Challenge

Short-form video platforms such as TikTok have become a major source of health and nutrition information, in particular for younger users. Generative AI (GenAI) has pushed the cost of producing such videos close to zero. Prior work of the supervising team conceptualizes the resulting flood of high-volume, low-effort AI-generated content (AIGC) as AI slop and identifies prevalence measurement and engagement effects as open research gaps. First observations indicate that AI slop is especially common among health and nutrition videos on TikTok.

A specific and potentially harmful subtype is impersonation: AI-generated videos in which a synthetic doctor speaks directly to the viewer, gives medical advice, promotes supplements, or questions established treatments. Some of these videos clone the face and voice of real physicians. Others invent doctors who do not exist and dress them in white coats, titles, and clinic settings. In both cases, the videos borrow the authority of the medical profession. Once viewers notice the deception, or simply learn that such fakes are widespread, the damage may reach beyond the single video and erode systemic trust in doctors and health institutions. So far, there is no systematic measurement of how common these videos are, what they look like, and how audiences react to them.

Objective

The thesis aims to measure the prevalence of AI-generated health videos on TikTok, with a focus on doctor impersonation, and to explore indicators of trust erosion. Specifically, it could:

• measure the share of AIGC among health and nutrition videos, with English-language content as the main corpus and German-language content as an addition for comparison and contextualization,
• identify and characterize impersonation videos (e.g., cloned real doctors vs. invented doctors, authority cues such as titles and white coats, topics, calls to action, commercial intent),
• compare engagement with impersonation videos, other AIGC, and non-AI health videos,
• explore whether and how such videos are linked to signs of declining trust in doctors and health authorities.

Methodology & Expected Results

The work can be methodologically grounded in a computational content analysis combined with qualitative coding. TikTok is the primary platform, building on prior work of the supervising team. Data access is not yet settled and can be adapted to the interest of the candidate and to what is available at the start of the thesis. Possible data sources include:

• the TikTok Research API or data access via tools,
• optionally, the YouTube Data API (YouTube Shorts) for a cross-platform comparison,
• existing datasets, data donation, or independent scraping, subject to ethical and legal standards.

Possible approaches include:

• keyword- and hashtag-based sampling of health and nutrition videos (English first, German as an addition),
• detection of AIGC through platform labels, existing synthetic media and deepfake detectors, and manual annotation,
• a codebook for medical impersonation (e.g., claimed credentials, direct address of the viewer, product links, health claims),
• statistical comparison of views, likes, shares, and comments across content types.

For the trust erosion part, the candidate could choose one of two routes, depending on interest and scope (or follow a completely different approach):

• Comment analysis: classification of comments for expressions of trust or distrust toward the video, toward doctors in general, and toward health institutions (e.g., LLM-assisted classification with manual validation).
• Small user study: an online survey or experiment that tests whether exposure to fake doctor videos, or learning that a video was fake, reduces trust in real doctors.

Expected artifacts may include:

• an annotated dataset of health videos (AI vs. non-AI, impersonation vs. other AIGC),
• a taxonomy and codebook of AI-generated medical impersonation,
• prevalence estimates and a statistical assessment of engagement differences,
• an assessment of trust erosion indicators based on comments or a user study.

Impact

Health misinformation can cause direct harm, from wasted money on supplements to delayed treatment. Fake doctor videos add a second risk: they draw on the trust people place in the medical profession and may wear that trust down over time. The thesis provides empirical figures on a problem that is so far mostly described anecdotally. The results can inform platforms, medical associations, physicians whose identity is misused, and regulators working on the DSA and the transparency obligations for synthetic content under the EU AI Act. The thesis also contributes to the AI slop research agenda of the supervising team and is suited for joint publication.

Requirements for the Candidate

• Experience with data collection and analysis (e.g., Python or R)
• Basic knowledge of machine learning or computer vision; experience with LLM-based classification is an advantage
• Interest in health communication, disinformation, and generative AI
• For the user study route: basic knowledge of survey or experimental design
• Awareness of ethical and legal constraints (platform terms of service, privacy, handling of user comments)
• Interest in publishing the results is a plus