Loading summary
Transcript1 lines
- [00:01]
A
Imagine standing in front of a mirror every morning for 20 years and seeing nothing but a void. Then overnight, the mirror starts talking. It doesn't just tell you that you're there. It tells you that your skin's hydrated but lacks the glass like perfection of a beauty advert. It tells you your jaw is too long to be objectively beautiful in your culture. It scores your face out of 10. For the blind and low vision community, this isn't a speculative scenario. It's the current clinical reality of multimodal language models. This transformation is captured with striking clarity in a recent BBC report by Milagros Costabel, a journalist who also has blindness who shares her very personal experience using AI as a visual surrogate. Costabel's account highlights a shift in assistive technology recently where AI has transitioned from simply identifying something like a tin of soup to providing a subjective, emotionally charged critique of the human form. It a really excellent, insightful, thoughtful piece which raised numerous points I hadn't considered but are undoubtedly really important. I'll summarise some key themes here now, but would encourage anyone to read the original article for full detail and context. For clinicians and health tech leaders, this development signals a new frontier where technology impacts patients psychology as much as it assists with their daily living. For decades, assistive technology for the visually impaired community was strictly functional. In 2017, the main utility was something called optical character recognition. This helped with things like reading mail and identifying medication labels. The output was very factual and binary. However, as Costable outlines, the integration of advanced models like GPT4 Vision into apps such as the one that she uses called Be My Eyes has changed the nature of the feedback. It's moved from what is this thing? To how do I look? This is a move from data processing to social interpretation. For the first time, blind users have access to visual economy the subtle cues of fashion, grooming and facial expressions that define model social interaction. This access is inherently empowering, providing a level of independence in self presentation that was previously impossible. Yet, as we'll explore, this newfound agency is currently tethered to algorithms that lack human context and objective neutrality. The primary concern for health professionals lies in the psychological impact of these AI mirrors. Costervald describes a skincare ritual where the AI provides very critical feedback, comparing her skin to the almost perfect example of reflective skin seen in advertisements. This is a direct injection of idealised Western beauty standards into the internal monologue of a user who can't visually verify the claim. Psychological research, including work by Helena Lewis Smith at the University of Bristol, suggests that frequent feedback about body image correlates with lower self satisfaction. Sighted individuals are exposed to the same idealized images, but they also have lots more visual context to recognize. Filters, lighting and staging. A blind user, by necessity must trust the AI's description as a baseline reality. When the AI describes someone's jaw as elongated or a face as jarring, it isn't delivering a medical diagnosis, it's regurgitating a biased training set. The risk of inducing anxiety, depression or body dysmorphia or tendencies towards any of those is a hurdle that needs to be addressed before these tools are recommended within clinical pathways or for regular daily use. Why does AI offer such specific critiques? It's a product of its training data. Most multimodal models are trained on Internet scale datasets which are historically skewed towards Eurocentric, thin and youth oriented people. Beauty Standards when the AI attempts to define what is traditionally beautiful, it isn't using some universal metric. It's using a statistical average of the images that it's seen most often. Merrill Alper, a researcher at Northeastern University, points out that the AI lacks an understanding of subjectivity and individuality. It treats beauty as a solvable equation rather than a contextual human experience. For a blind user, this means that the mirror they're using is a pre calibrated tool to favour a specific demographic. And this highlights an important need for empathy. First design in medical AI, where the model recognizes that beauty and other things are not a fixed single data point. Beyond bias, there's also the persistent issue of hallucinations, where the AI presents false information as fact. In a clinical or assistive setting, they're not just errors, but they're breaches of trust. Costabel cites the experience of Joaquin Valentinozzi, who used AI to select photos for a dating profile. The AI misidentified his hair color and incorrectly described his facial expressions, labelling a smile as a neutral expression. In a social context, these inaccuracies are very significant. Choosing a profile picture based on a hallucinated description can lead to profound insecurity. If the technology is meant to act as a sensory prosthetic, it must meet a higher standard of reliability than a standard consumer chatbot. Currently, the technology is showing us what's possible, but it's not yet showing us what is consistently true. To move these sorts of innovations from a high risk consumer tool to a reliable clinical asset, several hurdles need to be overcome. First, we need better communication about certainty and confidence. For assistive AI, the interface should communicate when the model is guessing or when lighting conditions make a description unreliable. Transparency about the model's limitations is really important for managing user trust. Next as mentioned earlier, I think the next leap in the field isn't just about a better description, but about historical context. An AI should not just describe user's current state, it should compare it to the user's own history. Instead of saying your skin's not glass like a sophisticated model should say your skin looks more hydrated than it did last Tuesday. This moves the benchmark from an impossible societal standard to an individual's own health baseline. And more generally, we need to think a bit more carefully about the interaction between where AI and humans can perform tasks. AI can be a helpful preliminary filter, but not necessarily the final word. So the rise of AI mirrors as detailed by Milagros Costobel, represents a monumental achievement in accessibility. It's demonstrating that large language models can function as powerful sensory prosthetics, offering the blind community a seat at the table of visual culture. The sheer scale of the Be My Eyes volunteer network and the integration into wearable tech like Ray Ban's metaglasses shows a clear path towards total independence. But we do need to be clear sighted about what this current iteration does and doesn't show. It shows a technology in its infancy, one that's prone to bias, susceptible to hallucination, and capable of inflicting psychological harm through unfiltered critique. It doesn't yet show an objective truth of the human form. The future of health AI in this space could be bright, provided that we approach it with a pragmatic focus on bias mitigation and user psychology. Carefully and thoughtfully, with patients input, we're moving towards a world where the mirror doesn't just talk, but understands. Until then, our role is to guide patients and users through this transition, ensuring that we use algorithmic mirrors as tools for empowerment, not as arbiters of self worth. Again, I'd really recommend reviewing the original article, which I'll put a link to in the notes. It's a really good, insightful piece.