The goal of the nudge is to guide the AI towards more cautious and reliable responses by embedding its autonomous behavior into a controlled process. Instead of providing an immediate answer to a customer inquiry, the system initially remains in a waiting state and requests specific clarifications when necessary to clear up the context, a step based on the ECLAIR model. For internal verification, the system performs a parallel comparison of multiple answer drafts, which is implemented through the LaMSeI model. If a high degree of internal uncertainty is detected during this process, the AI blocks the direct output of text and instead poses follow-up questions to the customer. If no sufficient clarity is reached after a maximum of three clarification attempts, the system terminates the automated questioning and seamlessly forwards the inquiry to a human employee. This automatic human-handover mechanism ensures that only verified information reaches the customer in critical cases.
Accept the cookie setting to see Panopto content.
Unlock PanoptoThe topic being addressed is the use of AI chatbots in customer service. A significant issue arises because so-called “AI hallucinations” can lead to the generation of false information. Since companies are legally liable for the statements made by their AI, this creates substantial risks. An example of the potential consequences is the sentence of the company „Aesthetify“, which involved the unauthorized attribution of incorrect medical specialist titles by a chatbot.
The goal of the nudge is to effectively reduce AI hallucinations and enable the system to precisely recognize and consider its own uncertainty. This process promotes legally secure and verified responses, thereby minimizing direct liability risks for companies. At the same time, this approach helps to sustainably strengthen the trust of customers in the company.
- AI tends to provide answers even when it lacks sufficient knowledge.
- It presents incorrect information with high confidence.
- AI overinterprets existing information.
- It generates plausible but unsupported facts.
- Context is lost, leading to misleading statements.
The target audience for this nudge primarily includes companies that utilize AI chatbots in their customer service. This approach also benefits customers indirectly, as the measures are intended to ensure they receive more reliable information.
adelina-maria.corsaro@stud.hshl.de
marie.kampen@stud.hshl.de
dylan-valentin.melo-schaenzlin@stud.hshl.de
mohammed.al-haj-khalaf@stud.hshl.de
