A Hearing Officer’s Warning About AI

On September 15, 2026, a Bureau of Special Education Appeals (BSEA) hearing officer issued a ruling that included the following warning to the parties about artificial intelligence:

“[T]he Parties are placed on notice that they are to be extremely cautious if and when using artificial intelligence (AI) in the drafting of any submissions. Overreliance on this tool may make their submissions unnecessarily convoluted, unclear and reliant on cases and regulations that either do not exist or do not support or address the issues at hand.”

South Hadley Public Schools – BSEA # 26-14590

It is not the first time the BSEA has written about this issue. In January 2026, a different hearing officer wrote that a filing prepared by an advocate “potentially appears to have been drafted using artificial intelligence supports as some of the legal claims/citations contained within it are wholly inaccurate.”

Dudley-Charlton Regional School District – BSEA # 26-05107

Over the past year, I’ve heard growing concerns from hearing officers about the prevalence of AI in BSEA matters, the amount of unnecessary work it creates, and how often it gets special education wrong.

AI can be a helpful tool for families, educators, and attorneys. It can explain unfamiliar terms, help organize a timeline, and make the special education process feel less overwhelming. But AI can also cause significant misunderstandings.

What happened in South Hadley

A parent representing herself filed three hearing requests between May and August 2026. She also filed a series of motions, notices, and status updates, and many of them were labeled “emergency.” Almost every request was denied.

According to the hearing officer, the parent’s sense of urgency “appears to stem from a misunderstanding that Student would start high school with none of his previous services.” The law is clear that the student was entitled to placement pending appeal, or “stay-put,” meaning he was entitled to the services in his last agreed-upon IEP.

The parent believed her son would start high school without an IEP and would have to be observed for 30 days before the Team could meet, so she kept him home from school. The hearing officer noted the parent “provided no legal basis” for that belief, and AI may have failed to accurately inform the parent about the law.

Why AI gets it wrong

I ran an experiment. I asked the three leading AI chatbots¹ questions as if I were the parent in South Hadley learning what services my son would receive in high school after an IEP expired or was rejected. Two completely misunderstood stay-put in Massachusetts and advised me to file an emergency hearing request.

At the risk of sounding empathetic to a text predictor², I understand where the chatbots were coming from and why they made this mistake.

Large Language Models (“LLMs”), the systems that power chatbots, acquire what is often referred to as “world knowledge” from their training data.³ They take in a large body of text from books, laws, internet articles, social media posts, etc., and compress it into a single model. That can create amazing outcomes and tools, but it can also create a lot of mistakes, especially when it comes to something as nuanced as the law.⁴ Most chatbots also search the web at answer time, but they don’t always pull accurate sources.

On stay-put, two of the chatbots cited federal regulations that require a family to file a request for a hearing to invoke their right to stay-put. That’s not the case in Massachusetts. In Massachusetts, stay-put attaches as soon as there’s a disagreement about an IEP.⁵

In South Hadley, under Massachusetts law, the student’s services did not depend on his mother filing anything, and his last agreed-upon IEP stayed in place when he started high school, even though it had expired.

I explain stay-put in Massachusetts here. Unfortunately, it seems these LLMs missed my blog post when they were building their world knowledge.

AI tends to miss exceptions to the general rule and focuses on the most common, which makes sense given that AI relies on patterns. In my experience, LLMs miss many basic legal principles, including that federal laws create a floor of rights, not a ceiling, and states can build on those rights. Massachusetts provides greater protection for families than federal law does when it comes to stay-put.

Even the chatbot that got it right offered to draft and file a hearing request, as if filing a lawsuit were as simple as ordering a pizza. The more emotion I conveyed in my questions, the more urgency I received in the answers.

A frightened parent with a tool that turns out confident, legal-sounding language on demand can turn a misunderstanding into a months-long fight.

What’s the risk?

In rare and extreme circumstances, the IDEA allows a court to order parents to pay a school district’s attorney’s fees. Under 20 U.S.C. § 1415(i)(3)(B)(i)(III), a court can award fees against a parent if the hearing request “was presented for any improper purpose, such as to harass, to cause unnecessary delay, or to needlessly increase the cost of litigation.”

In Lincoln-Sudbury, a matter where a family represented themselves at the BSEA, the hearing officer found their claim to special education eligibility “patently frivolous” and wrote, “If ever a matter merited a finding that it was brought for an improper purpose it is this.” Lincoln-Sudbury Regional School District – BSEA # 15-02427

The district then sued the parents in federal court. The court agreed with the hearing officer and ordered the parents to pay $188,996.15 in attorney’s fees plus $2,052.21 in costs. Lincoln-Sudbury Regional School District v. Mr. and Mrs. W., No. 16-10724-FDS (D. Mass. 2018)

Nothing in these recent rulings comes close to Lincoln-Sudbury. Pro se families and families represented by advocates are doing their best with the resources they have available, and families should not be punished for exercising their due process rights in good faith. However, I worry school districts might start arguing that they should be.

The Lincoln-Sudbury matter predated consumer access to AI and included what the hearing officer described as “insults, threats, distortions, misleading and tautological arguments.” But if a frustrated parent asks for that kind of writing, AI can produce a lot of it very quickly.

The warning in South Hadley went to both parties, and it applies to attorneys too. We should proceed carefully with this new, helpful, and potentially harmful technology.

Notes

1. Pew Research Center, Americans and AI 2026: Chatbots, Smart Devices and Views on Impact (June 17, 2026) (survey of 5,119 U.S. adults finding that OpenAI’s ChatGPT, Google’s Gemini, and Microsoft’s Copilot are the three most widely used AI chatbots, used by 44%, 24%, and 17% of adults, respectively), https://www.pewresearch.org/internet/2026/06/17/americans-and-ai-2026-chatbots-smart-devices-and-views-on-impact/. ↩
2. Stephen Wolfram, What Is ChatGPT Doing … and Why Does It Work? (Feb. 14, 2023) (explaining that when ChatGPT writes, it is essentially asking “over and over again ‘given the text so far, what should the next word be?’” and adding one word at a time), https://writings.stephenwolfram.com/2023/02/what-is-chatgpt-doing-and-why-does-it-work/. ↩
3. Tom B. Brown et al., Language Models Are Few-Shot Learners (2020) (describing training data for GPT-3 drawn from web pages, books, and Wikipedia), https://arxiv.org/abs/2005.14165; Fabio Petroni et al., Language Models as Knowledge Bases? (2019) (finding that language models “may also be storing relational knowledge present in the training data”), https://arxiv.org/abs/1909.01066. ↩
4. Damian Curran, Vanessa Sporne, Lea Frermann & Jeannie Paterson, Place Matters: Comparing LLM Hallucination Rates for Place-Based Legal Queries (Nov. 10, 2025) (testing GPT-4o, Claude 3.5 Sonnet, and Gemini 1.5 Pro on everyday legal questions and finding that each hallucinated legal facts in roughly half of its answers, and that the hallucination rate “is significantly associated with place”), https://arxiv.org/abs/2511.06700; Damien Charlotin, AI Hallucination Cases (database last updated Sept. 14, 2026) (collecting 2,041 court and tribunal decisions worldwide, 1,395 of them in the United States, in which a court found that a party relied on AI-hallucinated cases, quotations, or authorities; 1,174 of those filings came from self-represented litigants and 811 from lawyers), https://www.damiencharlotin.com/hallucinations/; Lnu v. Blanche, No. 24-4790 (9th Cir. June 3, 2026) (published order fining two attorneys $2,500 each and suspending them from practice before the court for six months after their briefs cited two nonexistent cases and misquoted real ones: “Lawyers using generative AI must thus be aware of the tendency of generative AI to make these mistakes and guard against them.”), https://cdn.ca9.uscourts.gov/datastore/opinions/2026/06/03/24-4790.pdf; Whiting v. City of Athens, Nos. 24-5918/5919, 25-5424 (6th Cir. Mar. 13, 2026) (ordering two attorneys to pay $15,000 each, plus the other side’s attorney’s fees and double costs, for appellate briefs containing more than two dozen fake citations and misrepresentations of the record), https://www.opn.ca6.uscourts.gov/opinions.pdf/26a0080p-06.pdf; Wilder v. President & Fellows of Harvard College, No. 2384CV01461-BLS2 (Mass. Super. Ct. May 18, 2026) (Salinger, J.) (refusing to let an out-of-state attorney appear in a Massachusetts case because he had been sanctioned in 2025 for filing motions citing eight AI-hallucinated cases and had “made no showing that he has learned from his mistakes”), https://www.courtlistener.com/opinion/10867360/glenn-wilder-and-others-v-president-fellows-of-harvard-college-and/; Heather L. LaVigne, Two Years of Fake Cases and the Courts Are Ratcheting Up the Sanctions, Mass. Office of Bar Counsel (2025) (surveying the Massachusetts sanctions orders for AI-generated fictitious citations in Smith v. Farwell (Norfolk Super. Ct. 2024) and Dastou v. Holmes (Middlesex Super. Ct. 2025)), https://bbopublic.massbbo.org/web/f/Two%20Years%20of%20Fake%20Cases%20and%20the%20Courts%20are%20Ratcheting%20up%20the%20Sanctions.pdf. ↩
5. 603 CMR 28.08(7) (“during the pendency of any dispute regarding placement or services, the eligible student shall remain in his or her then current education program and placement unless the parents and the school district agree otherwise”); 603 CMR 28.05(7)(b) (a district “shall implement all accepted elements of the IEP without delay”); compare 20 U.S.C. § 1415(j) (stay-put applies “during the pendency of any proceedings conducted pursuant to this section”). ↩