Improving the analysis, research, collection, and engagement of publicly available information

Problem:

The increasing volume, velocity, and variety of publicly and commercially available information challenge the military’s ability to track and make sense of the information environment. Foreign disinformation and influence campaigns are used to undermine American security. In the past four years, there has been a significant proliferation of tactics and the development of new, artificial intelligence (AI)-enabled techniques (e.g., synthetic media and DeepFakes created through generative AI). Given this new reality, improved tools are needed to provide situational awareness of prevailing narratives in the age of generative AI.

A key problem is the intense human labor and expertise required to monitor publicly available information across many languages, media types, and diverse cultural geographies while simultaneously aggregating this information that improves both understanding and operational effectiveness in the information environment.

Specific challenges include:

• Rapidly and sufficiently understanding the information environment to enable operations. 
• Accessing, collecting, discovering relevant publicly available information. 
• Analyzing diverse media types, including video and audio, requires huge amounts of time, localized expertise, language aptitude, and other skills that are in short supply.
• Operators require extensive training and background knowledge to be able to perform their jobs effectively.
• Preparing for operations in the information environment.
• Analytic report generation is constrained by manpower and not able to keep pace with the speed of relevance.
• Assessing friendly, adversary and neutral activities.
• Determining potential audiences and effectively engaging with them.
• Lack of a unified platform for compiling, analyzing, and visualizing operations in the information environment (OIE) to provide both situational awareness to commanders and the tools necessary to execute OIE tasks.

Proposed Solution:

The Department of Defense (DoD) seeks solutions that will provide significant advantages to personnel conducting operations in the information environment. The prototype should leverage some combination of large language models (LLMs), foundation models, and/or generative AI to advance some or all of the following operational needs and technical gaps (any one vendor need not address more than one area to be considered part of a potential solution):

1. Information operations discovery and evaluation:
   1. Automate analytics and reporting:
      1. Automated summarization across modalities (e.g. text, image, video, audio)
      2. Translation, transcription and description across modalities
   2. Identify, attribute, assess, and visualize information operations/campaigns.
      1. Entity resolution across data sources (e.g., person, organization) 
      2. Semantic search across multiple data modalities
      3. Identify, attribute, map, and assess foreign and non-state actor information campaigns
      4. Early detection and attribution of new information operations/campaigns.
      5. Geolocation: Identifying the geographic location of online content or users (e.g., automated geoguessr).
   3. Synthetic media / Deepfake detection: Identifying, explaining, summarizing and mitigating artificially generated media, such as images, text, or videos.
   4. Contextualizing activity in the information environment:
      1. Narrative analysis 
      2. Target audience analysis
      3. Tracking messaging efforts and correlations between events/actions and information campaigns
2. Operating information environment visualization:
   1. Provide commanders with easily understandable representations of the current and predicted information environment.
   2. Interactive visualizations of campaign emergence, spread, growth, and audience characteristics.
   3. Multiple UI/UX interfaces tailored for different user groups (e.g., analysts vs. commanders).
   4. Provide live visualizations and alerts of the information environment for use in current operations.
3. Content and messaging capabilities to support the work of analysts:
4. Generate a narrative strategy.
5. Assist in drafting content
6. Suggest a dissemination plan.

The solution may leverage and collaborate with tools already in the DoD inventory and should be compatible with Command and Control of the Information Environment (C2IE) tool. 

To demonstrate these capabilities vendors will either need to bring their own data feeds, partner with data providers or potentially use government data. Analytics vendors do not need to bring data feeds, nor do data providers need to bring analytic tools, however they should be interoperable. 

Model Agnosticism

Vendors should present flexible solutions that are not wholly dependent on a specific generative model, model provider, vendor or model architecture. The government should have the ability to update and/or change models if desired. Additionally, vendors must be prepared to disclose, but not necessarily share, the datasets which were used to train any model(s) included in the solution.

Responsible AI

All projects will be expected to comply with DIU’s Responsible AI Guidelines. Companies are encouraged to read the guidelines and to develop a plan for implementation, which may be presented as part of the submission. Particularly relevant issues for the use of large language models in this project include:

• Misinformation: Providing false or misleading information; not including appropriate caveats
• Opacity: Failure to provide references for claims; failure to disclose training methodologies
• Unintended bias: Discriminatory biases as well as more general biases that may be present in the training corpus
• Data drift: Failure to incorporate / adapt to new events and information
• Data leakage: Unintended release of sensitive information

Vendors should consider how they would address these issues in the context of this project from the outset. 

Additional Information

• Submissions must specify what large language model or family of models is incorporated into the prototype, whether such model is open or closed source, and what commercial arrangements (if any) exist with the model provider.
• Although models are not required to be run on premise this could be a requirement at a future date.
• A DD254 Facility Security Clearance (FCL) may be required, if the Government determines that it is necessary to complete the prototype effort. Therefore, vendors who receive awards may be required to ensure they are eligible to receive and maintain an FCL.

Awarding Instrument:

This Area of Interest solicitation will be awarded in accordance with the Commercial Solutions Opening (CSO) process detailed within HQ0845-19-S-C001 (DIU CSO), posted to FBO in Apr 2019. Additionally this document can be found within the DIU Library at www.diu.mil/library.

Follow-on Production:

Companies are advised that any prototype Other Transaction (OT) agreement awarded in response to this Area of Interest may result in the award of a follow-on production contract or transaction without the use of further competitive procedures. The follow-on production contract or transaction will be available for use by one or more organizations in the Department of Defense and, as a result, the magnitude of the follow-on production contract or agreement could be significantly larger than that of the prototype OT. As such, any prototype OT will include the following statement relative to the potential for follow-on production: "In accordance with 10 U.S.C. 2371b(f), and upon a determination that the prototype project for this transaction has been successfully completed, this competitively awarded prototype OTA may result in the award of a follow-on production contract or transaction without the use of competitive procedures.”