This paper explores the ‘cobra effect’ in human-robot collaboration, where interventions intended to improve trust inadvertently lead to worse outcomes. We conducted a controlled six-round CAPTCHA-solving experiment in which participants worked with a QT-Robot. Participants could solve each CAPTCHA themselves or delegate the task to the robot. The robot deliberately failed in rounds 4 and 5, triggering a trust repair strategy. Participants in the control group received a static pre-scripted apology, while those in the experimental group received adaptive apologies that incorporated user feedback. The results showed that adaptive apology led to moderate to large improvements in trust dimensions such as reliability, capability, and transparency. In contrast, static apologies resulted in reduced trust. These findings suggest that apology strategies must be sensitive to context and user perception. In addition, we identify the specific dimensions of trust negatively affected by static apologies, highlighting key areas for designing more effective context-aware trust calibration cues in collaborative robotics.