UNITED STATES DISTRICT COURT EASTERN DISTRICT OF MICHIGAN SOUTHERN DIVISION
SEITHER & CHERRY QUAD CITIES, INC., Case No. 23-11310
Plaintiff, F. Kay Behm v. U.S. District Judge
OAKLAND AUTOMATION, LLC, et al.,
Defendants. ___________________________ /
and
AP ELECTRIC, INC., Case No. 23-11342 Plaintiff, F. Kay Behm OAKLAND AUTOMATION, LLC, U.S. District Judge et al.,
CONSOLIDATED OPINION AND ORDER IMPOSING SANCTIONS FOR USE OF ARTIFICIAL INTELLIGENCE (“AI”)
This consolidated order is issued related to the parties’ briefings on pending Motions for Summary Judgment brought by the Defendants in two related cases. As communicated to the parties off the record, the court identified
a series of citations in Plaintiffs’ responsive briefings that appeared to be false citations generated by so-called artificial intelligence (AI) (with real case citations, but with fabricated quotations or explanatory
parentheticals that did not accurately reflect the holding of the case cited). These citations primarily appear on page 12 of both response briefings, e.g. Seither & Cherry, ECF No. 68, PageID.2056, though the
court did not verify every other page of Plaintiffs’ briefing at the time. The court then ordered Plaintiffs’ counsel to show cause why they should not be sanctioned pursuant to Rule 11(b)(2) & (c) of the Federal
Rules of Civil Procedure, and the inherent power of the court, for attributing fictitious quotations to court decisions and misrepresenting the holdings of various cases in reliance on AI-generated content. See,
e.g., Mata v. Avianca, Inc., 678 F. Supp. 3d 443, 2023 U.S. Dist. LEXIS 108263, 2023 WL 4114965 (S.D.N.Y. June 22, 2023). At this point, it “is no secret that generative AI programs are
known to ‘hallucinate’ nonexistent cases, and with the advent of AI, courts have seen a rash of cases in which both counsel and pro se litigants have cited such fake, hallucinated cases in their briefs.”
Sanders v. United States, 176 Fed. Cl. 163, 169 (2025). Attorneys who rely on “artificial intelligence (“AI”) to supplement their research” (ECF
No. 86, PageID.2373) do so in light of the well-known risks inherent in using those tools. The court also notes that the mere fact that the cases themselves that counsel cited were not fictitious (rather, only the quotes
or parentheticals) does not help matters; if anything, it highlights the risks of AI usage and reliance on these tools. When a case cite is “real,” an attorney, or for that matter a judge, might see a case they recognize
and assume the quote or holding has been accurately represented, where a case that an attorney does not recognize might, at least at first blush, trigger more exacting scrutiny.
The court does not say this to place more blame on Plaintiffs’ counsel for what appears to be a one-time event, and takes them at their word that they have taken steps to ensure it does not happen
again. The court merely uses the opportunity, for the benefit of pointing out an issue in a sure-to-be emerging trend in this area of the law, that as engineers “improve” their AI tools by ensuring its citations
at least use “real” cases, the accuracy and reliability of those citations, quotes, and explanatory parentheticals remain utterly suspect absent an attorney’s independent verification. Attorneys should understand
that chatbots, including legal “AI” chatbots, are large-language models (LLMs), not a true “artificial intelligence” out of the pages of science
fiction. They are not designed to answer questions factually.1 They are designed to mimic patterns of words, probabilistically. When they are “right,” it is because correct things are often written down in the
dataset they were trained on, not because they can independently fact- check themselves in the same way a human would. So when an LLM “explains” the holding of a case, it does so because it predictively strings
together a group of words that, when read by a lawyer, happens to represent something that is either true or false about that case. But on the back end, all the model was meant to do was make a sentence that
“looks” correct, which it did. They are word guessers; they are trying to predict what the next words would be if that sentence appeared on the internet. It has no way of “knowing” whether that sentence it created
about the case’s holding was in fact true or false.2
1 For a primer, try Mark Reidl, A Very Gentle Introduction to Large Language Models Without the Hype, Medium, https://mark-riedl.medium.com/a-very-gentle- introduction-to-large-language-models-without-the-hype-5f67941fa59e [https://perma.cc/284C-FBXG]. Riedl is a professor in the Georgia Tech School of Interactive Computing and associate director of the Georgia Tech Machine Learning Center. See https://www.cc.gatech.edu/people/mark-riedl [https://perma.cc/E7G6- 9D4S].
2 For example, “[t]here are things that we hold to be facts, like the Earth being round. An LLM will tend to say that[,]” because that is, on average, what an Having reviewed counsel’s response to the court’s order, the court
does not find that these citations were submitted in bad faith, which is required for sanctions under the court’s inherent power. See, e.g., Red Carpet Studios Div. of Source Advantage, Ltd. v. Sater, 465 F.3d 642,
646 (6th Cir. 2006). However, Rule 11 sanctions may be imposed regardless of whether an error was made in good or bad faith. “At the very least, the duties imposed by Rule 11 require that attorneys read,
and thereby confirm the existence and validity of, the legal authorities on which they rely.” Park v. Kim, 91 F.4th 610, 615 (2d Cir. 2024); see Mata v. Avianca, Inc., 678 F. Supp. 3d 443, 448 (S.D.N.Y.
2023); Willis v. U.S. Bank Nat’l Ass’n, No. 3:25-cv-516, 2025 U.S. Dist. LEXIS 92650, at *4-5 (N.D. Tex. May 15, 2025) (“’[c]onfirming a case is good law is a basic, routine matter and something to be expected from a
practicing attorney,’ especially because ‘[c]arelessness, good faith, or ignorance are not an excuse for submitting materials that do not comply with Rule 11.’”); Dehghani v. Castro, No. 2:25-cv-00052, 2025 U.S. Dist.
LEXIS 90128, at *12 (D.N.M. May 9, 2025) (“the standard under Rule
LLM will predict based on the sheer quantity of text affirming that. “But if the context is right, it will also say the opposite because the internet does have text 11 is one of objective reasonableness—the imposition of sanctions does
not require a finding of subjective bad faith by the offending attorney”), 5A Wright & Miller, Fed. Prac. & Proc. Civ. § 1335 (4th ed.) (“courts have regarded the presence of good faith as insufficient to excuse what
otherwise is a violation of one of the duties imposed by Rule 11”). Regardless of Plaintiffs’ counsel’s contrition, harm was done as a result of their acts. Defendants’ counsel were required to expend
additional resources responding to Plaintiffs’ updated briefing that the court ordered upon discovering the false citations. Judicial resources were diverted from considering pending cases to addressing Plaintiffs’
counsel’s misrepresentations. Regardless of counsel’s intent, greater care is warranted in the future. “Our system of justice must be able to rely on attorneys complying with their duty of candor to the court.”
Garner v. Kadince, Inc., 2025 UT App 80, ¶ 14 (May 22, 2025). The court finds that monetary sanctions are warranted to deter future conduct. Plaintiffs’ counsel – and not Plaintiffs – are responsible
for paying this penalty.
Free access — add to your briefcase to read the full text and ask questions with AI
UNITED STATES DISTRICT COURT EASTERN DISTRICT OF MICHIGAN SOUTHERN DIVISION
SEITHER & CHERRY QUAD CITIES, INC., Case No. 23-11310
Plaintiff, F. Kay Behm v. U.S. District Judge
OAKLAND AUTOMATION, LLC, et al.,
Defendants. ___________________________ /
and
AP ELECTRIC, INC., Case No. 23-11342 Plaintiff, F. Kay Behm OAKLAND AUTOMATION, LLC, U.S. District Judge et al.,
CONSOLIDATED OPINION AND ORDER IMPOSING SANCTIONS FOR USE OF ARTIFICIAL INTELLIGENCE (“AI”)
This consolidated order is issued related to the parties’ briefings on pending Motions for Summary Judgment brought by the Defendants in two related cases. As communicated to the parties off the record, the court identified
a series of citations in Plaintiffs’ responsive briefings that appeared to be false citations generated by so-called artificial intelligence (AI) (with real case citations, but with fabricated quotations or explanatory
parentheticals that did not accurately reflect the holding of the case cited). These citations primarily appear on page 12 of both response briefings, e.g. Seither & Cherry, ECF No. 68, PageID.2056, though the
court did not verify every other page of Plaintiffs’ briefing at the time. The court then ordered Plaintiffs’ counsel to show cause why they should not be sanctioned pursuant to Rule 11(b)(2) & (c) of the Federal
Rules of Civil Procedure, and the inherent power of the court, for attributing fictitious quotations to court decisions and misrepresenting the holdings of various cases in reliance on AI-generated content. See,
e.g., Mata v. Avianca, Inc., 678 F. Supp. 3d 443, 2023 U.S. Dist. LEXIS 108263, 2023 WL 4114965 (S.D.N.Y. June 22, 2023). At this point, it “is no secret that generative AI programs are
known to ‘hallucinate’ nonexistent cases, and with the advent of AI, courts have seen a rash of cases in which both counsel and pro se litigants have cited such fake, hallucinated cases in their briefs.”
Sanders v. United States, 176 Fed. Cl. 163, 169 (2025). Attorneys who rely on “artificial intelligence (“AI”) to supplement their research” (ECF
No. 86, PageID.2373) do so in light of the well-known risks inherent in using those tools. The court also notes that the mere fact that the cases themselves that counsel cited were not fictitious (rather, only the quotes
or parentheticals) does not help matters; if anything, it highlights the risks of AI usage and reliance on these tools. When a case cite is “real,” an attorney, or for that matter a judge, might see a case they recognize
and assume the quote or holding has been accurately represented, where a case that an attorney does not recognize might, at least at first blush, trigger more exacting scrutiny.
The court does not say this to place more blame on Plaintiffs’ counsel for what appears to be a one-time event, and takes them at their word that they have taken steps to ensure it does not happen
again. The court merely uses the opportunity, for the benefit of pointing out an issue in a sure-to-be emerging trend in this area of the law, that as engineers “improve” their AI tools by ensuring its citations
at least use “real” cases, the accuracy and reliability of those citations, quotes, and explanatory parentheticals remain utterly suspect absent an attorney’s independent verification. Attorneys should understand
that chatbots, including legal “AI” chatbots, are large-language models (LLMs), not a true “artificial intelligence” out of the pages of science
fiction. They are not designed to answer questions factually.1 They are designed to mimic patterns of words, probabilistically. When they are “right,” it is because correct things are often written down in the
dataset they were trained on, not because they can independently fact- check themselves in the same way a human would. So when an LLM “explains” the holding of a case, it does so because it predictively strings
together a group of words that, when read by a lawyer, happens to represent something that is either true or false about that case. But on the back end, all the model was meant to do was make a sentence that
“looks” correct, which it did. They are word guessers; they are trying to predict what the next words would be if that sentence appeared on the internet. It has no way of “knowing” whether that sentence it created
about the case’s holding was in fact true or false.2
1 For a primer, try Mark Reidl, A Very Gentle Introduction to Large Language Models Without the Hype, Medium, https://mark-riedl.medium.com/a-very-gentle- introduction-to-large-language-models-without-the-hype-5f67941fa59e [https://perma.cc/284C-FBXG]. Riedl is a professor in the Georgia Tech School of Interactive Computing and associate director of the Georgia Tech Machine Learning Center. See https://www.cc.gatech.edu/people/mark-riedl [https://perma.cc/E7G6- 9D4S].
2 For example, “[t]here are things that we hold to be facts, like the Earth being round. An LLM will tend to say that[,]” because that is, on average, what an Having reviewed counsel’s response to the court’s order, the court
does not find that these citations were submitted in bad faith, which is required for sanctions under the court’s inherent power. See, e.g., Red Carpet Studios Div. of Source Advantage, Ltd. v. Sater, 465 F.3d 642,
646 (6th Cir. 2006). However, Rule 11 sanctions may be imposed regardless of whether an error was made in good or bad faith. “At the very least, the duties imposed by Rule 11 require that attorneys read,
and thereby confirm the existence and validity of, the legal authorities on which they rely.” Park v. Kim, 91 F.4th 610, 615 (2d Cir. 2024); see Mata v. Avianca, Inc., 678 F. Supp. 3d 443, 448 (S.D.N.Y.
2023); Willis v. U.S. Bank Nat’l Ass’n, No. 3:25-cv-516, 2025 U.S. Dist. LEXIS 92650, at *4-5 (N.D. Tex. May 15, 2025) (“’[c]onfirming a case is good law is a basic, routine matter and something to be expected from a
practicing attorney,’ especially because ‘[c]arelessness, good faith, or ignorance are not an excuse for submitting materials that do not comply with Rule 11.’”); Dehghani v. Castro, No. 2:25-cv-00052, 2025 U.S. Dist.
LEXIS 90128, at *12 (D.N.M. May 9, 2025) (“the standard under Rule
LLM will predict based on the sheer quantity of text affirming that. “But if the context is right, it will also say the opposite because the internet does have text 11 is one of objective reasonableness—the imposition of sanctions does
not require a finding of subjective bad faith by the offending attorney”), 5A Wright & Miller, Fed. Prac. & Proc. Civ. § 1335 (4th ed.) (“courts have regarded the presence of good faith as insufficient to excuse what
otherwise is a violation of one of the duties imposed by Rule 11”). Regardless of Plaintiffs’ counsel’s contrition, harm was done as a result of their acts. Defendants’ counsel were required to expend
additional resources responding to Plaintiffs’ updated briefing that the court ordered upon discovering the false citations. Judicial resources were diverted from considering pending cases to addressing Plaintiffs’
counsel’s misrepresentations. Regardless of counsel’s intent, greater care is warranted in the future. “Our system of justice must be able to rely on attorneys complying with their duty of candor to the court.”
Garner v. Kadince, Inc., 2025 UT App 80, ¶ 14 (May 22, 2025). The court finds that monetary sanctions are warranted to deter future conduct. Plaintiffs’ counsel – and not Plaintiffs – are responsible
for paying this penalty. Pursuant to the court’s order, defense counsel submitted a declaration totaling his costs in relation to this issue at $1,485.00. Plaintiffs’ counsel shall pay that amount directly to defense
counsel.3 The court encourages, but does not require, Plaintiffs’ counsel to both complete a CLE course on the use of LLMs in legal research and
legal ethics. See Bunce v. Visual Tech. Innovations, Inc., No. 23-1740, 2025 U.S. Dist. LEXIS 36454, at *9 (E.D. Pa. Feb. 27, 2025). SO ORDERED.
Date: July 28, 2025 s/F. Kay Behm F. Kay Behm United States District Judge
3 Although “the purpose of Rule 11 sanctions is to deter rather than to compensate, [and] the rule provides that, if a monetary sanction is imposed, it should ordinarily be paid into court as a penalty[,]” here the court prefers that the penalty go to the party most directly injured by the offending conduct. See Fed. R.