A common question is whether machine learning development can qualify as research and development for the federal R&D tax credit under Section 41. The short answer is that machine learning development may constitute qualified research when it involves a technical uncertainty and a process of experimentation. Implementing standard ML models for known applications generally is not qualified research, and internal-use software rules may apply. This page explains the framework in general terms. It is educational and is not individualized advice. For the foundational framework, see our page on qualified research.
When ML Development May Warrant Review
Machine learning development may warrant review when the work involves a genuine technical uncertainty and a process of experimentation. Under the four-part test, the work must be for a permitted purpose, be technological in nature (relying on computer science), be intended to eliminate uncertainty, and be conducted through a process of experimentation.
Common scenarios that may warrant review include:
- Model accuracy — evaluating alternative model architectures to resolve uncertainty about whether a model can achieve the required accuracy for a new problem.
- Training approach — testing alternative training approaches to resolve uncertainty about whether a model can be trained effectively for a new dataset.
- Inference performance — evaluating alternative approaches to resolve uncertainty about whether a model can achieve the required inference performance.
- New problem class — testing alternative approaches to resolve uncertainty about whether ML can be applied to a new problem class.
Routine Implementation vs. ML Development
A central distinction is between implementing standard ML and developing new ML:
- Routine implementation — applying a standard ML model (e.g., a pre-trained model or standard architecture) to a known problem. There is no technical uncertainty about whether the approach will work. This is implementation, not research.
- ML development — developing a new ML approach where there is a technical uncertainty about whether the approach can achieve the required performance, and evaluating alternatives to resolve that uncertainty. This may warrant review.
Hypothetical Example
Consider a company that is developing a new ML model to classify a new type of data and is uncertain whether any available model architecture can achieve the required accuracy. The company evaluates alternative architectures, tests each, and systematically varies the approach to resolve the uncertainty. This systematic evaluation of alternatives may warrant review as qualified research.
By contrast, if the same company applies a standard pre-trained image classification model to a known problem, that is routine implementation, not research.
This example is illustrative only and does not state that the activity definitely qualifies.
Documentation That May Help
Records that can help support ML development claims include design records identifying the uncertainty and alternative approaches, model-performance test results, training records, and records of how results informed model decisions. For more, see our page on R&D tax credit documentation.
Key Takeaway
Machine learning development may constitute qualified research when it involves a technical uncertainty and a process of experimentation. Implementing standard ML models for known applications generally is not qualified research, and internal-use software rules may apply. Because the distinction is fact-specific, professional review is appropriate before claiming the credit.