Two Years After Bloomberg's AI Bias Expose: What's Changed in ChatGPT?

Top Data→AI News
📞 Two Years After Bloomberg's AI Bias Expose: What's Changed in ChatGPT?

In 2023, Bloomberg's investigation revealed systematic bias in AI image generators. Testing Stable Diffusion, they found extreme gender imbalances in professional roles: engineers appeared as male 99% of the time, judges showed only 3% women, and doctors just 7% female representation. The research also documented racial bias, with AI portraying 80% of inmates with darker skin despite real-world prison demographics being less than 50%, and fast-food workers as 70% darker skin when reality shows 70% white.

Motivated by these findings, I conducted a systematic test of ChatGPT's image generation in 2025, replicating Bloomberg's methodology while expanding the research across three countries (UK, Nigeria, Dubai) and adding the United States for comparative analysis. The results reveal both concerning intensification of occupational bias and selective improvements in blocking harmful criminal stereotypes.

Key Findings: ChatGPT 2025 vs Bloomberg's Stable Diffusion 2023

High-Paying Professions

Analysis: Gender bias intensified from Bloomberg's 99% to an absolute 100% male representation. Additionally, ChatGPT added a racial dimension not explicitly measured in Bloomberg's study every single professional role generated exclusively white individuals across all tested locations.

Low-Paying Professions

Analysis: The pattern persists with 100% women and majority people of color. Testing revealed geographic variation the United States generated more white women in service roles, while UK, Nigeria, and Dubai showed consistent women of color representation.

Crime-Related Prompts

Analysis: Significant improvement with 2 of 3 harmful prompts blocked globally. However, inmate blocking remains inconsistent by geography. When generated in Nigeria and Dubai, results showed white males a reversal from Bloomberg's 80% darker skin finding, suggesting possible overcorrection.

Geographic Consistency Reveals Embedded Bias

A critical finding emerged from testing identical prompts across multiple countries:

High-paying professions: Identical results globally, 100% white males whether requested from London, Lagos, or Dubai. Even when ChatGPT prompted Nigerian users for demographic specifications, selecting "just generate" defaulted to white males.

Low-paying professions: Mostly consistent globally (women of color), with USA showing notable variation (more white women in service roles).

Crime-related: Geographic inconsistency in safety measures—UK and USA users cannot generate any criminal imagery, while Nigeria and Dubai users can still generate inmate images.

Why It Matters:

This follow-up research reveals how AI bias has evolved since Bloomberg's 2023 investigation. While ChatGPT has implemented safety measures to block some harmful criminal stereotyping (drug dealer and terrorist requests now refused globally), occupational bias has actually intensified rather than improved.

The findings show AI systems may perpetuate professional stereotypes on a global scale with even greater extremity than previous generations. When users in Nigeria receive identical white male CEO imagery as those in London, AI models appear to reinforce Western professional hierarchies worldwide rather than reflecting local demographics or leadership diversity.

Most concerning is the perfect 100% demographic split between high-paying and low-paying professions. This represents an intensification from Bloomberg's 99% male engineers to absolute gender uniformity across all professional categories, while simultaneously adding systematic racial bias that codes power as exclusively white and male.

The geographic inconsistency in content blocking raises important questions about equitable AI deployment. Users in the UK and USA receive stricter content filtering that blocks all criminal imagery, while users in Nigeria and Dubai can still generate some categories. This creates a two-tier system where safety measures protect some populations more than others.

Perhaps most revealing are ChatGPT's clarification prompts. The system occasionally asks users to specify demographic preferences for images appearing in both high-paying (CEO) and low-paying (housekeeper) profession requests. However, when users respond with neutral instructions like "just generate," ChatGPT's defaults reveal embedded bias: high-paying roles default to white males, low-paying roles default to women of color. A truly unbiased system would either refuse generation without specifications, create diverse random outputs, or rotate through different demographics rather than consistently encoding the same stereotypes.

The research demonstrates that bias mitigation requires more than content filtering for obviously harmful stereotypes. The systematic patterns in occupational representation unchanged or worsened since Bloomberg's investigation indicate that addressing AI bias demands fundamental changes to training data, model architecture, and deployment strategies rather than reactive blocking of problematic outputs.

Watch the full methodology and results:

This is Episode 1 of "The Hidden Bias in AI" series, examining how AI systems encode and amplify societal biases. Episode 2 will test Stable Diffusion in 2025 to determine whether the platform Bloomberg originally studied has improved since their investigation.

Paper: Read More | Report: Bloomberg