Can AI Explanations Be Manipulated?

MSc research · Findings and supporting materials are unpublished

The question

Can an explanation generated by an AI system still be trusted when an adversary can influence or manipulate that explanation? My MSc research investigates the tamper resistance of explainable-AI methods under adversarial conditions.

The study

I am developing a reproducible benchmark for assessing how selected explainability methods behave when explanation outputs are subjected to adversarial manipulation. The work is designed around controlled comparison, repeatable experiments, and transparent evaluation criteria.

Why this matters

Explanations are often used to support human understanding, oversight, and confidence in AI systems. If those explanations can be manipulated without clear warning, they may create a misleading appearance of transparency or reliability.

What the work contributes

The contribution is an evaluation approach: a way to test explanation robustness using clearly defined conditions and reproducible benchmarking, rather than relying on a single persuasive explanation or demonstration.

Research status

This work forms part of an MSc dissertation that has been graded and formally accepted; the formal degree award is expected in November 2026. Findings and supporting materials are unpublished. This page describes the research question and approach only; it does not claim peer review, publication, or generalised results.