How RankRx Meets the AAMC's Principles for Responsible AI in Residency Selection

1. Balance Prediction and Understanding

AAMC's Concern:

AI should predict success while staying transparent and interpretable. Systems must measure qualities that actually align with an institution's definition of an effective student or resident.

Questions:

• How do you ensure the characteristics measured align with our institution's definition of an effective student or resident?
• How do you balance the complexity of your tool with the need for interpretable results?
• (Follow-up) How does your AI handle different data sources (e.g., academic, clinical, documents) in its decision-making?
• (Follow-up) Can you give an example of making your tool's output understandable to nontechnical users?
• (Follow-up) How do you incorporate our subject matter experts into the model-building and interpretation process?
• (Follow-up) What validation methods and metrics do you use to ensure large language model (LLM) outputs are accurate? Walk us through how you detect and prevent hallucinations or factual errors.

How RankRx Addresses This:

At RankRx, we don't decide what "success" looks like for your program, you do. Each institution defines its own success factors (e.g., USMLE scores, clinical rotations, prior research, volunteer service, or other custom experiences). RankRx then provides the framework for you to assign points, weights, and thresholds to those criteria. This ensures that all measurements align directly with your program's mission and values, not with a generic or external definition.

To maintain interpretability, RankRx avoids opaque "black-box" scoring. Every decision made by the system can be traced back to the inputs and weights your program defined. For nontechnical users, we generate clear explanations such as: "This applicant received X points for U.S. clinical experience and Y points for specialty-aligned research, based on the program's scoring rules."

Our system also supports multiple data sources, whether academic performance, clinical evaluations, or document-based materials like letters of recommendation and CVs. All are parsed, structured, and presented in a uniform, customizable scoring model.

Importantly, subject matter experts are involved from end to end: faculty and administrators define the metrics, review scoring outputs, and refine the system for each application cycle. RankRx acts as the engine that applies these definitions consistently, at scale.

Finally, to ensure reliability, we address the risk of LLM inaccuracies by grounding every output in program-defined categories and point systems. In addition to providing summary of text information, our system extracts and displays information exactly as it appears in the application. For example, with letters of recommendation, ERAS experiences, and personal statements, we provide the exact quotes from the documents so faculty can easily verify accuracy. For publications, the platform outputs the full list of citations instead of summaries, ensuring nothing is fabricated. Because all scoring is transparent and based on criteria defined by the program, any discrepancies are straightforward to identify. In addition, we run multiple assessments each cycle to validate reliability. This structured, evidence-based approach minimizes hallucinations and ensures all outputs remain directly traceable to applicant documents.

In short, RankRx makes prediction transparent and understandable because the logic is your own, our platform simply executes it with speed, consistency, and clarity.

2. Protect Against Algorithmic Bias

AAMC's Concern:

The AAMC emphasizes that AI tools in admissions must actively identify, measure, and prevent bias. Institutions want to know how vendors address fairness, representative data, communication, and accessibility.

Questions:

• What are the historical biases found in your selection tool (e.g., how are they defined and measured)? How do you prevent biases from affecting your AI tool?
• How does the AI tool ensure fairness for all demographic groups, including underrepresented in medicine?
• (Follow-up) How do you ensure your training data are representative?
• (Follow-up) How do you communicate your bias mitigation efforts to users and applicants?
• (Follow-up) How does your AI system accommodate user needs, including accessibility features and assistive technology compatibility?

How RankRx Addresses This:

RankRx is intentionally designed to minimize historical biases by removing factors that often introduce inequity in applicant screening. Our system does not rely on external "training data" or historical admissions outcomes, which can reflect past bias, but instead applies program-defined scoring criteria consistently across all applicants. Because institutions set their own weights and categories (e.g., USMLE scores, U.S. clinical experience, research), RankRx avoids embedding hidden assumptions from prior datasets.

To ensure fairness across demographic groups, applications are anonymized before processing, removing names and date of birth that could contribute to bias. Scoring is then applied uniformly, so every applicant is evaluated only on the factors programs have chosen. For transparency, programs can always trace back how points were awarded, and they can adjust scoring rules if any unintentional disparities are observed.

We communicate these fairness safeguards directly to our institutional partners, highlighting how anonymization, program-defined criteria, and transparent scoring mitigate bias risks. If programs choose, they can disclose these protections to applicants as well, reinforcing trust in the fairness of the process.

In terms of accessibility, RankRx is built to work with assistive technologies and complies with modern accessibility standards (e.g., WCAG 2.1). The platform is structured to ensure that coordinators, faculty, and administrators can easily review outputs regardless of device or ability.

In short: RankRx prevents bias not by hiding behind "black box" models but by ensuring program control, anonymization, transparency, and accessibility at every step.

3. Provide Notice and Explanation

4. Protect Data Privacy

5. Incorporate Human Judgment

The Challenge

The AAMC emphasizes that AI should support admissions and selection, not replace human decision-making. Institutions need to see clear safeguards for oversight, training, and balance.

How RankRx Addresses This

RankRx is built to complement, not replace, human judgment. The platform applies program-defined scoring rules to ensure every application is reviewed consistently, but it does not make final admissions decisions. Instead, RankRx provides recommendations and structured outputs that help staff quickly identify applicants aligned with the program's criteria.

Programs remain fully in control: if there is ever a disagreement, the human evaluators' decision always takes precedence over AI outputs.

To ensure effective use, RankRx provides training and onboarding support for program staff and faculty. This includes walkthroughs of how scoring rules are defined, how outputs can be traced back to applicant documents, and best practices for integrating the system into the admissions workflow. Ongoing support is also available throughout each application cycle.

Subject matter experts are incorporated directly into the setup process. They determine which criteria matter, how points are assigned, and how the model should interpret application materials. This ensures the AI reflects each institution's unique mission and values.

Moreover, RankRx includes a dedicated "PD Score" column where program directors or faculty can add additional points or adjustments for individual applicants. This ensures that faculty oversight directly shapes the final score within the platform itself.

Finally, safeguards are in place to prevent over-reliance: all outputs are fully transparent and traceable to applicant documents, reminding evaluators that RankRx is a decision-support tool. Faculty and staff are encouraged to use it to streamline review, while maintaining ultimate responsibility for holistic and fair selection.

6. Monitor and Evaluate

The Challenge

The AAMC stresses that AI in admissions must be continuously monitored, audited, and improved — not just implemented once. Institutions need reassurance that vendors have processes for fairness, accountability, and alignment with evolving goals.

How RankRx Addresses This

RankRx includes ongoing monitoring and evaluation as a core feature of the platform. Each cycle, programs can review how scoring criteria performed, identify if adjustments are needed, and refine their point systems for the following year.

Because all scoring is transparent and traceable, it is straightforward to audit results for fairness and consistency, ensuring adherence to responsible AI practices.

We also gather feedback from program directors and coordinators to assess both the effectiveness of the scoring outputs and the user-friendliness of the platform. Training and support are updated regularly to reflect system improvements, so users are always aligned with the latest version without losing clarity or ease of use.

With RankRx, programs can check applicant scoring at any time during the cycle through real-time AI monitoring. If a program sees something that doesn't align with their scoring rules, they can request adjustments or fixes from us immediately. This ensures transparency and accuracy while the process is still underway.

In addition, we are expanding to participate in fellowship applications, which run on a different timeline than residency. This gives us the opportunity to test, refine, and improve the system in between residency cycles, so programs benefit from a continuously evolving platform.

Finally, we demonstrate that improvements lead to better outcomes by measuring efficiency and consistency gains (e.g., reduced screening time, clearer applicant comparisons) and by working with programs to track alignment with their institutional goals. This cycle of monitoring, feedback, and iteration ensures RankRx evolves responsibly alongside each program's needs.