Reclaiming AI for Health Equity: NAACP’s Playbook for Ethical Innovation

Why AI Stagnates Without Equity

Healthcare AI content often mentions bias across race, gender, class, etc. What’s often missing: a guide showing us what we can do about it.

The NAACP flips this script with a health equity AI action plan whitepaper for healthcare leaders: Building a Healthier Future: Designing for AI Health Equity

Their whitepaper is honestly fantastic. As someone who's led health equity AI initiatives since 2019, I've seen how well many of the steps it recommends can work in the real world:

• Building AI algorithms that expanded preventive care outreach to underserved Black and Asian patients for a health system in New Jersey: STAT News

• AI for identifying social determinants of health related to billing codes for users of a digital mental health application

• Ongoing advising for the AI Hub at NYU McSilver Institute for Poverty Policy and Research

This in-depth guide mixes highlights from NAACP’s whitepaper with real-world learnings from building and selling health equity AI solutions in provider/payer markets. These stories come from companies without mountains of cash, so this guide is here to prove that you can launch impactful work without increasing headcount or bureaucracy:

• Health equity AI as a business development and growth strategy

• Embedding health equity AI into the product and data science roadmap

• Coaching commercial and customer teams 

  • Example: The role the health equity approach had in removing skepticism and driving AI adoption from 30% to 100% of hospital clients

Regardless of your company stage or budget, you CAN make powerful steps to achieve your health equity AI mission. This guide shares examples building the teams, community engagement, and evaluation partnerships necessary to make it real.

Before we dig in, you may be wondering, “What does NAACP have to do with the latest in machine learning methods?”

The short answer: EVERYTHING. The long answer…

NAACP's Origins in Data Science

What we call “AI” today is mostly based on statistics, often with methods from hundreds of years ago. NAACP founding members W.E.B. DuBois and Ida B. Wells were pioneers at this.

100 years ago, they were applying data models and data storytelling that outclasses much of the work you see in the field today. Let’s look at their discoveries…

DuBois: Visualizing Black America

When mainstream scientists and statisticians claimed, “Blacks are inferior,” DuBois and his team at Atlanta University produced charts and graphs proving otherwise.

Today, we visualize data using specialized programming languages and advanced statistical packages. In the late 1800s/early 1900s, Dubois’s team conducted this work by hand.

The team analyzed census reports and other data publications to produce the measures for these charts. There’s a powerful irony that they didn’t just disprove racists. They demonstrated the leading thinkers at the time had failed to crunch the numbers properly. As we’ll see in the next section, these failures were devastating for black patients in need of healthcare.

Weaponizing Statistics Against Black Healthcare

Race Traits and Tendencies of the American Negro (Friedrich Hoffman, 1896) was a 330 page tome of statistics and analytics designed to “prove” that the Black race is inferior and disease-prone.

Hoffman, then considered a leading voice or even “genius” in statistics, worked at Prudential, a major insurance institution. Hoffman led policies that did the following:

• Black Adults: Premiums (costs to hold an insurance plan)  held the same as White policy holders, but reducing their payouts by 1/3rd

• Black Children: Held payouts the same, but raised premiums

W.E.B. DuBois and Atlanta University mathematician Kelly Miller attacked Hoffman’s lethally flawed “science”:

“If the population were divided as to social and economic condition, the matter of race would be almost entirely eliminated.” -The Health and Physique of the American Negro - W.E.B. DuBois

DuBois demonstrated Hoffman’s failure to understand cause versus correlation. Hoffman’s analysis was completely in the aggregate. This means he looked at average measures of health across the total population, disregarding differences by city, income class, etc.

DuBois got more specific, comparing health outcomes among people with similar economic characteristics. This “stratification” approach explained away the “inferiority” that Hoffman found.

The most disturbing part is that the causal errors that DuBois called out a century ago are present in the vast majority of today’s machine learning models.

The most disturbing part is that the causal errors that DuBois called out a century ago are present in the vast majority of today’s machine learning models. The cost of these continued errors? Take for instance the 2019 case where a healthcare algorithm was 1/3rd as likely to flag Black patients for help versus a comparatively ill White patient: Dissecting Racial Bias in Healthcare Algorithms

The algorithm used healthcare spending as the stand-in measure for “sickness.” This failed to take into account that people who are sick may not be able to spend, and that mistake hit Black patients the hardest.

Would their machine learning team have made that same mistake under DuBois, who would have been 150+ years old when they built the model?

Fortunately, Causal AI methods have recently picked up steam, but the vast majority of today’s models are still correlative — with grave consequences such as the one above.

Wells: The Horrifying Data Behind Southern Lynchings

In 1892, Ida B. Wells conducted perhaps the darkest and most emotionally draining data exercise of all: Compiling and analyzing data on lynchings of Black men in the South.

We can only look at lynchings with pain and horror today, but at the time, news media would advertise when lynchings were scheduled, and people would bring their families to watch.

These outlets portrayed a false narrative of Black men as rapists and brutes, and these targeted lynchings were accepted as legitimate, even with no evidence of alleged crimes. 

In her newspaper, “Free Speech,” Wells spoke damningly of these lynchings and the lies around them. 

Prominent media outlets called for extreme violence on Wells, demanding she be tied to a stake, branded, and castrated. She was forced into exile from her home, and a mob of leading businessmen destroyed and sold her company.

Despite these attacks and losing everything, Wells was resolute: “They had made me an exile and threatened my life for hinting at the truth. I felt that I owed it to myself and my race to tell the whole truth.”

In an era long before an easy “search” button for this, she built her dataset story by story.

She tracked down newspapers and publications and cross-referenced these with eye witness accounts to get the full picture on hundreds of lynchings. In an era long before an easy “search” button for this, she built her dataset story by story.

She published her findings in Southern Horrors: Lynch Law in All its Phases (1893). It slammed the banner narrative for lynching, which was that Black men were assaulting White women. In her analysis, only one third of the 700+ lynchings Wells examined had those charges (before disputing the validity of said charges). The rest contained all manner of reasons, including political reasons and trivial offenses such as theft of hogs or “being drunk and ‘sassy’ to white folks.”

Her follow-up publication, “A Red Record,” further outlined and visualized these records.

Health Equity AI In Practice: NAACP’s Steps to Reclaim the Algorithm

Wells and DuBois fought loud and proud racism, but today, healthcare leaders are up against a more insidious discrimination. The echoes of the history outlined above—and continuing decades of bias and miseducation—show up in healthcare data in ways that are hard to identify.

To help us see these patterns more clearly, NAACP’s whitepaper asks us to look at AI as a sociotechnical system, rather than the purely technical framing it’s had in the past. It then uses 3 major sections to outline the health equity vision:

• AI for Health Equity: Sample Illustrations, Risk, & Opportunities

• Health AI Governance: Health Systems

• Maternal Health & AI: A Case Illustration

This companion guide covers the first two sections with personal examples from working on AI technologies within digital health & digital mental health companies since 2018. These stories reflect the experience where money, time, and team sizes are tight. 

The third section is a powerful model example of what NAACP’s governance framework looks like in action.

Solving the Real World Data Void

A great deal of healthcare AI depends on the notes, images, conversations, billing codes, diagnoses, biomarkers, and procedures recorded at times when people seek care. 

Real World Data

There’s no clean line between perceiving a medical need and seeking care.

Barriers like cost, coverage, and culture aren’t equal across ethnicities, genders, and sexes. As a result, there is less real world data from disadvantaged communities. 

Trial Representation Data

NAACP’s whitepaper discusses how a bias toward conducting research on white male populations further biases healthcare data. If models are trained on research that fails to represent clinical needs across race and gender, how might that impact the decisions they guide?

Innovation Crossroad - Beat or Bolster Bias?

Despite the biases we’re discussing, NAACP’s whitepaper discusses the life-saving possibilities from AI use cases: predicting the onset of various diseases, helping overcome healthcare literacy barriers, support in drug manufacturing processes, claims processing, etc. 

The biased history and data isn’t a stop sign to innovation, it’s a crossroad.

“Algorithms can reinforce structural biases - or fight against them. Which one is up to us, and the small seeming technical choices we make when building them.” -Dr. Ziad Obermeyer

Do we reinforce structural biases or fight against them? You can select actions and implications below.

For those who selected option three, the NAACP whitepaper walks you through 6 areas of health equity AI application:

• Patient Engagement with AI

• Healthcare AI Views from Safety-Net Professionals

• Use of Large Language Models within Health Care

• Synthetic Data, Health Equity, & Pharma AI

• How Pharmaceutical Companies Approach Health Equity & AI/ML

• AI & the Future of Clinical Trials

These sections describe the problems, and the prescriptive “What types of actions should we take?” parts come in the later sections.

As someone working with these models primarily in the healthtech, payer, and provider spaces, I found some sections to be a good refresher, and others to be a great lens into parts of the healthcare industry (pharma) I’m less familiar with.

But the section I found most powerful highlighted research from over 230 safety-net hospital clinicians.

Why Hype is Not Enough for Safety Net Professionals

In a 2025 study, researchers at The University of Texas at Austin sought health AI perspectives from providers who serve uninsured, marginalized patients.

They summarized key findings in a table, where the % figure represents the proportion of surveyed providers endorsing this view:

This reinforces how, in healthcare AI, we should gather input from our users and stakeholders. This is especially true for safety-net providers, who often don’t have a seat at the table.

And from personal experience, these ring true at many types of organizations that are outside of safety net as well. So, the better you are at addressing administrative needs, privacy/security, and training for safety net providers, the better you will be at addressing these same needs for your whole client base!

In the benefits column on the left, note the large margin of professionals who want to spend less time on administrative tasks.

In the barriers column on the right, note the second entry, “Insufficient Staff Training/Knowledge.” In building for healthcare, this is where you’ll spend the most time. Today’s AI tools all deal in uncertainty and probability, rather than absolutes. This means different trust and training needs for AI tools vs traditional software.

A common question you’ll hear among healthcare AI vendors struggling for clinical/healthcare leadership adoption: “How do we get healthcare to trust our solutions?”

The first and most important step is to be worthy of their trust.

NAACP’s whitepaper extensively outlines HOW to be worthy in the section on AI governance.

Health AI Governance

Progress in Healthcare AI is Too Slow

People often discuss the rate of progress in AI as “fast” and “overwhelming,” but I argue the opposite, that it is slow and underwhelming.

Consider these three questions:

• Are there widespread algorithms that engage people in healthy behaviors?

• Have commercial chatbots created a widespread improvement in mental health?

• Can care providers and organizations readily match people to available community based supports?

The answer to all three questions is currently no. This is where you, the healthcare innovator, come in. It’s up to you to build inclusive systems that address the gaps in our healthcare. The latest headline releases from Google, X, and Anthropic don’t—and can’t—drive this to the finish line.

Worthy of Trust: Cracking the Code for Healthcare Leader Buy-in

Years of working with clinicians, behavioral health professionals, and healthcare leaders has taught me that, even if hype opens doors, it doesn’t drive healthcare adoption. Is it because healthcare hates progress and tech leaders are far ahead of them in their thinking?

Not necessarily! It’s more that they’re making decisions that impact lives, licenses, and lawsuits. Demonstrating grounding in ethics, responsible AI, and health equity are key to working with people who take their community’s health seriously. 

What’s being presented instead?

• Models that generate decisions and outputs with no explanation as to why

• Insistence on “trust” from companies who are unwilling to share performance details overall or by subgroups such as ethnicity, sex, and gender

• Lack of investment in well-designed, inclusive studies/metrics to measure impact

  • Clinical impact

  • Financial returns

  • User experience & time savings

• Sales teams who have a hard time articulating value, safety, and effectiveness and can’t provide answers when pressed on common health equity questions. Examples:

  • How diverse was the data that this was trained on? (I’ve frequently seen this one)

  • How does the impact vary by community?

  • How does this help us account for access barriers?

• Lack of third party evaluation for health equity & performance standards

This isn’t to say that healthcare’s incumbents are free from blame in the lack of progress. Major hurdles healthcare needs to overcome for meaningful innovation:

• Vendor lock-in patterns, where innovation is delayed indefinitely until firms like Microsoft, Epic, or Oracle offer a solution

• Limited or siloed evaluation of health equity standards, rather than treating health equity as core to delivering great care and service

• Pursuit of “AI” as a check-the-box initiative with no specific outcome or equity goals

  • Opposite version: blanket bans on AI purchases without considering the wide nuances and differences between use cases & vendors (scribes vs chatbots for example)

Solving the above issues isn’t the panacea for healthcare AI adoption, but it does drive serious progress. AI isn’t new anymore, and sophisticated teams are demanding transparency and responsible approaches from their vendors.

I’m a big proponent of moving fast, but if you want to rush into healthcare without governance, you’ll “move fast” into bankruptcy.

Governance as a Growth Tool

Governance is often seen as a barrier to innovation, but done properly, it’s an accelerant for growth!

For tech leaders: Healthcare leaders and policy makers are EXCITED when health AI tools drive equity and inclusion. For example, following a “Reclaim the Algorithm” presentation with NYU McSilver Institute at the Puerto Rico’s SOMOS Policy Summit, leaders told us they changed their opinion from blocking AI initiatives to embracing thoughtful deployments.

Amidst all the hype and BS in AI, you have the chance to cut through the noise and reshape the possibilities these leaders see in sociotechnical solutions. NAACP’s governance framework is a powerful starting point.

The framework and language could be intimidating for cash-strapped innovators who likely don’t have a Chief Ethics Officer or internal fairness auditors.

If this is you, don’t let it stop the show. In my work in the field, we didn’t always have the resources to hire more people and create new departments. We turned this into a strength. Health equity became a core offering in our products, not a separate faction. Keep that in mind with the building blocks of health equity AI governance.

The following sections cover NAACP’s recommended governance layers, and how these have looked in real world projects:

• Ethical and Normative Layer

• Organizational Governance Layer

• Operational and Lifecycle Layer

The Third Party Validation Layer is not an official layer in NAACP’s framework, but having an external partner is critical to health equity initiatives. 

Ethical and Normative Layer

This is where leadership establishes health equity AI as core to the company. NAACP’s whitepaper calls for establishing mission alignment, setting standards that promote explainable AI, and building the governance team:

• Translate equity goals into Key Performance Indicators (KPIs)

• Ensure clinicians and other stakeholders understand AI inputs & outputs

• Establish clear accountability for succeeding or failing at equity goals

However, this is difficult from day 1, even if you have “C” in your title. Here are some tips that should work no matter where you sit in the company hierarchy.

Find Collaborators on Every Team

At Actium Health, our data science team’s consensus, all the way up to the CTO, was that we should investigate and mitigate bias in our algorithms, but that wasn’t nearly enough.

To enact practices such as AI model “nutrition facts” (descriptions of how well models perform and the data that informs them), bias detection/optimization, and interpretability research, we had extensive 1:1 conversations with product, marketing, sales, and customer success.

Small, focused conversations allowed us to educate on bias, build consensus for the investments we’d need to make to address it, and get support for the initiative:

• Product Support: Incorporating health equity practices into the product planning pipeline

• Clinical Support: Clinicians in your company can help you validate whether your use case actually promotes health equity, how it may fit in a workflow, and barriers to adoption other clinicians may see

• Marketing Support: Preparations on case studies, webinars, and conferences to educate the market on these practices

• Sales Support: Partnering with their team on how these practices could help win deals and adoption

• Customer Success: Driving client adoption of health equity enhanced models and practices

The NAACP whitepaper calls for governance structures, transparency, and mission statement alignment. This collaboration helps diverse teams, technical and non-technical, to understand and contribute to the health equity AI goals!

Tie Health Equity AI KPIs to Growth, Revenue, and Expenses

It will never be enough to focus on health equity because of its virtues, ethics, or clinical outcomes without a clear business case. This work requires additional investment and focus, so you have to demonstrate what the return is for those efforts. Examples of supporting business cases:

• Some billing codes, such as HCPS G0136, relate directly to health equity, establishing direct ROI for these efforts

• Many of your health system partners commit to missions for diversity and inclusion. Identify opportunities for your systems to help grow into underserved communities. The methods discussed in the STAT News article helped us expand client AI adoption

• Your team is recognized as a thought leader in an industry starved for stories on health equity innovation with provable impact

Example Goals/Metrics:

Internal Data Science Team Metric: Improve multiclass accuracy or recall by 25%

KPI: Reduce client population health disparity by 5%

KPI: 4 speaking sessions regarding health equity accomplishments

KPI: 200 inbound leads to health equity innovation whitepaper

Business Goals: Showcase governance and inclusive practices to reduce churn

Organizational Governance Layer

If the “Ethical/Normative Layer” is the foundation, the organizational governance layer is where the work happens. NAACP’s guide calls for equity impact assessments, forming an AI ethics & equity review board, and forming a data governance council.

Equity Impact Assessments (EIAs)

In an AI ethics partnership with University of Chicago Booth’s Center for Applied AI (CAAI), the first exercise the CAAI led was a full algorithmic inventory. This was to account for potential health equity challenges wherever decisions are automated, instead of just focusing on AI models. 

Taking inventory consists of documenting where algorithms are used, identifying how they’re used, identifying the types of bias they may perpetuate, and speculating on the opportunity to improve health equity. The length of the inventory may be overwhelming, but an easy first step toward prioritizing the list is to identify what has the biggest impact to the KPIs from the ethical layer.

Before asking how well any particular algorithm is performing at its task, this is the opportunity to ask, “How does it impact a user if the algorithm performs well, and what’s the impact if it fails?”

Your team may find that some use cases have a very high impact on equity, such as clinical predictors, whereas others have a low impact.

Review Board and Data Governance Council

In the ethical layer, you’ve done the work of communicating across teams. You can now work with those leaders to identify the right coalitions and cadence for these overarching review processes.

This review board doesn’t just start and stop with your company, it should involve the communities you’re building for. Without community input, developers, researchers, and designers are making assumptions about people with vastly different lives and income brackets.

Companies like Modality, a company that helps train Community Health Workers by allowing them to interact with personas, depend on these approaches. Their development process involves careful conversations and feedback from members of the communities they serve.

Community interaction eliminates stereotypes, archetypes, and assumptions and forces a focus on what’s meaningful and real to the people you’re building for.

Operational and Lifecycle Layer

This layer calls for inclusive data practices, applying fairness metrics within models, and conducting monitoring and auditing. The rubber meets the road as you’re evaluating the actual output from your models.

NAACP’s Health Equity AI guide calls for mainly technical stakeholders at this stage, but don’t forget that the work is not finished just because the models get high marks on health equity measures.

Every technical accomplishment needs a corresponding translational accomplishment for these approaches to work and sell in the market. Examples:

AI Practitioners may see the largest challenges here - translating their work and its outputs in a way that makes sense to business and clinical stakeholders.

It’s difficult, but critical, as a simple misunderstanding or blackbox approach is enough for companies to abandon thousands of hours of work.

Third Party Validation Layer

Throughout this guide, I referenced third parties, such as the University of Chicago, that helped lead algorithmic audits as well. Trusted third party evaluation is the difference between rigorous, good-faith evaluation or simply giving oneself a pat on the back.

The NAACP doesn’t explicitly identify this layer, but I believe that no matter how good or well-meaning your team is, outside analysis will elevate their work. In algorithmic auditing, there is too much risk that your team can’t see past their own norms and standards.

AI use cases vary, and so does the expertise needed to judge them. So, when identifying a third party auditor, be sure to check for healthcare experience, their familiarity with your types of models, and their familiarity with the market you serve.

Final Thoughts

NAACP’s Building a Healthier Future: Designing for AI Health Equity is the most detailed guide currently available  for health equity AI practitioners and leaders.

This companion guide adds some additional color, but the most powerful examples, such as the case study on health equity in maternal health AI, are in the main guide.

Please keep NAACP’s guide and this document in mind if you find yourself stuck/challenged in health equity AI initiatives. Recall the options above:

• Purchase and build solutions without health equity in mind

• Ban AI solutions

• Build AI with a health equity framework

The first two options are the most common in the status quo. They’re tantamount to doing nothing at all.

The third option elevates the experience for underserved patients and positions you and your company to set the standard in your corner of healthcare. Choose wisely!

Collaborators

Editor: Michael Barett

Scientific Review

Kara Emery, PhD - Director of Data Science, AI Hub at NYU McSilver Institute

Dr. Kay Nikiforova

Special thanks to the NAACP and Dr. Craig Watkins, lead author of this guide.

Disclosure: In September 2025, I participated in a paid speaking engagement at an NAACP event, but this companion guide was not in any way funded by NAACP.