Government Contracting

Introduction

In 2025, federal contractors relied on large language models (“LLM”) for assistance in preparing a significant number of bid and size protests, monetary appeals, and related federal procurement litigation (together “Disputes”). We know that because at least twenty 2025 public Decisions showed the “hallmarks” of Generative Artificial Intelligence misuse (“Gen-AI Misuse”), defined as the intentional or negligent use of an LLM resulting in inaccuracies that waste public and private resources.^[1]

LLMs are prone to inaccuracies, fabrications, and confabulation. An anthropomorphic catch-all term is “hallucination.” Because LLMs fundamentally are word prediction machines, they generate likely but false statements and citations around 30 percent of the time.^[2]

This is a new phenomenon in federal procurement: prior to 2025, no public decisions discussed Gen-AI Misuse and procurement tribunals are only slowly recognizing the threat posed by Gen-AI Misuse. As such, 2026 is likely to be much more problematic. In 2025, procurement tribunals collectively issued just two sanctions for misuse. Tribunals are unlikely to remain so lenient in 2026.

The immediate consequences are clear: procurement tribunals are already under pressure to resolve disputes quickly and have limited resources to do so. Brandolini’s Law says that energy spent to refute false claims is an order of magnitude higher than to create the falsehoods. This concept applies with equal force to frivolous bid protests and wasteful monetary appeals. A rapid increase in new filings may overwhelm the system with many flawed filings. While these are being resolved, many procurements are paused pending resolution. In every case, the Government, private parties, and tribunals will waste time and resources dealing with these filings. The long-term consequences are still emerging.

Identifying Gen-AI Misuse

Federal government contract Disputes are litigated at specialized judicial or quasi-judicial tribunals such as the Government Accountability Office (“GAO”), the Court of Federal Claims (“COFC”), the Armed Services and Civilian Boards of Contract Appeals (“ASBCA” and “CBCA”), and the Small Business Administration’s Office of Hearings and Appeals (“OHA”). These Disputes are rare. Together, the five tribunals handle only about 3,600 annual Disputes compared with over four million new federal contracts and orders annually.^[3]

Of these Disputes in 2025, thirty-one Filings involved Gen-AI Misuse with most (23) at GAO. GAO likely deals with the most Gen-AI Misuse because it hears about half of the approximately 3,600 Disputes and it has more pro se filers. Pro se contractors are more susceptible to being misled by sycophantic LLMs and less able at identifying and fixing hallucinations. Nineteen of the 20 final decisions involving Gen-AI Misuse were filed by pro se litigants.

There were five Gen-AI Misuse decisions at COFC and two at the ASBCA as well. Ultimately, the 31 filings resulted in 20 public decisions, 15 warnings, and 2 sanctions. In only three cases the litigant admitted using an LLM. When confronted, eight of the 20 implausibly denied using an LLM.^[4] For the remainder, it is unclear from the publicly available documents whether the tribunal asked about LLM usage.

2025 Gen-AI Misuse	Total	GAO	COFC	ASBCA	CBCA	OHA
Filings[5]	31	23	5	2	1	0
Decisions[6]	20	12	5	2	1	0
Warnings[7]	15	10	5	0	0	0
Sanctions[8]	2	1	0	1	0	0

 

What are the “hallmarks” of Gen-AI Misuse? Software detecting LLM writing is not reliable. In fact, humans are often better at identifying Gen-AI writing. In the Gen-AI Misuse decisions, tribunals focused on a more objective method of detecting LLM use: hallucinated case law citations.

GAO has divided these into three groups, including “(1) refer[ences] to fabricated or non-existent decisions; (2) fabricated or non-existent legal conclusions of decisions issued by our Office; or (3) otherwise objectively do not stand for the factual or legal propositions asserted. . .”^[9] Humans make errors, but not fabrications of this kind. COFC says that “Fake cases generated by AI often have reporter citations that lead to cases with different names, in different courts, and about different subjects.”^[10]

Another “hallmark” is the speed with which litigants file responses to briefing. For instance, in one COFC case, the Court remarked that the filings were likely Gen-AI because (in addition to fake cases) they were prepared and filed “within hours of Defendant filing its briefs.”^[11]

GAO

GAO’s fiscal year ran from October 1, 2024 to September 30, 2025. GAO’s annual report was silent on the trend of Gen-AI Misuse.^[12]  Nevertheless, GAO has been the quickest to call out hallucinations and warn protesters that inaccurate legal filings may be dismissed as a sanction. GAO has issued warnings to protesters in ten publicly available bid protest decisions.^[13] It is hard to tell how many other protests involved Gen-AI Misuse because GAO only makes certain decisions public.^[14]

Notably, GAO is the only tribunal to dismiss a protest or appeal as a sanction for repeated Gen-AI Misuse. In September, GAO dismissed four protest filings in a single scathing decision.^[15] The sanctioned firm, Oready, LLC (“Oready”), had filed 15 pro se GAO bid protests in a period of eight months, including multiple filings after it was sanctioned in September 2025.^[16]

Because Gen-AI use is not inherently objectionable, procurement tribunals do not seem interested in detecting or punishing any other type of Gen-AI Misuse at present. For instance, this Report considers one case involving a factual error to be Gen-AI Misuse, but GAO treated it like any other error.^[17] Oready was caught misusing Gen-AI in numerous other protests. In this protest, it claimed that certain information supporting its position could be found on pages that did not exist in a document, an obvious factual hallucination.

GAO denied or dismissed all 12 protests containing Gen-AI Misuse for various reasons, including the merits, timeliness, or standing defects.

COFC

COFC was the first federal procurement tribunal to note Gen-AI Misuse, in March 2025. Since then, the Court has warned five pro se litigants for including fabricated or inaccurate citations in filings.^[18]

In one case, COFC articulated the dominant tribunal stance on Gen-AI: “use of AI, by itself, does not violate this Court’s Rules. . . citation to fake cases does.”^[19] Likewise, COFC’s position on pro se Gen-AI Misuse has been taken up by other tribunals. “While courts afford pro se litigants considerable leeway, that leeway does not relieve pro se litigants of their obligation under Rule 11 to confirm the validity of any cited legal authority.”^[20]

Curiously, two of the five pro se litigants at COFC cited to the nonexistent case Tucker v. United States.^[21] The Court mused that because two different pro se plaintiffs cited to the same hallucinated case in separate COFC cases, “[p]erhaps both AI programs hallucinated this case name based on the Tucker Act, this Court’s jurisdictional statute.” Notably, one of those pro se plaintiffs filed similar actions in at least “three [additional] federal courts.”

All five Gen-AI Misuse cases were unsuccessful for the contractor.

ASBCA

The ASBCA has referenced the issue twice, sanctioning a law firm for Gen-AI Misuse in October 2025.^[22] The ASBCA struck a reply brief as a sanction because “over seventy (70%) of the citations were inaccurate.” Government counsel spent significant time identifying and explaining the fabrications. But the ASBCA conducted its own review, finding even more “citations not identified by the [Government] in its motion.” The law firm that filed the reply brief admitted using “AI technology,” claiming that it had “implemented safeguards to ensure the accuracy of the citations[.]” The ASBCA called the safeguards “woefully insufficient.”

Following these cases, the ASBCA’s website now displays a new page warning of Gen-AI Misuse. The page states: “We caution the parties that the current generation of AI tools are known to sometimes create materially false characterizations of legal precedent, misquote cases, and even create non-existent case citations. Thus, any party which uses AI tools to assist in drafting filings before the Board is expected to take independent steps to ensure the accuracy of such filings and may be subject to appropriate sanctions if their filings mischaracterize the law, misquote cases, or cite to nonexistent cases.”^[23]

CBCA

While the CBCA has not discussed any Gen-AI Misuse in its decisions so far, it appears to be aware of the issue, stating in its annual report that the “technology is ripe for abuses but also opportunities.”

The CBCA dismissed an appeal by Oready without prejudice in August. The decision itself was silent on the issue of Gen-AI Misuse. Nevertheless, this Report counts the Filing and Decision as Gen-AI Misuse because a review of appeal records obtained under the Freedom of Information Act confirms that Oready included hallucinated case law in its Certified Claim.^[24] Furthermore, Oready’s $7,947.50 appeal was dismissed because it was premature.^[25] The records also reflect that Oready appealed once again in CBCA 8563 sometime in August 2025. Oready’s second appeal is not counted as Gen-AI Misuse because it has not been resolved and there are no publicly available filings yet.

OHA

Each Dispute tribunal has commented on Gen-AI Misuse publicly except for OHA. The fact that OHA has not remarked upon any Gen-AI Misuse is surprising because many size protests are filed pro se. It seems likely that there are instances of Gen-AI Misuse at OHA that have gone unremarked upon.

Conclusion

Gen-AI Misuse almost certainly existed prior to 2025, but its recognition is an important turning point as tribunals consider how to deal with the relatively new technology. Contractors, particularly pro se contractors, are likely to continue using LLMs in Disputes. There is every reason to suspect that Gen-AI Misuse will grow in 2026. It is also reasonable to anticipate tribunals will warn and sanction more contractors for Gen-AI Misuse. Tribunals should consider processes for dismissing frivolous litigation when it identifies Gen-AI Misuse.

This Report proposes that the Tribunals address this problem systematically by requiring (1) disclosure of Gen-AI use in filings and (2) certification that the facts and citations have been verified. This pre-filing certification requirement would put contractors on notice of the risk of hallucinations. Tribunals would also be permitted to swiftly identify and dismiss frivolous protests.

_______________

[1] A “Decision” means a final or interim disposition of a one or more legal filings in a written finding by the tribunal.

[2] See e.g., Ren, Gruhlke, and Lauscher, Detecting Hallucinations in Authentic LLM-Human Interactions (October 12, 2025), https://arxiv.org/abs/2510.10539.

[3] In FY24 about 1,800 bid protests were filed at GAO. Half as many monetary appeals and bid protests were filed at COFC in FY24 (886). The CBCA and ASBCA, together, saw fewer than 800 monetary appeals in FY25. OHA regularly decides under 100 size appeals per year. There is no reliable data on Agency-level bid protests, but experts estimate these are rare. See Yukins, Christopher R. “STEPPING STONES TO REFORM: MAKING AGENCY-LEVEL BID PROTESTS EFFECTIVE FOR AGENCIES AND BIDDERS BY BUILDING ON BEST PRACTICES FROM ACROSS THE FEDERAL GOVERNMENT.” Public Contract Law Journal, vol. 50, no. 2, 2021, pp. 197-250. JSTOR, https://www.jstor.org/stable/27243262. Accessed 25 Nov. 2025. See https://web.archive.org/web/20241127180854/https://www.whitehouse.gov/omb/briefing-room/2023/11/08/fact-sheet-biden-harris-administration-announces-new-better-contracting-initiative-to-save-billions-annually/

[4] See e.g., Assessment and Training Solutions Consulting Corp., B-423398 (June 27, 2025) (claiming the fabricated citations were included because the protester had “limited access to subscription-based legal research databases”).

[5] Even when only one final decision is generated, if there are multiple distinct cases filed each is counted as a “Filing.” For instance, while GAO issued only one final decision related to KE System Services, Inc., there were three distinct case numbers meaning the contractor filed three separate protests. 

[6] A “Decision” means a final or interim disposition of a one or more legal filings in a written finding by the tribunal.

[7] A “Warning” is where a tribunal states that it could have but chose not to exercise its authority to sanction a contractor for Gen-AI Misuse. If a contractor was sanctioned, it is not considered a warning.

[8] A “Sanction” is a specific exercise by the tribunal of its inherent authority to punish the contractor for Gen-AI Misuse. For instance, GAO dismissed four protests at once as a sanction for repeated Gen-AI Misuse in Oready, LLC, B-423649.1; B-423650.1; B-423670.1; B-423670.2 (Sep. 25, 2025).

[9] Oready, LLC, B-423649; B-423650; B-423670; B-423670.2 (September 25, 2025).

[10] Sanders v. United States, 176 Fed. Cl. 163 (2025) (describing the “hallmarks of cases generated by AI.”).

[11] Parra v. United States, No 25-cv-431 (June 27, 2025).

[12] GAO-26-900695 (December 12, 2025).

[13] Raven Investigations & Security Consulting, LLC B-423447 (May 7, 2025); Oready, LLC, B-423524.1 (June 5, 2025); Assessment and Training Solutions Consulting Corp., B-423398 (June 27, 2025); Wright Brothers Aero, Inc., B-423326.2 (July 7, 2025); BioneX, LLC B-423630 (July 25, 2025); Oready, LLC, B-423524.2 (August 13, 2025); Helgen Indus. d/b/a/ DeSantis Gunhide, B-423635 (August 26, 2025); IBS Government Services, Inc., B-423583 (August 29, 2025); Supply Demanding, LLC, B-423810, B-423810.2 (November 24, 2025); KE System Services, Inc., B-423881; B-423881.2; B-423881.3 (December 22, 2025).

[14] See e.g., Oready, LLC, B-423524 (June 5, 2025) (unpublished). The majority of GAO bid protest filings do not result in a published decision because they are either unremarkable and GAO uses its discretion to not publish a public decision, or they are resolved via corrective action by the agency or withdrawal by the protester before a decision is rendered.

[15] Oready, LLC, B-423649.1; B-423650.1; B-423670.1; B-423670.2 (Sep. 25, 2025).

[16] Oready, LLC, B-423524.1 (June 5, 2025); Oready, LLC, B-422910.2 (June 18, 2025); Oready, LLC, B-423524.2 (August 13, 2025); Oready, LLC, CBCA 8427 (August 27, 2025); Oready, LLC, B-423823.1 (September 11, 2025); Oready, LLC, B-423905.1 (September 22, 2025); Oready, LLC, B-423650.1 (September 25, 2025); Oready, LLC, B-423649.1 (September 25, 2025); Oready, LLC, B-423670.1 (September 25, 2025); Oready, LLC, B-423670.2 (September 25, 2025); Oready, LLC, B-423906.1 (September 26, 2025); Oready, LLC, B-423758.1 (December 8, 2025); Oready, LLC, B-42394.1 (Open, filed September 22, 2025); Oready, LLC, B-424126.1 (Open, filed December 1, 2025); Oready, LLC, B-424096.1 (December 11, 2025); Oready, LLC, B-424096.2 (Open, filed December 23, 2025).

[17] Oready, LLC B-422910.2 (June 18, 2025) (identifying a factual error that was very likely a hallucination but not ascribing the error to Gen-AI Misuse).

[18] Sanders v. United States, 176 Fed. Cl. 163 (2025); Parra v. United States, No 25-cv-431 (June 27, 2025); Polinski v. United States, No. 25-cv-783 (October 15, 2025); McCaster v. United States, No. 24-cv-2027 (October 23, 2025); Strong v. United States, No. 24-cv-1537 (November 13, 2025).

[19] Polinski v. United States, No. 25-cv-783 (October 15, 2025)

[20] Sanders v. United States, 176 Fed. Cl. 163 (2025).

[21] See Sanders v. United States, 176 Fed. Cl. 163 (2025); see also Parra v. United States, No 25-cv-431 (June 27, 2025).

[22] Medical Receivables Solutions, Inc., ASBCA No. 64036 (July 15, 2025) (declining to address the agency’s allegations of Gen-AI Misuse); Huffman Constr., Co., ASBCA Nos. 62591, 62783 (October 23, 2025) (striking a reply brief as a sanction for Gen-AI Misuse).

[23] https://www.asbca.mil/Use-of-AI/ (accessed last on January 5, 2025).

[24] In its Certified Claim, Oready two of the three citations were fictitious, including the nonexistent: “Peter Kiewit Sons’ Co. v. United States, 74 Ct. Cl. 393 (1962)” and “General Dynamics Corp. v. United States, 17 Cl. Ct. 704 (1989).”

[25] Oready, LLC, CBCA 8427 (August 27, 2025).