Artificial intelligence: what about legislation?

If you want to get started with AI, the Dutch AI regulation, first recorded in April 2021, requires you to meet four conditions:Minimaliseer de bias (vertekening door systematische fouten): bouw in je AI verantwoordelijkheid in om ervoor te zorgen dat de algoritmen (en onderliggende data) zo onbevooroordeeld en representatief mogelijk zijn.

1. Minimize bias (bias due to systematic errors): build accountability into your AI to ensure that the algorithms (and underlying data) are as unbiased and representative as possible.
2. Ensure transparency of AI: to build trust with employees and customers, develop explainable AI that is transparent across all processes and functions.
3. Protect the privacy and security of the data used: use a privacy- and security-first approach to ensure that personal and/or sensitive data is never used unethically.
4. Keep algorithms ethical: by creating an ethical rationale for AI, you can mitigate risks and create systems that benefit your shareholders, employees and society as a whole.

Upcoming regulations (the AI regulation will become a law once it is further developed and ratified) also point in the direction of mandating explanations on the use of algorithms within the Dutch government. A Dutch national algorithms register is already operational.

Major players like Microsoft are also making significant strides in the area of guidelines. Read here how Microsoft thinks about responsible AI.

Regulatory AI European Commission

Our European Commission has also been working on legislation to limit the risks of working with AI for several years (see the AI Act timeline). There are calls within Europe to address the opacity, complexity, bias, some degree of unpredictability and partially autonomous behavior of some AI systems. This will make AI systems compatible with fundamental rights and facilitate enforcement of legal requirements.

1. Unacceptable risk

Systems with these algorithms are prohibited:

• Social Scoring algorithms (e.g. monitoring personal behavior in public places)
• Predicting personal criminal or fraudulent behavior
• Cognitive behavioral manipulation of individuals and vulnerable groups in particular, such as toys that respond to a voice and provoke dangerous behavior
• Personal classifications based on gender, sexuality, ethnicity or political affiliation
• Facial recognition in public places

2. High risk

These systems must adhere to restrictive rules and regulations, such as human rights assessment and elimination of bias in the data and algorithms used. This means full transparency in the data sources used and how the algorithms work. Human oversight of the systems is mandatory.

Examples:

• Control of critical infrastructure, such as road traffic and utilities
• Determining education level or, for example, a health insurance premium
• Recommendation algorithms with social impact
• Assessing people's suitability, e.g. for a job application or promotion

3. Limited risk

These systems must be transparent. The results may be used without human supervision:

• Systems for Chatbots on commercial websites

4. Minimal risk

No rules are imposed on systems, such as:

• Recommendation systems for online stores
• Recommendations for listening to music, watching movies et cetera
• Use of AI in the flow of computer games

Generative AI systems

Generative AI systems, such as ChatGPT, were outside the scope of the European Commission's original draft (April 2021). Basic rules for the application of generative systems, were not given and require extra attention, as their application is particularly broad. A separate chapter was recently included in the regulation for this purpose.

The European Parliament agreed to include rules for generative AI. Providers of it should assess and mitigate potential risks to health, safety, fundamental rights and/or democracy. For instance, it must be clear which datasets were used during the training of the AI system. The data should be objective, diverse and accurate, and it should be clear that the outcome was generated by AI.

Generative AI will have to meet a number of transparency requirements:

• Applications must include a statement that the content was generated by AI.
• The model should be designed to prevent illegal content being generated.
• There should be a record of copyrighted data used to train the system.

Balancing potential and risks

Despite all the focus on AI, the world has been caught off guard by the rapid rise of generative AI in particular. This technique is now finding wide application, with OpenAI making its ChatGPT solution available for free. This includes services in areas such as voice generators, image generators and music generators.

Questions such as what is true and what is fantasy? and is this real or not? are now commonplace. The challenge will be to prevent abuse of the unbridled enthusiasm of innovative organizations, without compromising the added value of AI.

A good example of how to deal with uncertainties and probabilities is the application of Microsoft's OpenAI-based Copilot that is now making its appearance in Europe. For now, the last word has not been said on this.