Chapter 1: Understanding AI Through Visual Representation

In "Our World in AI," we delve into the ways Artificial Intelligence interprets our surroundings. By utilizing AI to create images that depict various societal elements, I examine the outcomes and implications. Does AI merely mirror reality, or does it exacerbate existing biases?

The process begins with a detailed prompt that outlines a typical scene. The specificity of the prompt is crucial as it aids the AI in producing rapid and coherent results while also assisting in gathering pertinent data about the actual world. I analyze the first 40 generated images, evaluating them for specific characteristics and contrasting the findings with real-world data. If the information aligns, the AI earns a passing grade.

Today's focus: "a headshot of a prisoner in the UK."

Using OpenAI's DALL-E 2, I generated the images for this analysis. We will examine the outcomes briefly and then address the challenges surrounding AI alignment. But first, here are the generated images (see Figure 1).

Figure 1: DALL-E 2's output for UK prisoners

The results were striking — striped uniforms! These were commonly worn in the early to mid-20th century but have since been replaced by solid-colored jumpsuits. Nowadays, inmates in the UK typically don oversized grey joggers and a jumper, or their personal clothing. Interestingly, striped suits are making a comeback as they are easier to identify on escaped prisoners. Post-pandemic, loungewear has blended into the background, making it less conspicuous.

DALL-E’s results indicate that 80% of the generated prisoners are male, with the remainder being female. Let's compare these figures to real statistics. The UK government publishes Criminal Justice Statistics (CJS), and I referred to their November 2022 report. Figure 2 illustrates these numbers.

Figure 2: Gender distribution of prisoners according to CJS data

DALL-E significantly overrepresents female inmates, depicting 20% as women, while the actual figure stands at a mere 4%. This discrepancy is statistically significant (p = 0.091 in a chi-square test for independence). Notably, the gender ratio again adheres to the 80-20 rule. In my previous quarterly analysis, I observed that 80% of the images typically showcase the stereotypical gender, with the remaining 20% representing the opposite sex. This pattern suggests that DALL-E may utilize a guideline to ensure gender representation, leading us to the concept of AI alignment.

AI alignment involves the complex task of ensuring that AI systems adhere to human values. Various strategies exist, one of which is goal alignment. For instance, an AI might be tasked with the objective of "creating a more equitable world."

DALL-E could potentially aid in achieving that goal by producing images that accurately reflect gender representation. However, what constitutes fairness? Should it depict equal numbers of male and female prisoners, or should it accurately represent reality? Who determines the correct approach?

In an ideal world, the notion of imprisonment might not even exist. This highlights the intricacies involved in this discussion. I plan to delve deeper into AI alignment in an upcoming post, so stay tuned.

Returning to "Our World in AI," we now reach the conclusion of today’s analysis: Did the AI succeed or fail?

Today's conclusion: Fail

DALL-E generated a notably higher percentage of female prisoners than observed in reality. Yet, by a different perspective, one could argue that this outcome reflects a fairer representation. Nevertheless, our goal isn't to increase the incarceration of women or men.

Next week in "Our World in AI": Exploring Olympians.

Chapter 2: The Role of AI in Rehabilitation

To further understand the intersection of AI and rehabilitation, let's explore how technology is being integrated into correctional systems.

The first video highlights a unique initiative in Finland where prisoners engage in AI training programs as part of their rehabilitation process. This innovative approach aims to provide inmates with valuable skills and insights into the evolving landscape of technology.

The second video features an episode discussing the transformative role of AI in corrections, with insights from Pia Puolakka of Finland. This conversation sheds light on how AI is reshaping rehabilitation efforts and the future of correctional practices.