When courts sanctioned the attorneys in <\/i>Mata v. Avianca, Inc. for submitting artificial <\/i>intelligence (AI)-generated citations to nonexistent cases, they reached for the vocabulary of <\/i>negligence. That characterization is wrong. Large language models do not forget cases the way <\/i>attorneys do; they generate citations that look real but correspond to nothing in any reporter, and <\/i>they do so, as studies have since documented, at rates between fifty-eight percent to eighty-eight <\/i>percent on legal queries. The tools\u2019 general propensity to fabricate was publicly known by 2023, <\/i>before most of the incidents now filling the reporters. An attorney who submits AI output without <\/i>verification does not merely fail a diligence standard. They submit a representation to a court <\/i>while consciously declining to establish any basis for its accuracy, satisfying the elements of <\/i>reckless misrepresentation under Restatement (Second) of Torts \u00a7 526(b)\u2013(c). The cases that have <\/i>followed <\/i>Mata reach instinctively for language that exceeds negligence, calling citations \u201cbogus\u201d <\/i>and monetary sanctions \u201cinsufficient deterrents,\u201d and some have begun to reach for the word <\/i>\u201creckless\u201d without grounding that intuition in a cross-regime doctrinal framework. This Essay <\/i>supplies that framework.<\/i><\/p>\n
This Essay makes three original contributions. First, it introduces the concept of \u201cstructurally <\/i>predictable error\u201d to explain why AI hallucination is categorically different from human memory <\/i>failure: the error follows from the tool\u2019s architecture rather than from individual cognitive lapse, <\/i>its rate is publicly documented, and a reasonable practitioner in the field either knows or should <\/i>know its hallucination rate before using the tool. Second, it builds the doctrinal case for reckless <\/i>misrepresentation by connecting Section 526(b)\u2013(c) to the philosophy of AI output, arguing that <\/i>an attorney who relies on a tool architecturally indifferent to the distinction between truth and <\/i>falsity adopts that indifference as their own representational posture toward the court. Third, it <\/i>proposes a graduated three-tier accountability framework that calibrates sanctions to actual <\/i>culpability, preserving space for good-faith verification efforts and the access-to-justice benefits <\/i>that AI tools offer to under-resourced practitioners, while refusing to treat willful blindness toward <\/i>a known category of fabrication as mere carelessness.<\/i><\/p>\n