Why we can't fix bias with more AI w/ Patrick Lin

Overview

The video argues that “fixing bias with more AI” is fundamentally limited because AI bias reflects human and societal bias. It also emphasizes that the concept of bias itself is ambiguous and highly context-dependent—so there is no single, universally agreed standard for what “the fix” should be.

What sparked the discussion (Gemini image “scandal”)

• In early 2024, Google integrated an AI image generator into its chatbot (Gemini).
• Public reaction was intense and chaotic. People shared screenshots showing the model produced historically inaccurate images—such as:
  • “Founding fathers” depicted incorrectly
  • Vikings depicted with anachronistic traits
  • A Pope depicted with people of color
• In some cases, the results were highly problematic, including instances where people of color were generated in Nazi uniforms.

Google’s response included a claim that the team tried to avoid “traps” from prior image generators (such as producing violent or sexual content). However, critics on both the left and right accused Google of “overcorrecting,” arguing it advanced “woke” or biased agendas.

Broader takeaway: even if everyone agrees bias is a problem, different groups interpret what counts as a “fix” through their own biases rather than a shared definition.

Evidence that bias shows up across AI systems

The host and guest cite research and reporting showing that bias persists across multiple AI domains:

• A joint study (late 2023) found that text-to-image systems often generated surgeons/trainees as white men, even when real-world demographics differ.
• Washington Post reporting described biases triggered by prompts—for example when users asked for:
  • “Muslim people”
  • “attractive people”
  • “productive person”

The guest also highlights life-altering domain examples:

• Hiring systems (e.g., penalizing applicants based on certain words or college contexts)
• Lending and underwriting (e.g., zip code acting as a proxy for race)
• Criminal sentencing and AI policing
• Healthcare misdiagnoses when training data lacks diversity
• Facial recognition failures when certain groups are underrepresented (including an example from China)

Core thesis from Patrick Lin: AI ethics is an ecosystem issue

“AI ethics” isn’t limited to whether an individual model produces wrong outputs. Instead, it involves:

• Developers
• Users
• Stakeholders
• Surrounding institutions

Bias is also not only about direct legal discrimination based on protected classes. It can arise indirectly through:

• Proxies
• Correlations
• “Implicit bias”

The guest argues that humans are pattern-making, stereotyping creatures with deep psychological and historical roots to bias—making it unrealistic to expect AI alone to eradicate bias.

Why more AI/data often won’t solve it

Several reasons are offered:

• The “hammer” problem: if the only tool you have is AI, you may treat complex social problems primarily with AI approaches.
• Unclear definitions: bias is difficult to define precisely, and “unfairness” vs. “fairness” lacks a single universally accepted definition.
• Non-scalable data sources: training on “less biased” data is not scalable because the data originates from humans and therefore carries human bias.
• Ambiguity and value judgments: machines cannot reliably infer, in every context, what counts as “racist,” “misogynist,” or “unjust,” because those judgments depend on social and moral context.

Nuanced framing: sometimes discrimination may be legitimate

The guest argues that a simplistic rule—such as “never treat protected groups differently”—can create false positives.

For instance, a filmmaker casting actors for a historically specific role (e.g., Martin Luther King Jr.) might legitimately filter by attributes like age, gender, and ethnicity to match historical reality. The point: bias cannot be reduced to one blanket rule without context.

Proposed direction: pluralism and localization (“AI sovereignty”)

The guest supports tuning models to cultural context rather than using a one-size-fits-all approach.

• Different societies may reasonably disagree on ethical priorities.
• What is offensive or acceptable can vary by region and history.

The host adds “AI sovereignty” as a near-term approach: regional systems should reflect local norms, values, and cultures rather than inheriting default assumptions from Western training pipelines.

What individuals (and users) can do

The video concludes that solutions require human-level work—not only model changes.

The host recommends “AI literacy” and responsible usage, including:

• Treat outputs as fallible
• Scrutinize results
• Avoid impulsively amplifying mistakes
• Revise prompts and request better representation when needed
  • e.g., ask for images that reflect the real demographics of a profession

It also argues companies should improve transparency—not only about training data, but about:

• what they are trying to correct
• how they intend to correct it

Presenters or contributors

• Patrick Lin (guest): Professor of philosophy, California Polytechnic State University; Director of the university’s Ethics and Emerging Sciences Group
• Belaval Sadu (host)
• Praka Ragavan (quoted): Google senior vice president (from the Gemini response)