Can AI Tell Whether You're A Criminal From Your Face? | Machine Learning and Physiognomy

The video discusses the resurgence of pseudosciences like phrenology and physiotomy in modern AI applications, particularly in predicting criminality based on facial features. It critiques the ethical and methodological flaws in such studies, highlighting the dangers of perpetuating biases and stereotypes. The video also examines a controversial preprint claiming high accuracy in criminal prediction using machine learning, which faced significant backlash. It emphasizes the inherent flaws in these studies due to biased datasets and the potential misuse of AI in law enforcement and other areas, urging caution and ethical considerations.

What is phrenology primarily concerned with?

• Studying palm lines to predict future events

• Measuring bumps on the head to determine mental traits

• Analyzing handwriting to assess personality

• Predicting criminal behavior through facial features

Why was the preprint claiming to predict criminality using facial images criticized?

• It had no practical applications

• It was too expensive to implement

• It used outdated technology

• It was considered ethically unsound and poorly executed

In the thought experiment, what does labeling all images as 'criminal' imply?

• That the person has committed a crime

• That the person is destined to become a criminal

• That the person has been arrested

• That the person has been convicted

What is a common application of physiotomy in the 1800s that continues today?

• Predicting job performance

• Determining educational potential

• Assessing mental health

• Confirming biases about criminal tendencies

What is a major flaw in using mug shots for machine learning datasets?

• Mug shots are not evidence of a crime

• Mug shots are difficult to process

• Mug shots are always of low quality

• Mug shots are too similar to non-criminal images

What is a potential consequence of using flawed machine learning systems in law enforcement?

• Perpetuation of biases and stereotypes

• Increased accuracy in criminal predictions

• Improved public trust in technology

• Reduced need for human oversight

What does the critique of these studies suggest about the datasets used?

• They are universally applicable

• They are scientifically validated

• They reflect cultural biases and perceptions

• They are comprehensive and unbiased