As part of a commitment to responsible innovation, the U.S. National Science Foundation is advancing the use of Artificial Intelligence (AI) to support its mission in science and technology. This page serves as a central hub for agency-wide AI efforts, including strategic direction and compliance guidance.
On this page
NSF's approach to AI
NSF is focused on deploying AI in ways that are ethical, transparent and aligned with our core values. The agency's AI initiatives are designed to enhance operational efficiency, improve service delivery and ensure equitable outcomes for all stakeholders.
Below you will find key documents that guide NSF's work with AI.
AI strategy
Office of Management and Budget (OMB) Memorandum M-25-21, Accelerating Federal Use of AI through Innovation, Governance, and Public Trust, directs each Chief Financial Officers Act (CFO Act) agency to develop an AI Strategy and post it publicly on the agency’s website.
The NSF AI strategy outlines the agency’s vision, goals and priorities for the development and use of AI technologies in satisfaction of the requirements of Section 2(a) of the Appendix to M-25-21. It provides a roadmap for integrating AI into NSF operations while ensuring alignment with both federal mandates and scientific integrity.
AI compliance plan
The AI in Government Act of 2020 and U.S. Office of Management and Budget (OMB) Memorandum M-25-21, Accelerating Federal Use of AI through Innovation, Governance, and Public Trust, direct each agency to submit to OMB and post publicly on the agency’s website a plan to achieve consistency with M-25-21, unless the agency makes and posts a written determination that the agency does not use and does not anticipate using AI.
NSF's AI compliance plan satisfies the requirements of Section 3(b)(ii) of the Appendix to M-25-21 and Section 104(c) of the AI in Government Act. The agency will report on compliance with the case-specific practices mandated in Section 4 of the M-25-21 Appendix separately, through its annual AI use case inventory, included below.
AI use case inventory
Executive Order 14179, "Removing Barriers to American Leadership in Artificial Intelligence," requires agencies to publish annual inventories of nonclassified and nonsensitive artificial intelligence use cases.
In compliance with this executive order and further guidance provided by the Chief Information Officer Council, this page lists AI use cases that the U.S. National Science Foundation uses to advance its mission, enhance decision making or otherwise benefit the public. This list is also linked in a tabular format.
NSF has developed an AI compliance plan and an AI strategy to comply with the Office of Management and Budget Memorandum 25-21.
For more information, please contact NSF Chief Information Officer Clyde Richards at clyde.richards@nsf.gov or (703) 292-7011.
NSF's 2024/2025 AI use case inventory
Updated December 2025
Use Case ID: AII-7
Agency: NSF
Bureau / Department: OCIO
Summary of Use Case: ServiceNow GenAI is a FedRAMP High-certified system to accelerate digital transformation using Now Assist. Automatically generate content such as responses, work notes, and knowledgebase articles related to customer service experience. Increase productivity and speed to delivery with intelligent recommendations. Resolve issues swiftly with chatbots that understand human language.
Stage of System Development Life Cycle: Deployed
Date Implemented: 2025
Use Case ID: AII-8
Agency: NSF
Bureau / Department: CISE/OAD
Summary of Use Case: The Budget Area Topic Classification Tool uses a pre-trained BERT model ('bert-base-uncased') to classify proposals into 1 of 8 budget areas: Advanced Manufacturing, AI, Advanced Wireless, Clean Energy, Microelectronics & Semiconductors, Quantum, or Trustworthy AI. This approach is semi-supervised in nature as the training data is comprised of pre-tagged NSF Directorate for Computer and Information Science and Engineering proposals from prior years. Upon training the model to learn the multi-label structure of the underlying data, the model's accuracy and discriminatory power were tested using a hold-out sample (i.e. 80/20 split). The model outputs a numerical vector of probabilities representing the likelihood that a given proposal belongs to 1 or more categories. A cut-off value can then be set for determining the final classification for a given proposal.
Stage of System Development Life Cycle: Pre-Deployment
Date Implemented: 2023
Use Case ID: AII-9
Agency: NSF
Bureau / Department: OLPA
Summary of Use Case: Wellsaidlabs has been used in our last projects to generate temporary placeholder voice-over, which will be replaced later by real voice-over, so it doesn't cause much AI policy concerns.
On top of the typical text to voice conversation, 11 Lab allows users to train their own voice-over, which sounds to have the potential to reshape our workflow, with the proper supervision and regulation.
Stage of System Development Life Cycle: Deployed
Date Implemented: 2023
Use Case ID: AII-10
Agency: NSF
Bureau / Department: OLPA
Summary of Use Case: NSF Office of Legislative and Public Affairs Creative Services team will use Photoshop's AI-powered tools to refine edits of photos and graphics, particularly the retouching of portraits.
Stage of System Development Life Cycle: Pre-Deployment
Date Implemented: 2023
Use Case ID: AII-19
Agency: NSF
Bureau / Department: EDU/DRL
Summary of Use Case: The resubmit checker takes a list of proposal IDs and returns a list of similar proposals submitted to NSF in the past, scored by their similarity. This is useful for identifying resubmitted proposals for several purposes: evaluating whether to return without review for lack of revision, ensuring continuity of merit review, and the analysis of broader resubmit trends.
Stage of System Development Life Cycle: Pre-Deployment
Date Implemented: 2024
Use Case ID: AII-22
Agency: NSF
Bureau / Department: OCIO
Summary of Use Case: Within software development, AI-driven tools, such as AWS CodeWhisperer, Amazon Q and Bedrock, and integration with development workflow, promise to elevate developer productivity and refine code quality. The objective of this pilot is to explore potential integration of these AWS services with DIS Developers IDEs, Jenkins automation, Bitbucket, laying the groundwork for Secure AI Assisted Development at NSF.
Stage of System Development Life Cycle: Pre-Deployment
Date Implemented: 2024
Use Case ID: AII-24
Agency: NSF
Bureau / Department: TIP-ITE-ENGINES
Summary of Use Case: Airtable is a low-code platform that incorporates OpenAI. The OpenAI services are purchased by and the data is kept private to Airtable. The NSF Directorate of Technology, Innovation and Partnerships has been approved to use Airtable for several usecases and Airtable has been certified as secure by the NSF OCIO Security Team. We would like to get approval to use Airtable to trial Airtable's AI's ability to match potential reviewer expertise to the reviewer.
Stage of System Development Life Cycle: Pre-Deployment
Date Implemented: 2024
Use Case ID: AII-25
Agency: NSF
Bureau / Department: TIP-ITE-ENGINES
Summary of Use Case: The NSF Directorate for Technology, Innovation and Partnerships is authorized to use Airtable for project management, including reviewer volunteers and recruitment. Airtable is a no-code platform that includes a built-in database. We do not store any PII in our database except for "public PII" such as name, position, work email address, areas of professional expertise, etc. We have two specific use-cases for which we would like to employ Airtable AI:
- People who volunteer sometimes to not list all areas of their expertise or expertise becomes stale. We have been able to do a proof of concept where Airtable AI is prompted to look at the potential reviewer's Linked In profile and identify or update the potential reviewer's expertise in its database based on recent updates to professional experience in LinkedIn. We would like permission to have AI complete this work. Typically, an admin person or program director will do this work and it is time-consuming and burdensome. AI can do it much faster and a member of the staff can check it rather than complete it for 400-500 reviewers.
- We also have a database of potential Engines submitters in Airtable -- people who have attended a webinar, roadshow, or submitted at an earlier time and asked to be contacted when new funding opportunities are available. We track very similar information in the potential submitters database as we do for reviewers who volunteer to review for the program. The information is used to support outreach to these potential submitters when new funding opportunities are announced or a new Engines event is announced. We need to know the state associated with the particular submitters' institution to determine which regional Engine they should be mapped to geographically (state in the U.S.). We have found that we can prompt the AI to look at the potential submitter's institution and the AI can locate the state 100% of the time to identify the region that the submitter would be associated with if they were accepted to the program. Currently, a staff person does this work and it is burdensome. The AI can do this work much faster so that a member of the staff can redirect their time to more value-added work and then go back and validate the AI's work. This permits the Engines program to segment outreach by event and regional engine.
The Airtable AI is an Airtable private instance of OpenAI.
Stage of System Development Life Cycle: Pre-Deployment
Date Implemented: 2024
Use Case ID: AII-26
Agency: NSF
Bureau / Department: TIP-ITE-ENGINES
Summary of Use Case: Per the "CHIPS and Science Act of 2022" mandate, the NSF Directorate for Technology, Innovation and Partnerships needs to know our investment portfolio based on the 10 KTA.
Stage of System Development Life Cycle: Pre-Deployment
Date Implemented: 2024
Use Case ID: AII-27
Agency: NSF
Bureau / Department: SBE/NCSES
Summary of Use Case: The key objective of this effort is to create and test machine-learning-backed or artificial intelligence-backed user experiences with federal statistical data. This experience shall improve on the current state of user interactions based on obtaining answers to questions via search engines or emailing federal staff or contractors. This project seeks to develop and pilot an AI chatbot (or the like) that answers users text queries submitted via an interface. Answers will be obtained from public statistical data of federal statistical agencies in this project.
Stage of System Development Life Cycle: Ideation
Date Implemented: 2024
Use Case ID: AII-30
Agency: NSF
Bureau / Department: OGC
Summary of Use Case: We don't have a tool yet. However, we are in the process of rolling out FOIAXpress and do not know yet of any AI capabilities
Stage of System Development Life Cycle: Ideation
Date Implemented: 2024
Use Case ID: AII-32
Agency: NSF
Bureau / Department: OLPA
Summary of Use Case: The NSF Office of Legislative and Public Affairs (NSF OLPA) is interested in using a ChatGPT integration with NSF's Digital Asset Management system (DAM), for the images we host there. We'd use it to generate alt text for images we'll be using in digital communication products (such as nsf.gov) to ensure they're accessible to individuals using assistive technology.
While we'd have ChatGPT generate the alt text, it would be reviewed by NSF OLPA staff before being used in any communications materials. We've confirmed with Acquia that they don't retain our images or use them for model training.
Stage of System Development Life Cycle: Ideation
Date Implemented: 2024
Use Case ID: AII-33
Agency: NSF
Bureau / Department: OIA/EAC
Summary of Use Case: We are exploring whether the application of BERTopic modeling (or a similar natural language processing method) is applicable to a portfolio analysis.
Stage of System Development Life Cycle: Ideation
Date Implemented: 2024
Use Case ID: AII-35
Agency: NSF
Bureau / Department: OIA
Summary of Use Case: We propose using unsupervised natural language processing to identify topical themes within responses submitted to the NSF public engagement questionnaire. The AI system will include multiple open-source techniques used together to achieve this goal. Details of the system are as follows: To ensure high-quality topics, we will first programmatically remove responses that are blank or noninformative. The National Center for Health Statistics Semi-Automated Non-Response Detection for Surveys model will be used to filter out nonresponse responses (e.g., "sdfsdfa," "idkkkkk"). To reduce duplicate responses potentially caused by bots or submitted to boost relative importance, responses with the exact duplicate question wording will be removed over a threshold of three responses per IP address. If an individual has several different questions they would like to submit, all of their questions will be included. We will use an open-source profanity filter such as python's profanity-filter to identify toxic language. Any responses that are primarily toxic language will be dropped from analysis. To group questions into themes, we will use a word-embedding model to convert text responses into a multidimensional space for topic modeling. Because we anticipate responses in multiple languages, we will use a multilingual word-embedding model, such as "paraphrase-multilingual-MiniLM-L12-v2" or the most up-to-date version of this model at the time of analysis. Embeddings will be reduced in dimensions using a dimension reduction algorithm such as t-SNE (t-distributed Stochastic Neighbor Embedding) to prepare for clustering. Reduced response embeddings will be clustered into groups using the ClassTfidfTransformer clustering algorithm from the python package BertTopic. The algorithm will be seeded with words corresponding to a list of NSF programs corresponding to different research areas to map the clusters to these established categorizations. To aid in the review of topics generated by the clustering algorithm, an open-source large language model (LLM), such as Meta Llama, will be used to summarize the responses grouped within a cluster. Using a prompt like the example below, this strategy will produce a plain language summary and title of each of the clusters of responses. The clusters produced will be reviewed by RTI staff for accuracy and appropriateness of naming. Staff will make edits to the LLM-generated summaries as necessary and note any clusters that are candidates for merging with other clusters into a single topic. The topics will then be validated by NSF experts and the advisory committee. Once the existing cluster model has been finalized, the cluster model will be applied to all responses. Where feasible, the RTI STEM education experts/analysts will manually map the research topics to NSF research programs.
Stage of System Development Life Cycle: Pre-Deployment
Date Implemented: 2025
Use Case ID: AII-38
Agency: NSF
Bureau / Department: OIA
Summary of Use Case: We propose using a Commercial Off the Shelf AWS product, Textract, as a character recognition tool to convert handwritten text into machine-encoded text for a postcard. Handwritten text will be an open-format question for science and engineering and some optional check boxes for demographic and other data questions.
Stage of System Development Life Cycle: Deployed
Date Implemented: 2025
Use Case ID: AII-48
Agency: NSF
Bureau / Department: OCIO
Summary of Use Case: The NSF Office of the Chief Information Officer is working to provide AI-ready RPPR data – a key component for determining outcomes of NSF investments. Note: Information in the RPPR may be used to identify "outputs" and be linked to other data sets (e.g., NSF PAR, Patent, etc.) to assess NSF's "broader impacts"; however, it is not currently available in an accessible format. Problem Statement: RPPR data, a critical resource for tracking the outcomes of NSF-funded projects. Data are currently stored in various formats across multiple locations, making it inaccessible for analytic purposes and difficult to reach without impact to current operations. The lack of centralized, accessible data limits NSF's ability to assess impacts, make informed decisions, and fully leverage data for insights. Request approval to use AWS Bedrock, AWS SageMaker, AWS Neptune & AWS Comprehend from AIDG.
Stage of System Development Life Cycle: Pre-Deployment
Date Implemented: 2025
Use Case ID: AII-49
Agency: NSF
Bureau / Department: TIP & OCIO
Summary of Use Case: Copilot can be used to rewrite and generate text based on user prompts in Microsoft 365 services, including Microsoft Word, Microsoft Excel, and PowerPoint. Copilot for Microsoft 365 uses Microsoft Graph, an API, to evaluate context and available Microsoft 365 user data before modifying and sending user prompts to the language model. After receiving its output, Microsoft Graph performs additional context-specific processing before sending the response to Microsoft 365 apps to generate content.
Stage of System Development Life Cycle: Pre-Deployment
Date Implemented: 2025
Use Case ID: AII-51
Agency: NSF
Bureau / Department: ENG/CBET
Summary of Use Case: The AI system utilizes state-of-the-art transformer-based models (e.g., BART, Pegasus, T5, SciBERT) to perform text summarization. These models are deployed entirely within the agency's secure firewalls, ensuring no external data exposure and no external access. BART, Pegasus, T5 and SciBERT are state-of-the-art, open-source AI models designed for natural language processing tasks. The model will be able to summarize lengthy internal documents, helping employees quickly extract key insights from large volume of information. This tool automates summarization, reducing the time employees spend reading lengthy documents, enhancing productivity by accurately identifying key points in such documents. The traditional extractive summarization (e.g., via TF-IDF) selects key sentences verbatim from the source, often misses several key elements of the document, and results in an inaccurate representation of the content. Unlike traditional extractive summarization, generative summarization (used by BART, Pegasus and T5) creates modified sentences that better capture the meaning and context of the document. This leads to more natural, coherent and concise summaries which enhanced readability and contextual accuracy.
Stage of System Development Life Cycle: Pre-Deployment
Date Implemented: 2025
Use Case ID: AII-52
Agency: NSF
Bureau / Department: ENG/CBET
Summary of Use Case: The AI system utilizes a sentence-transformer model (all-mpnet-base-v2) and a statistical keyword method (TF-IDF) to perform document similarity tasks. These models are deployed entirely within the agency's secure firewalls, ensuring no external data exposure and no external access. These AI models are designed for natural language processing tasks. Via all-mpnet-base-v2, the model will utilize deep learning models to capture semantic meaning, ideal for conceptual similarity beyond exact wording. Via TF-IDF, the model will focus on lexical overlap and is effective for direct technical document comparison. Both models calculate Cosine similarity, with scores ranging from -1 to 1. Scores can be interpreted as follows: identical (1), highly similar (0.8-0.99), somewhat similar (~0.5), mostly dissimilar (0.1-0.3), completely dissimilar (0) and opposite (-1 to 0). Document similarity analysis can significantly boost productivity since a large volume of documents can quickly and efficiently be analyzed for the extent of contextual or lexical similarity and can aid in routine planning for proposal administration. Cosine similarity is also a well-established method to quantify similarity and is already used within NSF tools.
Stage of System Development Life Cycle: Pre-Deployment
Date Implemented: 2025
Use Case ID: AII-53
Agency: NSF
Bureau / Department: OIA
Summary of Use Case: "I request to use the existing AI Builder pre-built GPT functions within Microsoft Power Platform to perform summarization tasks on text and tag data in M365 to support:
Creation of Planner tasks from emails and
Creation of draft weekly and monthly accomplishment reports from text and tag data in email, meeting invitations, and tasks.
The scope of data is limited to that available to an individual M365 account."
Stage of System Development Life Cycle: Deployed
Date Implemented: 2025
Use Case ID: AII-56
Agency: NSF
Bureau / Department: BIO/IOS
Summary of Use Case: The goal of the analysis is to compare how any two programs, divisions and directorates in NSF are similar to each other to assess uniqueness and overlap between units. The comparison is based on a similarity score (cosine similarity) of the embeddings of any two proposal summaries. The embeddings are generated using SentenceTransformer.