MPhil research on AI's perception of animacy and vulnerability

Compare AI and human neuroimaging data on animacy and biological vulnerability, integrate brain sensory representations into AI, and publish open-source biological validation guidelines.

Endorsed by+1

Compare AI and human neuroimaging data on animacy and biological vulnerability, integrate brain sensory representations into AI, and publish open-source biological validation guidelines.

Endorsed by+1

Project Details

During the MPhil studies, I will conduct research with a comparative analysis of current AI systems of animacy, physiological and biological vulnerability of living beings and compare it to neurobiological human data. Using human visuo-semantic neuroimaging datasets, I will integrate the sensory representations of the human brain into AI systems and test their alignment. I am to produce neuroscientifically based learning algorithms that can perceive the fragility of biological organisms and the precipice of animacy (livingness) of different entities: humans, animals, insects, and other organisms.

The research will be carried out at Prof. Kamila Jozwik’s lab at the University of Cambridge. Prof. Jozwik specializes in biologically-inspired deep learning and AI models. My previous research includes studies of human-machine interactions and the usage of AI chatbots in mental health. In this study, I saw that AI is not ready yet to fully understand the emotional spectrum and sometimes even physiological boundaries of humans. It inspired me to conduct the proposed research on AI’s concept of animacy and contribute research findings on how neuroscience can help us build more empathetic and intelligent AI systems.

The research will be conducted within 12 months, where the first 9 months will be for full-time work on the research. The last 3 months will be for writing AI design frameworks for tech companies based on the results of the experience and sharing them with the AI safety community through the Leverhulme Centre for the Future of Intelligence in Cambridge.

The end goal of the project is to develop guidelines for biological validation of AI models and publish them open-source.

Theory of Impact

The goal of my study will be to investigate how AI's internal representations are misaligned from human ones, which could lead to higher risks of physical harm to living things by AI in the future. Currently, many companies are integrating AI in various fields, including medicine, mental health, and psychological counselling. However, AI still can’t completely understand the fragility of human life, which can lead to injuries and even death.

With neurobiological validation of these concepts, we will be able to teach AI to be more perceptive of the vulnerability of living organisms and improve the safety of neural models in risky fields.

People

Elena Mishina

Team Member

Grants Received– no grants recorded

Add grant

Funding Asks

Cambridge MPhil research on AI's perception of animacy and vulnerability

Open

I’m seeking funding for UK student visa and travel, which will enable me to come to the University of Cambridge, UK, to conduct my MPhil research on how AI systems internally represent biological and psychological vulnerability

Minimum

$1,990

Ideal

$1,990

What this ask is

I’m seeking funding to cover UK student visa and travel costs, which will allow me to begin my MPhil research at the University of Cambridge on aligning AI systems’ internal representations with neurobiological data of cognitive processing of distress, vulnerability, and animacy concepts. This project develops a methodology for comparing human brain-imaging responses to how AI models internally process the same stimulus through representational similarity analysis. The outcome of this work will be an open toolkit with research-based guidelines for neurobiological ground-truthing to evaluate AI systems on their understanding of psychological and biological vulnerability states.

Theory of impact

This project will test if AI models’ internal representations are aligned with human perception of vulnerability or only mimic it. If AI systems diverge from biological reality, then we risk creating a misaligned AI that can be dangerous to implement in high-stakes fields, such as medicine, mental health, and emergency response, where even a slight mistake can lead to the death of a person. Neurobiological validation of models can provide a much deeper and detailed measurement of what is happening inside AI systems. It can help compare how AI processes information to human cognitive processes, assess how AI learns, and ensure that AI’s internal processing is in the same direction as human ones for its safe deployment.

How the money will be spent

The funding includes:

£558 ($755) fee to apply for a student visa from outside the UK,
£776 ($1050) fee for an immigration health surcharge (student rate),
£122 ($165) for a one-way economy flight ticket from Budapest to London with checked luggage,
£15 ($20) for a one-way economy bus ticket from London to Cambridge with extra luggage fee.

Discussion

Katja GorlinskiEndorsement1mo

Hello Elena,

You you rise a topic I find quite upstream. It is true, before any harm prevention regulation comes into play, there is the need for the system to recognize that there is something fragile in front of it which can actually be harmed. As far as I understand, you test the correctness of such a recognition through human brain imaging data, and therefore it becomes verifiable.

This problem was assigned to philosophy from the beginning, I guess.

I think you found an innovative solution to an extremely important and challenging risk.

One question came to my mind: what makes the right marking of user's fragility influence the behavior of the model? Recognition is different from caring, and I wonder where the point of contact between them is. Perhaps you see it yourself already.

I support your project.