit or translate a significant amount of content for presentations or documents, mention it during the presentation or at the beginning of the text.