Would You Trust a Chatbot With Your Sex Life? Implementing and Evaluating a Secure, Anonyms Chatbot for Young People
Translated title
Would You Trust a Chatbot With Your Sex Life? Implementing and Evaluating a Secure, Anonymous Chatbot for Young People
Author
Michaelsen, Frederikke Filtenborg
Term
4. term
Education
Publication year
2026
Abstract
This thesis examines to what extent a sexual-health chatbot can be perceived as trustworthy by young people and how it compares with established digital alternatives. Building on pre-thesis work and expert-informed requirements, the prototype SafeBubble was developed as a secure, on-premises chatbot using retrieval-augmented generation from medically validated Danish sources. The system prioritizes anonymity, data security, transparency, and responsible scoping, including flagging uncertainty and referring to professional help when needed. The chatbot was evaluated in a between-subjects study against two systems: an established sexual-health website (Sex & Samfund) and a general-purpose conversational AI (ChatGPT). Participants completed realistic information-seeking tasks and a questionnaire measuring trust as a multidimensional construct grounded in Ability, Benevolence, and Integrity, supplemented by trust-in-automation constructs such as understandability, reliability, and behavioral intention. Comparative statistical analyses showed that trust varied across systems and dimensions; SafeBubble was rated particularly high on benevolence, integrity, perceived reliability, and understandability, but differences were not significant across all dimensions, including perceived ability and general trust in automation. Moreover, higher trust scores did not automatically translate into stronger behavioral intention. The study contributes empirical insight into trust as a contextual and multidimensional phenomenon in conversational sexual-health interventions and highlights the importance of architectural transparency, interaction quality, and responsible scoping. Future work includes closer integration of medical expertise, improvements to language quality and retrieval control, expanded measures of anonymity and data security, and longer-term evaluations over time.
Dette speciale undersøger, i hvilket omfang en seksual-sundheds-chatbot kan opfattes som troværdig af unge, og hvordan den sammenlignes med etablerede digitale alternativer. Med udgangspunkt i et pre-thesis-arbejde og ekspertinformerede designkrav blev prototypen SafeBubble udviklet som en sikker, on-premises chatbot baseret på retrieval-augmented generation fra medicinsk validerede danske kilder. Løsningen prioriterer anonymitet, datasikkerhed, gennemsigtighed og ansvarlig afgrænsning, herunder at markere usikkerhed og henvise til professionel hjælp ved behov. Chatbotten blev evalueret i et between-subjects design mod to systemer: et etableret dansk seksual-sundhedswebsite (Sex & Samfund) og en generel conversational AI (ChatGPT). Deltagerne løste realistiske informationsopgaver og udfyldte efterfølgende et spørgeskema, hvor tillid blev målt som et flerdimensionelt begreb baseret på rammerne Ability, Benevolence og Integrity samt supplerende trust-in-automation-konstrukter som forståelighed, pålidelighed og adfærdsintention. Komparative statistiske analyser viste, at tillid varierer på tværs af både systemer og dimensioner; SafeBubble blev især vurderet højt på velvilje, integritet, oplevet pålidelighed og forståelighed, men forskelle var ikke signifikante på alle dimensioner, herunder oplevet evne og generel tillid til automation. Desuden omsattes højere tillidsvurderinger ikke automatisk til øget adfærdsintention. Studiet bidrager med empirisk viden om tillid som et kontekstuelt og flerdimensionelt fænomen i samtalebaserede seksual-sundhedsinterventioner og understreger betydningen af arkitektonisk gennemsigtighed, interaktionskvalitet og ansvarlig afgrænsning. Perspektiver for fremtidigt arbejde omfatter tættere inddragelse af medicinsk ekspertise, forbedret sproglig kvalitet og retrieval-kontrol, udvidede målinger af anonymitet og datasikkerhed samt længerevarende evalueringer over tid.
[This apstract has been generated with the help of AI directly from the project full text]