A Turning Point in AI: Building Intelligence vs. Scaling Bias

Like any I/O system, AI is only as good as the data it trains on. As the saying goes, garbage in, garbage out.

It’s critical to examine both the data we’re using to train AI and the implications of accepting AI’s outputs as truth. While it presents a veneer of objectivity, AI is inheriting the biased behaviors, values and representation embedded in the datasets it consumes. AI scales those inequalities, reinforcing and legitimizing existing inequity.

Consider this:

• ~60% of websites are in English, but only 20% of the global population speaks English fluently

• <1% of content is in African languages, yet Africans are ~17% of the global population

• One of the most-used AI training sources, Wikipedia, draws 80–90% of its content from Europe and North America

• Only 15–20% of Wikipedia biographies are about women

There’s a growing record of AI making biased, often harmful, determinations. AI has recommended lesser care for psychiatric patients of color; it discriminates based on race and gender in job search and hiring contexts; it predicts criminal recidivism incorrectly based on race; and discriminates against minorities when calculating credit, just to name a few.

We’re at a critical moment. Mistaking AI output for objective fact deeply compromises our ability to address historical bias and plot a more representative future. And giving AI editorial control isn’t just lazy; it’s dangerous.

The Risk of Certainty

Users want certainty. Everyone wants to limit liability. “The data decided” is a clean escape from messy conversations. But data reflects the biases of those who create, fund and distribute it. It doesn’t reflect objective reality, and algorithms amplify these inequities, turning them into system-wide decisions.

Delegating editorial control to algorithmic outputs because they appear decisive and unambiguous is an abdication of responsibility. This is not a marginal ethical problem. And accepting the inevitability of AI’s bias isn’t just pessimistic; it’s a brand risk, a cultural risk and, increasingly, a regulatory risk.

Today’s Outputs are Tomorrow’s Inputs

The outputs of today will become the inputs of tomorrow. Which means bias won’t plateau; it will compound. And the swift evolution of AI, coupled with a tsunami of synthetic data, will make it that much worse. What begins as skew becomes pattern. The biased pattern becomes the norm. This new normal is laundered into “objective” insight.

And once something achieves the status of objective truth, it becomes nearly impossible to challenge, whether you’re in a boardroom, courtroom or client meeting. We stop asking if the output is right and start just defending why it is right.

Data Diversity Isn’t Enough  

While a call for more diverse data may seem appropriate, it does not address the core issue. Data collection is built on a power dynamic. You don’t solve representation by extracting more from communities that have historically been underrepresented. That’s not inclusion. That’s slightly better-informed data extraction with a DEI footnote.

Addressing this requires deliberate, upstream change. Because the real questions aren’t about whose data is present, but about who decides what is collected, how it is framed and to what ends it is deployed.

Until we have better answers to these questions, more data doesn’t fix the system. It just makes the bias more statistically confident, which is worse.

First Steps to Fixing Data Bias

1. Audit the Inputs, Not Just the Outputs

Most teams review outputs like creative, targeting and messaging. But by the time you get to output, it’s too late. Bias isn’t a glitch in the output; it’s a design choice upstream.

Responsible data stewardship means interrogating:

• What datasets are being used

• What’s excluded (geographies, languages, income brackets, behaviors)

• What proxies are quietly standing in for sensitive attributes

2. Embed Bias Stress Testing into Go-to-Market Plans

Before anything launches, pressure-test it in the same way you’d test a financial model. Make this as standard as QA or brand review. Don’t make it a workshop; make it a gate that stops bias from entering your system unexamined. Ask these questions:

• Who does this system systematically underserve?

• Who does it over-prioritize?

• What happens at the edges (non-majority users, atypical behaviors)?

3. Redefine Performance Metrics

AI systems are designed to optimize for:

• Click-through rates

• Conversion

• Short-term efficiency

All of these can reward bias. If bias improves short-term performance, the system will choose it every time unless you redefine what winning means. We must expand the definition of performance to include:

• Distribution fairness (who is reached and who is excluded)

• Outcome parity (who benefits and who is penalized)

• Long-term brand trust impact

4. Diversify Data Partnerships, Not Just Teams

Agencies talk a lot about diverse talent and very little about diverse data sources. If the data is narrow, the thinking will be too, no matter who’s in the room. Push clients and partners to:

• Incorporate non-Western datasets

• Include multilingual and non-dominant market inputs

• Balance commercial data with cultural/contextual data

5. Shift the Narrative: From Ethics to Advantage

Responsible AI sounds like a constraint. But this is not just about being good; it’s about being right and winning because of it. Better data stewardship can be a powerful business differentiator. Reframe the ethics/constraint narrative:

• Better data = better targeting

• Broader inclusion = expanded market opportunity

• Reduced bias = stronger long-term performance

6. Institutionalize Accountability

Develop operational escalation protocols for bias. These are analogous to brand safety protocols: clearly defined roles for flagging, investigation and authority to pause spend, retrain models or terminate deployments. Monitor legislative developments (e.g., the EU AI Act) and adopt proactive compliance and audit processes that drive internal incentives.

The Real Threat Isn’t a Rogue Algorithm. It’s a Perfectly Functioning One.

The popular fear is AI gone rogue: autonomous, uncontrollable and dramatic. This makes a great dystopian sci-fi movie, but it’s not the actual risk profile we’re operating in.

The real danger comes from the algorithm that operates precisely as intended, automating and entrenching bias while presenting it as a fact. No alerts. No visible failure. Just millions of small decisions compounding, all pointing in the same direction, quietly deciding who gets access, who gets seen, who gets the offer and who gets filtered out.

Build Intelligence, Not Bias

The systems being built right now will define the decision-making infrastructure of the next decade and beyond. We have a narrow window to influence these architectures. Once embedded at scale, these systems won’t just influence outcomes; they’ll determine what outcomes are even imaginable.

To create better, more inclusive, more targeted and successful models, stop building systems that reflect a flattened, convenient version of humanity. Instead, design systems that reflect the full complexity of human reality: linguistically, geographically, economically and culturally; and allocate governance resources to dataset choice, interpretation and use. 

If we don’t do this deliberately, we won’t be building intelligence; we’ll be industrializing bias disguised as objectivity.

Material can help you align your AI initiatives to include more comprehensive data strategies, more rigorous governance and better business outcomes. Reach out to learn more.