Division of Protection (DoD) software program acquisition has lengthy been a fancy and document-heavy course of. Traditionally, many software program acquisition actions, comparable to producing Requests for Data (RFIs), summarizing authorities laws, figuring out related industrial requirements, and drafting undertaking standing updates, have required appreciable human-intensive effort. Nevertheless, the appearance of generative synthetic intelligence (AI) instruments, together with giant language fashions (LLMs), provides a promising alternative to speed up and streamline sure points of the software program acquisition course of.

Software program acquisition is one among many advanced mission-critical domains that will profit from making use of generative AI to reinforce and/or speed up human efforts. This weblog submit is the primary in a sequence devoted to exploring how generative AI, significantly LLMs like ChatGPT-4, can improve software program acquisition actions. Under, we current 10 advantages and 10 challenges of making use of LLMs to the software program acquisition course of and recommend particular use circumstances the place generative AI can present worth. Our focus is on offering well timed data to software program acquisition professionals, together with protection software program builders, program managers, techniques engineers, cybersecurity analysts, and different key stakeholders, who function inside difficult constraints and prioritize safety and accuracy.

Assessing the Advantages and Challenges of Generative AI in DoD Software program Acquisition

Making use of LLMs to software program acquisition probably provides quite a few advantages, which may contribute to enhancing outcomes. There are additionally necessary challenges and issues to think about, nonetheless, and the evolving nature of LLM know-how can pose challenges. Earlier than making an attempt to use generative AI to DoD software program acquisition actions, due to this fact, it’s important to first weigh the advantages and dangers of making use of these applied sciences to acquisition actions.

Our colleagues on the SEI not too long ago wrote an article that identifies some LLM issues that must be thought-about when deciding whether or not to use generative AI to acquisition use circumstances. Our weblog submit builds upon these and different noticed advantages and challenges when making use of generative AI to evaluate the professionals and cons for making use of LLMs to acquisition. Particularly, some advantages of making use of LLMs to software program acquisition actions embrace the next:

1. Effectivity and productiveness—LLMs can improve effectivity in software program acquisition by automating varied duties, comparable to producing code, analyzing software program artifacts, and aiding in resolution making. This automation can speed up processes and scale back handbook effort.
2. Scalability—LLMs excel in processing textual content and knowledge, making them appropriate for context-specific summarization and sophisticated inquiries. This scalability is effective when coping with intensive software program documentation, necessities, or codebases widespread in DoD acquisition applications.
3. Customization—LLMs could be custom-made by means of immediate engineering to refine context-specific responses. Acquisition applications can tailor the conduct of those fashions to go well with their particular software program acquisition wants, enhancing the relevance and accuracy of the outcomes.
4. Wide selection of use circumstances—LLMs have versatile functions in software program acquisition, spanning documentation evaluation, necessities understanding, code era, and extra. Their adaptability makes them relevant throughout a number of phases of software program acquisition and the software program improvement lifecycle. LLMs are educated on huge knowledge units, which implies they will contribute to a broad vary of software program acquisition subjects, programming languages, software program improvement strategies, and industry-specific terminologies. This broad information base aids in understanding and producing helpful responses on a variety of acquisition-related subjects.
5. Fast prototyping—LLMs allow speedy code prototyping, permitting mission stakeholders, acquirers, or software program builders to experiment with completely different concepts and approaches earlier than committing to a specific answer, thereby selling innovation and agile improvement practices.
6. Creativity—LLMs can generate novel content material and insights based mostly on their intensive coaching knowledge. They’ll suggest revolutionary options, recommend various approaches, and supply contemporary views throughout software program acquisition phases.
7. Consistency—LLMs can produce constant outcomes based mostly on their coaching knowledge and mannequin structure when immediate engineering is carried out correctly. LLMs have a configuration setting or temperature that permits customers to boost consistency in responses. This consistency helps enhance the reliability of software program acquisition actions, lowering the possibilities of human errors.
8. Accessibility and ease of use—LLMs are accessible by means of net companies, APIs, and platforms, making them available to acquisition applications. Their ease of use and integration into current workflows helps simplify their adoption in software program acquisition. LLMs are additionally accessible to people with various backgrounds utilizing a pure language interface. This inclusivity permits a variety of nontechnical stakeholders to take part successfully in software program acquisition.
9. Data switch—LLMs can facilitate information switch inside organizations by summarizing technical paperwork, creating documentation, and aiding in onboarding new workforce members, thereby selling information sharing and continuity.
10. Steady studying—LLMs can adapt and enhance over time as they’re uncovered to new knowledge and prompts by way of fine-tuning and in-context studying. This steady studying functionality permits them to evolve and turn into more adept in addressing software program acquisition challenges related to particular applications, laws, and/or applied sciences.

LLMs are nonetheless an rising know-how, nonetheless, so it’s necessary to acknowledge the next challenges of making use of LLMs to software program acquisition actions:

1. Incorrectness—LLMs can produce incorrect outcomes—typically known as hallucinations—and the importance of this incorrectness as a priority will depend on the precise use case. Errors in code era or evaluation can yield software program defects and points. The accuracy of LLM-generated content material should be verified by means of constant testing and validation processes. LLM governance for enterprise options requires constant monitoring and monitoring of LLMs as a part of a accountable AI framework.
2. Disclosure—Delicate data should be protected. Some software program acquisition actions could contain disclosing delicate or proprietary data to LLMs, which raises issues about knowledge safety and privateness. Sharing confidential knowledge with LLMs can pose dangers if not correctly managed (e.g., by utilizing LLMs which are in personal clouds or air-gapped from the Web). Organizations ought to concentrate on find out how to mitigate the enterprise safety dangers of LLMs and forestall entry to personal or protected knowledge. Information firewalls and/or knowledge privateness vaults can be utilized to implement some knowledge protections throughout the enterprise.
3. Usability—Though entry and ease of use are strengths of LLMs, some new abilities are required to make use of them successfully. LLMs require customers to craft applicable prompts and validate their outcomes. The usability of LLMs will depend on the experience of customers, and lots of customers will not be but proficient sufficient with immediate patterns to work together with these fashions successfully.
4. Belief—Customers should have a transparent understanding of the restrictions of LLMs to belief their output. Overreliance on LLMs with out contemplating their potential for errors or bias can result in undesirable outcomes. It’s important to stay vigilant to mitigate bias and guarantee equity in all content material together with methods produced by way of generative AI. Though LLMs can solely be efficient if bias is known, there are numerous sources for LLM bias analysis and mitigation.
5. Context dependency and human oversight—LLMs’ effectiveness, relevance, and appropriateness can range considerably based mostly on the precise atmosphere, use case, and cultural or operational norms inside a specific acquisition program. For instance, what could also be a big concern in a single context could also be much less necessary in one other. Given the present state of LLM maturity, human oversight must be maintained all through software program acquisition processes to make sure folks—not LLMs—make knowledgeable selections and guarantee moral compliance. The NIST AI Threat Administration Framework additionally supplies necessary context for correct use of generative AI instruments. When attainable, LLMs must be offered particular textual content or knowledge (e.g., by way of in-context studying and/or retrieval-augmented era (RAG)) to investigate to assist certain LLM responses and scale back errors. As well as, LLM-generated content material must be scrutinized to make sure it adheres to enterprise protocols and requirements.
6. Value—The prices of LLMs are altering with greater demand and extra competitors, however value is all the time a consideration for organizations contemplating utilizing a brand new software program software or service of their processes. Some techniques for addressing privateness issues, comparable to coaching customized fashions or growing compute sources, could be expensive. Organizations have to assess the entire prices of utilizing LLMs of their group, together with governance, safety, and security protocols, to completely think about the advantages and the bills.
7. Fixed evolution—LLM know-how is regularly evolving, and the effectiveness of those fashions modifications over time. Organizations should keep present with these advances and adapt their methods accordingly.
8. Mental property violations—The expansive coaching knowledge of LLMs can embrace copyrighted content material, resulting in potential authorized challenges when utilized to creating or augmenting code for software program procurement.
9. Adversarial assault vulnerabilities—Adversarial machine studying can be utilized to trick generative AI techniques, significantly these constructed utilizing neural networks. Attackers can use varied strategies, from tampering with the information used to coach the AI to utilizing inputs that seem regular to us however have hidden options that confuse the AI system.
10. Over-hyped LLM expectations of accuracy and trustworthiness—The most recent releases of LLMs are sometimes extremely succesful however will not be a one-size-fits-all answer to fixing all software program acquisition challenges. Organizations want to know when to use LLMs and what kinds of software program acquisition challenges are finest suited to LLMs. Particularly, making use of LLMs successfully as we speak requires a savvy workforce that understands the dangers and mitigations when utilizing LLMs.

Increasing Use Instances for Generative AI in Software program Acquisition

By contemplating the advantages and challenges recognized above, software program acquisition professionals can determine particular use circumstances or actions to use generative AI threat prudently. Generative AI will help on many actions, as indicated by ChatGPT in DoD Acquisitions or Assessing Alternatives for LLMs in Software program Engineering and Acquisition. Some particular software program acquisition actions we’re exploring on the SEI to find out the advantages and challenges of making use of generative AI embrace the next:

• Doc summarization—Understanding giant acquisition paperwork or a number of paperwork takes intensive and costly human effort. LLMs can present summaries of paperwork and supply an interactive atmosphere for exploring paperwork.
• Regulatory compliance—Maintaining with evolving authorities laws is crucial for DoD software program acquisition. LLMs can repeatedly monitor and summarize modifications in laws, making certain that acquisition actions stay compliant and updated.
• Normal identification—Figuring out related industrial requirements is a time-consuming process. LLMs can methodically parse by means of huge databases of requirements and supply suggestions based mostly on undertaking specs, saving time and lowering errors.
• RFI era—Producing RFIs is a vital step within the software program acquisition course of. LLMs can help in drafting complete and well-structured RFIs by analyzing undertaking necessities and producing detailed questions for potential contractors.
• Proposal analysis—Evaluating proposals from contractors is a important section in software program acquisition. LLMs can help in automating the preliminary screening of proposals by extracting key data and figuring out (non-)compliance with necessities.
• Threat evaluation—Assessing dangers related to software program acquisition is significant. LLMs can analyze historic knowledge and project-specific particulars to foretell potential dangers and recommend mitigation methods.
• Venture standing updates—Conserving stakeholders knowledgeable about undertaking standing is crucial. LLMs can generate concise undertaking standing studies by summarizing giant volumes of knowledge, making it simpler for resolution makers to remain up to date.

Authorities Rules and Steerage for Utilizing Generative AI

Publicly obtainable generative AI companies are comparatively new, and U.S. authorities laws and directives are altering to adapt to the brand new know-how. It is vital for any DoD acquisition stakeholders who’re contemplating utilizing generative AI instruments to concentrate on the most recent steerage, together with safety issues, to make sure compliance with the altering regulatory panorama. Some current examples of presidency steerage or rising coverage associated to generative AI embrace the next:

Trying Forward

Whereas generative AI provides many potential advantages for acquisition professionals, it’s important for DoD applications and acquisition professionals to judge how LLMs could (or could not) align with their particular software program acquisition wants critically and objectively, in addition to formulate methods to handle potential dangers. Innovation in software program acquisition utilizing generative AI is about growing productiveness for acquirers and stakeholders whereas mitigating dangers. People should proceed to have a central position within the software program acquisition actions, and people that may finest leverage new generative AI instruments safely can be essential to all stakeholders.

Deliberate exploration of LLMs inside the DoD’s acquisition processes is vital to gaining insights into each their advantages and potential pitfalls. By comprehending the capabilities and limitations of generative AI, software program acquisition professionals can discern areas the place its software is most advantageous and the dangers are both manageable or minimal. Our subsequent weblog submit on this sequence will delve into specific cases to facilitate cautious experimentation in software program acquisition actions, enhancing our grasp of each the alternatives and dangers concerned.