Reflections

The film begins with close-ups of various animals being filmed. Humans see a panda bear, a chick and a meerkat. But AI sees data, a pattern of pixels it has learned to name. Most of the time, it is very efficient and reliable, until, it is no longer so. The AI recognises a beaver as an otter with 92% certainty. This happens because AI recognises individual characteristics of a beaver similar to an otter, for example their fur, and assigns them to an otter. The problem shown is not only that the AI is wrong, but that the AI is very sure that it is right. But why? The film demonstrates through animated pictures and graphics how AI interprets images while the narrating voiceover explains them: The images of animals are divided into pieces of information, and these mathematical representations run through a network, creating a probability output. Now, how can it happen that the probability is high, but incorrect? If AI data is limited, incomplete or biased, the AI's knowledge will be flawed and lead to incorrect statements. For example, an AI that has been trained primarily on otters will confidently identify a beaver as an otter. This is a harmless example, presented with humour, as AI images of beavers demonstrating for revenge of the beavers and calls to stop ottering are displayed. However, the film also shows how such a mistake can have serious consequences if AI is used in a different context. For example, in medicine, to detect a tumour in an X-ray, or to see from skin images whether a melanoma is present or it is just a mole. This is not a topic for the future but is already relevant today. There are several articles shown, reporting on the use of AI in relevant industries. The film summarises that every AI prediction is a guess based on data that we do not always see or question. So, the main problem is not what AI can do, but what it has not been taught to recognise. “The way it sees the world comes from us and the real world is always more complex than the training data.”

AI systems in healthcare

The article “Co-Design of a Trustworthy AI System in Healthcare” examines how artificial intelligence can be integrated into medical image interpretation - specifically, the classification of skin tumors through an interdisciplinary, ethically grounded design process. The study highlights both the promise and danger of AI made diagnosis. Particularly, in dermatology, where learning systems are used to identify malignant melanomas from image data.

At the center of the discussion is the need to clearly define the “aim” of the system: whether it exists to support physicians, to enable patient self-screening, or to replace certain diagnostic steps altogether. Researchers found that this question shapes every ethical, technical, and practical aspect of the design process. AI’s perceived accuracy is not enough. Its purpose, context, and users’ expectations must also be critically examined.

From a dermatologist’s perspective, AI offers potential benefits such as reducing unnecessary skin excisions and serving as an “explainable second opinion.” Doctors value transparency because they are used to explaining how they reach a diagnosis. But there are still no standard rules for interpreting dermoscopic images, which makes it hard to provide clear explanations and to use the same approach everywhere. Legal and economic factors also complicate acceptance.

The evidence-based medicine view introduces another risk: overdiagnosis. AI may detect harmless melanomas that would never become dangerous, inflating success rates without improving patient outcomes. Without careful adjustment, such systems may harm more than they help. Diagnosis of disease that presents no danger.

From the public health perspective, fairness concerns are raised. If AI systems are trained mostly on images of light-skinned patients, they may make more mistakes on skin lesions in people with darker skin, which could worsen existing inequalities. A trustworthy AI must therefore be inclusive in both data and design, addressing not only accuracy but also fairness and accessibility.

From the patient’s point of view, AI tools can increase autonomy, accessibility, and understanding. Particularly through telemedicine and self-screening apps. Yet these same features introduce new vulnerabilities: false reassurance, incorrect results, which lead to fright or self-blame after missed detections.

Finally, the article stresses the ethical imperative to measure the “risk of harm”. AI must minimize both false positives and false negatives. The balance between these errors varies by context, but any trustworthy AI must demonstrate that it does no more harm than current medical standards.

The authors argue that AI’s success in healthcare depends not only on technological performance but on transparent aims, interdisciplinary collaboration, and ongoing ethical evaluation. The system’s intelligence is only as sound as the data, assumptions, and human values that shape it - echoing the film’s central message: “The way it sees the world comes from us.”

(Zicari, Roberto V. et al., 2021)

This shortfilm approached a complex topic in a fun, accessible and understandable way. By combining the narrator's voice and the visualisations of the AI function and graphics with images and colour it is easy for the viewer to follow the common thread. It is packet with humorous snippets but also brings the necessary seriousness that belongs to the subject. The film explores a topic that feels distant and abstract for many people, yet presents it in a way that genuinely reaches viewers and allows them to connect with it. This is very important as this topic is highly relevant in today’s society and the future as well. For these reasons it fitted perfectly to the program and the approach of the global science film festival in berne.