This thesis examines how audiences evaluate trust, credibility, emotions, and ethical acceptability in nonfiction films when comparing traditional indexical imagery with AI- generated synthetic visuals. The implicit contract of nonfiction, that what is shown has an indexical link to reality, is increasingly unsettled by AI. As synthetic media enters documentary and journalism, a pressing question arises: how do audiences evaluate trust and authenticity when the anchor of recording is replaced by machine inference?

Two short films were produced for the study: a real documentary constructed from conventional footage and a synthetic version generated using machine learning models. Participants did not know whether the film they viewed was real or synthetic, ensuring that their responses were shaped by the viewing experience itself rather than by disclosure. A mixed-methods design was employed. Quantitative survey data were analysed in Survey Monkey and Excel, while qualitative open-text responses were coded and thematically analysed in NVivo. The analysis was organised into six thematic areas: educational value and perceived reliability; emotional response and empathy; perception of ethical boundaries; realism and visual impact; trust and credibility; and viewer awareness and interpretive frames.

Findings highlight three overarching contributions. First, indexical anchors remain central to trust: technical flaws in the real film were tolerated as signs of authenticity, whereas anomalies in the synthetic film were interpreted as evidence of unreality. Second, in the AI- generated film, two deliberately designed reconstructions of a deceased individual produced heightened emotional closeness while also provoking strong ethical discomfort, a tension conceptualised here as the empathy–ethics paradox and one of the study’s most significant findings. Third, ethical boundaries were drawn more firmly around people than places: while AI use to recreate environments was broadly accepted if transparent, recreating deceased individuals was largely rejected.

Building on these insights, the study proposes Inference Journalism as a new professional genre term for nonfiction storytelling. Defined as the transparent use of AI/ML to infer and reconstruct places, events, or people from real-world anchors such as photographs, recordings, or data traces, Inference Journalism frames synthetic material as reconstruction rather than deception. By offering audiences a clear interpretive category, this genre proposal seeks to stabilise trust and expectations in nonfiction practices during the age of synthetic media.