- Anzeige -
read time: min
Ethics and Explainability related to Artificial Intelligence

Dec 10, 2019 | Allgemein, Topstory

Ethics & AI

Ethics and Explainability related to Artificial Intelligence

Since today’s artificial intelligence does not develop an awareness of what it learns, AI in principle is completely objective. Therefore the risks of using AI are related to the data and not in the AI itself. The current discussion of AI and ethics aims to prevent artificial intelligence systems from discriminating people.

(Image: ©Tatiana Shepeleva/stock.adobe.com)

To achieve this goal these systems should be designed to minimize the risk of unintended consequences. For example, a company would need to test upfront whether artificial intelligence discriminates or devalues certain groups of people.

In what context can ethics and AI be seen?

1.) Integration of ethics in AI

Since AI does not develop awareness of the learned content and ethics is not objectifiable, ethics can not be meaningfully integrated into artificial intelligence.

2.) Ethics of the data

The result of an AI is based on the data. Therefore the responsible use of data is an ethical approach to AI.

3.) Ethics of the results

This raises the question of how to assess the result of an AI and how AI and humans interact to maintain ethical principles.

If you compare artificial intelligence with the current methods, it becomes evident that AI basically fulfills the same purpose as “electronic data processing” for around 60 years. Data processing and evaluation. The two approaches differ only in the type of models used to evaluate the data.

Humans are limited in their ability to recognize complex relationships by Miller’s number (7+-2). Therefore in human models complex relationships are simplified and e.g. evaluated in the form of a

decision tree. Therefore results are explicable within the model. The question of whether the assumptions of the model are statistically valid or (implicitly) discriminating is usually neither asked nor validated. It is largely socially accepted to get a justification whose explanatory value, i.e “relevance with respect to the decision”, is typically not verifiable.

Models of artificial intelligence recognize relationships directly in complex data. Instead of simplified assumptions, they reflect the statistical reality of the data. As the evaluation is basically based on statistical correlations, this enables a fundamentally fairer assessment. However this becomes problematic when data is incomplete or contains apparent relationships that are not transferable to reality.

Due to the implied legal risks, it is a fundamental goal to develop non-discriminating systems. In order to prevent discrimination in AI system several techniques can be used such as:

1. Removal of potentially discriminatory features from the data

2. Add more of the underrepresented data features

3. Higher weighting of under-represented data features

Overall dealing with this kind of data issues is a common but usually well manageable problem. Yet AI does not provide a universal solution for every task. If no sufficient amount of data can be made available, AI usually does not represent a good solution approach. Human intuition and ethics complement perfectly with the speed and objectivity of AI. Since AI quickly uncovers false or incomplete human assumptions, not only the AI learns from humans, but humans can learn from AI. Since human intuition can not be replaced by artificial intelligence in the medium term,

we generally recommend the use of AI as an assistance system. This can not just increase quality and efficiency of a company. In times of shortage of skilled workers, companies can benefit by using AI to increase the productivity of key employees.

Legal risks related to a lack of nationwide guidelines in dealing with the GDPR already present a significant obstacle to the introduction of artificial intelligence in Germany. From our point of view the use of AI in a production environment is still an isolated case within the industry or the public sector. The requirement that systems of artificial intelligence fundamentally must correspond to ethical principles and must be explainable means an additional hurdle that is not generally met by today’s classical data processing systems.

Regardless of the technology used, the GDPR prohibits people being subject to fully automatic decisions, given the decision has a significant impact on the person. As therefore compliance related to ethical standards and explainability is an implied requirement for relevant decisions, we recommend to not differentiate in this respect between the classical data processing systems an AI.

In order not to further increase Germany’s enormous backlog related to artificial intelligence, the goal for the further discussion about AI should be from our point of view on how to reduce legal risks by creating legal guidelines related to the GDPR, how to reduce bureaucracy related to public funding as well as raising awareness within the industry and public sector about the specific advantages of AI.



Autonomous Picking
Autonomous Picking

Image: Universal Robots (Germany) GmbH Today, most bin...


this may also interest you

Safety First

For years, image processing has only marginally dealt with the subject of safety.

inVISION at home

Special events require special reactions. To ensure that you always have access to the current issue even during the Corona pandemic and that you can read our trade journals in your home office (or any other place), TeDo Verlag provides you with the upcoming issues of its trade journals free of charge as an eMagazine.

Fingers Crossed

Probably none of us imagined March 2020 at the beginning of the year like this.

Who’s next?

Unfortunately, the coronavirus currently has the trade fair landscape firmly under its control.

Access to AI

Meanwhile more and more users have made their first experiences with AI.

DGQ Quality Day

DGQ Quality Day For the sixth time, the Deutsche...