Comments on OSTP and NITRD: Developing a National AI R&D Strategic Plan
The administration’s dedication to enhancing America’s innovation capacity in artificial intelligence (AI) is commendable. The National AI R&D Strategic Plan has evolved through multiple administrations, reflecting their priorities. Looking ahead, a renewed focus is needed to fulfill the commitment of making AI truly beneficial for the American public.
This article outlines three key recommendations to sharpen the federal AI research and development strategy:
- Make unlocking AI the central goal of federal AI R&D
- Prioritize research connecting technical design to AI performance outcomes
- Invest in research that generates more and higher quality data for AI
1. Make Unlocking AI the Central Goal of Federal AI R&D
The current Strategic Plan covers a wide range of important topics, but risk reduction tends to dominate its focus. While managing risks is critical, assuming that harm prevention automatically leads to benefits is misguided. Safer AI does not guarantee usability, scalability, or adoption where it matters most.
A strategy focused mainly on avoiding failure is not the same as one designed to enable success. If the U.S. wants to boost innovation, strengthen its economy, and improve public services with AI, the strategy must actively aim to unlock AI’s full potential. This requires revisiting each of the plan’s nine strategies to ensure they promote positive outcomes, not just risk mitigation.
Even governments known for cautious approaches, like those in Europe, are shifting their research focus to accelerate AI adoption. For example, France’s Inria aligns public R&D with national AI ambitions, Germany is restructuring ministries to enhance industrial competitiveness through technology, and the European Commission is developing a Strategy for AI in Science to boost AI uptake in critical areas such as health and climate.
The lesson is clear: public R&D must act as a catalyst for deployment and competitiveness. The U.S. should ensure its federal investments do more than reduce harm—they should actively enable AI’s value across sectors.
2. Prioritize Research Connecting Technical Design to AI Performance Outcomes
Effective AI adoption depends on understanding how technical design choices translate into outcomes like fairness, reliability, and security. Take the Veterans Health Administration’s pilot of Pingoo AI, which provides diabetic veterans with current health information using a retrieval-augmented generation model.
The VA first needs to identify which performance outcomes matter most—such as validity and reliability. Agencies like the National Institute of Standards and Technology (NIST) have outlined key trustworthy AI traits including reliability, safety, privacy, fairness, and explainability.
Once outcomes are defined, the challenge is knowing which technical methods will best achieve them. There is a range of techniques—improving document retrieval, filtering hallucinations, flagging uncertainty—but little research guides which are most effective or how to prioritize them in specific AI systems.
The National AI R&D Strategic Plan should fill this gap by mapping technical parameters to measurable AI outcomes. This would help organizations make design choices based on evidence, supporting more reliable and practical AI deployment.
3. Invest in Research to Generate More and Higher Quality Data for AI
The Plan correctly highlights data’s importance but treats it as mostly existing and in need of unlocking or sharing. In reality, innovation is often limited by data that doesn’t exist or is of poor quality, lacking representativeness or interoperability.
The next plan should prioritize research not just on accessing data, but on generating it—both in quantity and quality—as a core driver of AI progress.
- Research on Synthetic and Real Data Integration: A 2024 study from Epoch AI warns that high-quality public web text for training AI may be depleted between 2026 and 2032. Overtraining models risks exhausting valuable data sources sooner. Synthetic data—AI-generated text or images—offers a way to fill gaps but presents challenges:
- Synthetic data often misses rare or unusual cases, risking poorer AI performance in edge scenarios.
- It can introduce hallucinations—factually incorrect or distorted information that compounds errors if used as training data.
- The interaction between synthetic and real data during training is not well understood; improper mixing can degrade model performance.
Focused research is needed to develop methods that preserve rare case representation, detect and filter hallucinations, and define safe mixing ratios of synthetic and real data.
This area is especially promising for healthcare AI, where real data is sensitive or fragmented. Synthetic data can simulate patient records or rare conditions without compromising privacy, potentially improving diagnostic tools for underrepresented groups.
- Improving Data Collection in Underrepresented Areas: In fields like education, infrastructure, and public health, usable data is often missing or fragmented. This fuels a data divide, contributing to social and economic inequalities.
The federal R&D strategy should support research into innovative data collection methods, building sustainable partnerships in neglected sectors, and improving data quality, interoperability, and documentation standards. Encouraging international collaboration can also enhance these efforts.
Such investments will help close gaps in data availability and quality, enabling AI systems that better serve all communities.
For those interested in further AI training and resources to support these advancements, explore comprehensive courses and certifications at Complete AI Training.
```Your membership also unlocks:
