Why is ethics in the context of AI important?

An AI system can work technically reliably and still deliver results that your company does not want to represent – especially when people are evaluated, sorted, or treated differently. Large companies often set up their own ethics or governance bodies for this; for most SMEs, it's enough to clarify the crucial questions directly at the specific application.

It's not just about how often a system is correct, but how serious the consequences of an error weigh: A wrong product recommendation is quickly corrected, a faulty pre-selection of applicants weighs heavier. The question "Is this allowed?" does not automatically answer "Do we want to do it this way?" Legal requirements set the framework within which your company bears its own responsibility. Ethics therefore remains an ongoing process, closely linked to specific AI applications and to be regularly reviewed.

Engaging with ethics is particularly meaningful when a company:

• Uses AI for the evaluation or pre-selection of people, such as in applications,

• Allows customers to interact directly with an AI assistant,

• Makes AI-supported decisions with financial or legal consequences,

• Uses generative AI for content that could appear real,

• Is unsure whether a planned AI deployment is socially acceptable, or

• Aims to deliberately strengthen trust with employees and customers.

The sooner you clarify ethical questions at the specific application, the easier it is to anchor trustworthy AI in everyday work.

How do I proceed?

In 5 steps to trustworthy AI:

• Step 1 – Clarify the actual role of the AI

Look at what the AI actually does in the workflow: Does it just provide information, make a recommendation, or does it effectively determine the outcome because its recommendation is almost always adopted? For example, if an AI sorts applications in advance and those rated lower are hardly considered, it effectively decides who gets an interview. With generative AI, there's also the question of whether it creates content that customers might mistake for statements from real people. The crucial factor is the actual impact on people, not just the official purpose.

• Step 2 – Consider the consequences for those affected and set a clear goal

Formulate a specific, verifiable goal instead of a general intention – for example: "By the end of the month, it will be clarified which applicants perform systematically worse in our AI pre-selection." Clarify who is affected and what consequences arise: Incorrect information in customer service can usually be quickly corrected, a missed offer or candidate elimination weighs heavier. Pay attention to the balance between benefits and possible consequences – if your company saves only a little time while a mistake brings significant disadvantages to others, the use should be critically examined.

• Step 3 – Choose an application and look for systematic disadvantages

Choose a narrowly defined use case as a starting point, instead of reviewing all AI systems at once. Even good systems cannot completely avoid errors; it becomes critical when certain groups repeatedly perform worse. Therefore, compare typical practical cases instead of just the average success rate: Are certain applicants regularly classified worse? Does speech recognition work equally well for all customer groups? Involve employees who work with the processes—their experiential knowledge recognizes borderline cases that metrics alone do not show.

• Step 4 – Establish responsibilities and clarify correction process

Appoint a fixed contact person who accompanies the use case, plans the timeline, and reports progress to management. They determine when a human checks and corrects the AI recommendation—a formal approval is not enough if it is always adopted anyway. If people interact directly with AI or its evaluation has significant impacts, it should be clear to them what role the system plays and who to contact in case of problems.

• Step 5 – Evaluate after 4 to 8 weeks and draw consequences

Set a fixed date at the beginning for the evaluation after 4 to 8 weeks: Was the goal from step 2 achieved? Are systematic disadvantages still apparent? Record the insights in writing. The decision at the end is whether and how your company uses AI—for example, only as a recommendation instead of a decision, or abstention if a minor economic advantage is outweighed by significant disadvantages for others.

Podcast recommendation

Building Trustworthy Open Source AI

Find this and many more pieces of information on the AI Compass topic "Ethics" in our download portal: