Master Thesis: Establishing Trust in AI Through a Practical Framework

Are you a master’s student looking for an innovative company to collaborate with on your master’s thesis during the spring semester of 2027? Look no further.

ASSA ABLOY is the global leader in access solutions and provides the latest technologies to give access to physical and virtual spaces for millions of people every day. Joining ASSA ABLOY means being part of a fast-moving and highly innovative company with diverse opportunities in more than 70 countries worldwide. What can you expect as a master’s thesis intern at ASSA ABLOY? Cutting edge technology, a steep learning curve, dedicated supervisors from within the enterprise, a group of fellow master's thesis students to share the experience with – and much more.

We have an exciting range of subjects. Please remember to indicate in your application if you're interested in other subjects as well and mark them in your order of preference. Please also let us know your thoughts on how you would approach the subject. Please note that most of your work will be done at one of our two sites in Stockholm.


Master Thesis Nr 5: Establishing Trust in AI Through a Practical Framework

With AI now closely connected to our daily work, new questions arise regarding whether a solution is trustworthy and the consequences of launching a solution where the users lose trust in the data and its output. In a mix of AI-powered and deterministic solutions, what is the risk of losing the user’s trust? As AI becomes increasingly integrated into human cognitive processes, unverified or untrustworthy information/ hallucinations may become embedded within our epistemic practices, potentially leading to cognitive dependency and skill degradation.

This master’s thesis aims to design and validate a framework for evaluating the trustworthiness of AI-powered solutions. The framework should provide an automated evaluation mechanism that establishes measurable benchmarks for data quality, model behavior, and answer quality. An initial approach is to create a large, representative evaluation set containing thousands of relevant questions, expected answers or acceptance criteria, known edge cases, and deliberately misleading inputs. The solution can then be evaluated repeatedly as the data, prompts, models, or system design change.

The framework could evaluate an AI solution across the following dimensions: 

1. Data trustworthiness
   Accuracy, completeness, freshness, ownership, lineage, privacy, and access control of the data used by the solution. 
   
   

2. Answer trustworthiness
   Factual correctness, relevance, completeness, consistency, groundedness in approved sources, and appropriate handling of unknown information. 
   
   

3. System reliability
   Repeatability where required, resilience to failures, performance, monitoring, logging, and traceability across deterministic and AI components. 
   
   

4. Safety and governance
   Privacy, security, bias, compliance, auditability, human oversight, and clear responsibility for decisions. 
   
   

5. User trust experience
   Transparency about AI use, explanation of outputs, visible sources, confidence or uncertainty indicators, feedback mechanisms, and easy reporting of incorrect answers. 

Suggested questions of Exploration: 

• How to identify the keypersons for a solution to build trust?
  This may include end users, domain experts, data owners, product owners, AI engineers, security and compliance teams, and operational support teams. 
  
  

• How can we verify that the output from AI is relevant every time?
  Possible methods include rule-based checks, retrieval-source validation, domain-expert review, LLM-as-a-judge techniques, semantic similarity measures, and comparison against a trusted reference answer. 
  
  

• In what way can we spot the hallucination and how can we avoid it from happening?
  This includes unsupported claims, fabricated citations, incorrect summaries, contradictions, and answers that go beyond the available evidence. 
  
  

• When should a human be involved in the decision process?
  The framework should define risk thresholds for human approval, review, correction, or escalation. 
  
  

• Which areas of an AI-powered solution are most critical for user trust?
  Examples include data quality, source traceability, factual correctness, relevance, consistency, security, privacy, explainability, fairness, and behavior under uncertainty. 
  
  

• How can a representative benchmark dataset be created and maintained?
  How should questions, expected answers, edge cases, adversarial prompts, and real user scenarios be selected? 
  
  

• Which metrics best represent trustworthiness?
  Potential metrics include groundedness, factual accuracy, answer relevance, citation correctness, consistency, completeness, refusal quality, calibration of confidence, fairness, latency, and cost. 
  
  

• How can the framework support continuous evaluation after release?
  This could include regression testing, production monitoring, feedback collection, drift detection, periodic benchmark runs, and alerting when quality drops below agreed thresholds. 
  
  

• How should uncertainty be communicated to users without unnecessarily reducing usability or trust? 

A useful outcome of the thesis would therefore be a practical evaluation framework, benchmark design method, automated test harness, scoring model, and governance recommendations that organizations can apply throughout the lifecycle of an AI-powered solution.

In our team, we are building valuable AI-powered solutions for our internal users. We stay at the forefront of emerging technologies by exploring and applying agentic AI approaches, Graph RAG applications, knowledge engineering, ontologies, knowledge orchestrators, and other advanced capabilities.

We look forward to receiving your application with your resumé and hearing your ideas on how you would approach this subject!