Broyles v. McNeal: Chatbot Citations in Custody Brief — Georgia Court Issues AI Verification Warning

Background: Legitimation, Custody, and a Social Security Benefit Dispute

Joshua Broyles, a man diagnosed with schizophrenia but found by his treating physician and therapist to be financially responsible and not a danger to himself or others, filed a legitimation petition seeking recognition as the biological father of G. N. B. He also requested liberal visitation or primary physical custody and shared legal custody. The child’s mother, Fayth McNeal, conceded paternity but requested sole legal and physical custody.

The trial court granted legitimation but awarded sole legal and physical custody to McNeal, restricted Broyles to limited daytime-hour weekend visitation, and ordered Broyles to direct the Social Security Administration to transfer the child’s monthly SSDI dependent benefit of $957 to McNeal. The proceeding was not transcribed, creating evidentiary gaps on appeal. Broyles appealed pro se.

The Court of Appeals vacated and remanded on multiple substantive grounds — the visitation restriction was not adequately supported, and the SSDI transfer order raised legal questions that required further consideration. The custody outcome was therefore sent back to the trial court for redetermination consistent with the appellate opinion.

The AI Issue: Chatbot Citations in a Pro Se Custody Brief

McNeal’s counsel argued that Broyles had used a chatbot to prepare his appellate brief. The court acknowledged that “there are incorrect citations in the brief” — consistent with AI hallucination — but also noted that “at least one is a correct citation.” This nuanced observation distinguishes Broyles v. McNeal from cases where every AI-generated citation is fabricated: here, the tool produced a mixed result — some real, some not.

The court issued a general caution in a footnote that is unusual for its pragmatic tone: “We caution parties that the products marketed to them as ‘artificial intelligence,’ like calculators, can be useful tools, but any language-like strings of words or citation-like strings of numbers and letters provided therefrom must be verified against actual authority and should not be assumed to exist or be correct propositions of law or fact.”

This framing — comparing AI tools to calculators, acknowledging they can be useful, but insisting on independent verification — is more measured than some courts’ reactions. The court did not impose sanctions and did not dismiss Broyles’s appeal despite the citation issues. Indeed, Broyles achieved a partial appellate victory: the custody order was vacated and remanded.

The court reserved equal criticism for the opposing party’s attorney: McNeal’s response brief, drafted by a licensed attorney, “contains no supporting citation to case law, not even to the standard of review, and contains only statutory authority.” The only cases McNeal’s brief referenced were those Broyles had cited himself. This is a remarkable judicial observation — the AI-using pro se litigant cited more case law than the licensed attorney on the other side.

Holdings

1. Custody order vacated and remanded. The trial court’s limitation of visitation was not adequately supported and the SSDI transfer order required further proceedings. The case was remanded for redetermination consistent with the appellate opinion.
2. No sanctions for AI-impacted citations. Despite acknowledging incorrect citations in Broyles’s brief, the court did not sanction him and did not dismiss the appeal. The partial accuracy of his citations (at least one correct) and his pro se status weighed in his favour.
3. AI verification duty stated in practical terms. The court’s chatbot warning is notable for its non-punitive, instructional tone: AI tools may be useful, but verification against actual authority is mandatory.
4. Attorney’s brief also criticised. McNeal’s licensed attorney was criticised for providing no case citations in the response brief — with the court noting that the prudent attorney provides supporting authority. This creates an implicit equivalence: AI-generated inaccurate citations and no citations at all are both inadequate.

“We caution parties that the products marketed to them as ‘artificial intelligence,’ like calculators, can be useful tools, but any language-like strings of words or citation-like strings of numbers and letters provided therefrom must be verified against actual authority and should not be assumed to exist or be correct propositions of law or fact.”

— Georgia Court of Appeals, Broyles v. McNeal, May 8, 2026

India Angle: AI in Family Law — Child Custody Proceedings

Broyles v. McNeal is directly relevant to Indian family law practitioners. India’s family courts and High Courts regularly hear legitimation (paternity), child custody, and maintenance proceedings where one party is unrepresented or only partially represented. As AI chatbots become more accessible, pro se litigants — and even some practitioners — will increasingly use them to draft petitions, replies, and written arguments in family proceedings.

Relevant Indian Law

• Guardian and Wards Act 1890 / Hindu Minority and Guardianship Act 1956: Child custody proceedings in India are governed by these statutes, as well as personal law codes. Citations to Supreme Court and High Court decisions on “best interests of the child” are critically important in these cases. AI hallucination risks are highest when the factual basis is complex and legal principles require nuanced citation — exactly the profile of a child custody case.
• CPC Order XLVII / Family Court Rules: In Indian family proceedings, written arguments and notes of submissions form part of the record. An incorrect citation in a written submission, even by a pro se litigant, can be raised by opposing counsel and used to undermine the credibility of the entire submission. Unlike appellate briefs in the US, Indian family court arguments are often informal and subject to less formal scrutiny — but this cuts both ways: a fabricated citation may go unchallenged and influence the court.
• Advocates Act 1961 Section 35 / BCI Rule 9: For advocate practitioners (not pro se litigants), citing an incorrect case — even from a chatbot — in a family proceeding where a child’s interests are at stake would be a serious professional duty breach, potentially engaging the BCI’s misconduct jurisdiction.

Three Practical Tips for Indian Practitioners

1. In family law cases, citation accuracy matters because judges rely on it directly. Family court judges in India often deal with heavy dockets and may accept case law citations in written arguments without independent verification. If you submit an AI-hallucinated citation, it may actually influence the decision — making the harm from the hallucination more direct than in adversarial litigation where opposing counsel checks your citations.
2. Brief preparation AI versus citation AI are different risks. Using an AI tool to help structure arguments, draft standard paragraphs, or identify relevant legal issues is legitimate and useful — this is the “calculator” function the Georgia court acknowledged. The risk is specific to using AI to generate case citations. Keep these roles separate: use AI for structure, use verified databases for citations.
3. If opposing counsel alleges you used a chatbot, the appellate record now includes that allegation permanently. The Broyles case shows that courts will note chatbot use allegations in published opinions. Indian advocates should be prepared for this reputational dimension: even if no sanctions are imposed, a published High Court observation that your brief appears chatbot-generated is professionally damaging.

Quick Takeaways

• AI tools can produce mixed citation results — some correct, some fabricated — making selective verification unreliable. All AI-generated citations require independent verification.
• Pro se litigants who use AI chatbots in custody proceedings may win their appeals on substantive grounds even when their citations are imperfect — courts are distinguishing citation accuracy from legal merit.
• Licensed attorneys who respond to AI-impacted pro se briefs with citation-free submissions may face comparable judicial criticism — the duty to support legal arguments with authority runs both ways.
• The Georgia court’s “calculator” analogy for AI tools is a useful framing: just as a calculator can produce accurate numbers from a wrong formula, an AI tool can produce plausible-sounding citations from a wrong legal premise.
• In India: chatbot use in family law is coming; the verification duty applies to every citation regardless of how the brief was drafted.

Deep Dive: The Partial Accuracy Problem in AI Citations

Broyles v. McNeal introduces a dimension of the AI hallucination problem that is underexplored: partial accuracy. Most AI hallucination cases in the law focus on entirely fabricated citations — cases that simply do not exist. Broyles’s brief contained at least one correct citation alongside the incorrect ones. This partial accuracy is actually more dangerous than total fabrication in some respects.

When a brief is entirely fabricated, careful review by opposing counsel or the court will expose it. When a brief is partially accurate — mixing real cases with invented citations — the presence of correct citations creates an impression of legitimate research that may make the incorrect ones less obvious. A reader who verifies the first three citations and finds them accurate may not check all twelve. This is the same psychological mechanism that makes social engineering attacks effective: a partially correct message is more credible than an entirely false one.

For Indian advocates, this has a specific implication: do not spot-check AI citations. Verify every one. The cost of verifying ten correct citations is trivial compared to the cost of missing the one that is fabricated.

The court’s comparison of AI tools to calculators also deserves extended analysis. Calculators are deterministic tools: the same input always produces the same output, and errors are systematic (a calculator that adds wrong will always add wrong in the same way). AI language models are probabilistic tools: the same prompt can produce different outputs on different runs, and errors are non-systematic (the model may generate a correct citation ten times and a fabricated one the eleventh). This means the “test before you trust” approach that works for a calculator does not work for an AI citation tool. Verification must occur at the output level, not the tool level.

Indian practitioners should also note the family law context: child custody proceedings involve judicial discretion over factual findings that are rarely overturned on appeal. If an AI-hallucinated citation slips through at the trial level and influences a custody order in favour of one parent, the appellate correction process is long, expensive, and uncertain. Getting the citations right the first time — before a child’s custodial arrangement is determined — is not merely a professional courtesy. It is a matter of substantive justice for the child and the parties.