Fairness Mirror

Fairness Mirror

Our nudge starts at a point where discrimination in customer service is often not openly visible, but quietly runs along in seemingly neutral AI decisions. When a customer writes briefly, incorrectly or very emotionally, when a customer has only a small order value or hardly any purchase history, such signals can be read by an AI unnoticed as indications that the case is less important or deserves less goodwill.

This is exactly where our fairness nudge intervenes, because it is not aimed primarily at customers, but at the AI itself. Before the AI proposes a solution, it should check whether it is actually evaluating the concrete problem or whether its recommendation is being influenced by characteristics such as language, place of residence, purchase value or presumed ability to pay. The decision should therefore not be guided by social signals, but by factual criteria such as the actual problem, the waiting time, existing claims and the available solution options.

This creates a brief moment of pause in a process that would otherwise run very quickly and automatically. The nudge is intended to prevent socioeconomic cues from silently leading to poorer service, and it ensures that comparable concerns also receive comparable chances of a fair solution.

Accept the cookie setting to see Panopto content.

Unlock Panopto

What does the topic mean?

Our nudge is applied in customer service. In the area of technical customer service, many companies face the challenges of a shortage of skilled workers and a high volume of information, which makes fast and reliable customer support more difficult. With the help of LLMs, however, it is possible for companies to offer more cost-effective and efficient support in customer service. According to a study by the Institute for Business Value, all companies stated that they are integrating generative AI into customer service; in addition, 67% of the surveyed companies have already begun introducing AI in customer service. Because of its high practical relevance, this field of application is particularly suitable for testing our developed nudge.

Goal of the nudge

Our nudge aims to reduce socioeconomic bias within chatbots in the field of customer service. This should eliminate potential differences in the quality, interpretation and quantity of responses in customer consultation.

    Needs analysis

    • The need arises where companies increasingly automate customer service, but fairness is often not yet considered systematically. LLMs can process inquiries quickly, but in doing so they also take over patterns from data that reflect social differences. As a result, customers with simpler language, a low order value or a weak purchase history may unconsciously receive poorer solutions, even though their concern is objectively of equal weight. Our nudge addresses exactly this gap. It creates a brief fairness moment in the automated service process and ensures that the quality of the solution is not determined by social signals, but by the actual customer problem.

    Cause analysis

    • Because LLMs are trained on datasets that contain human distortions and stereotypes, the AI also reproduces bias. LLMs learn statistical relationships between level of education, writing style, occupation or income and certain behaviors or characteristics. This can lead to socioeconomic bias, causing people from differently perceived socioeconomic statuses to receive different results generated by the AI. Ultimately, this leads to systematic distortions in the treatment of people based on characteristics that suggest economic or social status, such as occupation, language and education.

Target Group

Our target group primarily includes companies whose customer service is supported by LLMs.

Added value of the nudge

  • Fair Customer Service Through Responsible AI Use

    The added value of our nudge lies primarily in enabling companies to use LLMs in customer service not only faster, but also more fairly and responsibly. Especially where many inquiries are processed automatically, there is a risk that an AI unconsciously draws conclusions from seemingly incidental signals such as writing style, place of residence, purchase history or order value and thereby treats certain customers worse without this distortion being immediately noticeable.

  • A Protective Layer For Fair Service Decisions

    For companies, the nudge therefore provides an additional protective layer between the automated recommendation and the actual service decision. It helps the AI evaluate the case more strongly according to factual criteria and not according to socioeconomic clues that should not actually play a role in the solution. As a result, comparable concerns can be handled more consistently, which not only increases fairness but also creates trust in AI-supported customer service.

  • Conscious Review

    At the same time, the nudge protects companies from a problem that often only becomes visible when customers feel unfairly treated or complaints escalate. It makes the use of LLMs more controllable, more transparent and more compatible with a customer-oriented understanding of service. The nudge thus creates customer service that is not only efficiently automated, but also looks more closely before small social signals turn into major differences in treatment.

CONTACT US

Claire Daly

claire.daly@stud.hshl.de

Julian Bertram

julian.bertram@stud.hshl.de

Lilly Koschlakow

lilly.koschlakow@stud.hshl.de