AI Legal Error Hotspots: Where Court Filings Go Wrong

A Concentrated Pattern of AI Errors Is Taking Shape in U.S. Courts

A few widely reported citation mistakes in courts were easy to dismiss at first. But today, this is no longer the case.

Courts across the U.S. are now seeing a steady rise in filings that include made-up or inaccurate references produced by AI tools. These issues are showing up more often in certain states, especially where AI adoption is moving faster than review practices.

As AI becomes part of everyday legal workflows, understanding where filings go wrong is becoming essential for reducing risk and protecting clients and reputations.

 

Key Takeaways

• AI-related legal filing errors are rising quickly, with cases surging from 25 in early 2025 to 249 by Q4.
• A small group of states, led by California and New York, accounts for a large share of total incidents.
• Most errors come from pro se litigants, who represent over 60% of all cases involving AI-related filing issues.
• Fabricated legal citations remain the most common mistake, making up more than half of all recorded errors.
• The majority of filings do not disclose which AI tool was used, limiting accountability and traceability.

 

AI Legal Errors by State: The Fastest Climbers and Repeat Hotspots

Across the U.S., the data points to a certain group of states driving the recent surge in AI-related filing errors.

California leads with the most cases at 97, followed by New York with 69 and Texas with 49. Florida and Illinois round out the top five with 42 and 33 cases, respectively.

Together, these states account for 40% of all recorded AI-related legal errors in court filings. 

The trend also shows no sign of slowing. In the first quarter of 2026, California reported 33 incidents while New York and Texas each recorded 14, keeping all three at the center of the issue.

Florida stands out for a different reason. Several states have accumulated high total numbers over time, but the Sunshine State experienced one of the fastest increases in the dataset.

The state recorded only a single incident in 2024, but this number jumped to 28 in 2025, a 2,700% increase, and recorded 13 cases in the first quarter of 2026.

If the current pace continues, Florida could exceed its 2025 total well before the end of the year.

Another notable development is the emergence of a second wave of states in 2025. Illinois, Pennsylvania, New Jersey, Indiana, Nevada, Washington, Mississippi, Oklahoma, Oregon, and Minnesota had no recorded incidents in either 2023 or 2024. But within a single year, each reported meaningful case counts.

This pattern suggests that AI-related filing errors are no longer limited to a few established hotspots. This issue is spreading quickly to new jurisdictions as AI adoption gains traction.

For legal teams, these patterns offer a clear signal of where risks are emerging fastest and where closer oversight may be needed when using an AI tool for lawyers.

 

 

Quarterly Trends by State

The upward trend for AI-related filing errors continues in 2026. Q1 2026 recorded 226 incidents, nearly matching the 236 reported in Q4 2025, even though the data covers only part of the quarter.

The year-over-year trend is even more striking. Q1 2024 recorded just 7 incidents, but this number increased to 24 in Q1 2025 and then surged to 226 in Q1 2026.

This represents a 32-fold increase over two years when comparing the same quarter.

Washington stands out as the fastest-growing state with a sustained upward trajectory. No incidents were reported in this state in 2024, but it recorded its first incident in Q1 2025 and has increased every quarter since, from 1 case to 13.

The state’s Q1 2026 total (13) matches the total recorded in 2025. No other state demonstrates this combination of consistent quarter-on-quarter growth, and the number continues to grow.

California holds another distinction regarding AI-related filing errors. Its case count kept climbing throughout 2025, rising from 2 cases in Q1 to 29 by the final quarter. It then reached a new high of 33 cases in the first quarter of 2026.

The increase from Q2 to Q3 2025 (14 in just one quarter) remains the largest absolute single-quarter increase of any state, and it has continued rising every quarter since.

 

A Look at the Fines and Sanctions

States are taking very different approaches to penalizing AI-related filing errors, and the financial impact and sanctions vary widely.

 

Monetary Sanctions

California leads by a large margin, with $256,553 in fines in terms of nationwide financial sanctions. This amount accounts for 35% of the total AI-related filing error fines collected in the country, which is a total of $723,549. These fines were imposed across 22 different cases.

California is followed by Florida at $99,019, Illinois at $78,426, and Colorado at $34,000.

Florida ranks second despite having only six sanctioned cases because the average fine for each case is $16,503. The average fine in Colorado is $17,000. Both are higher than California’s median penalties.

Oklahoma’s average is $14,450, and Illinois’s is $9,803, rounding out the top five for average fine severity.

 

 

By Severe-Sanction Share

This metric tracks how often incidents result in serious consequences. Examples of severe sanctions imposed include monetary fines, professional sanctions, case-dispositive sanctions, and equally serious court-imposed penalties.

Less severe sanctions include warnings, corrective orders, remedial directions, or similar non-punitive actions.

Louisiana recorded only eight incidents of AI-related legal errors, but the courts imposed severe sanctions in five of them. This gives Louisiana a severe sanction share of 62.5%, the highest among states with a meaningful number of cases.

Wyoming follows at 50%, while New Mexico reports 44.4% and Georgia 43.75%.

These figures suggest that courts in some jurisdictions are more likely to respond with punitive measures when AI-related errors appear in filings. The approach differs considerably across the country.

In comparison, several states with much larger caseloads report lower rates of severe sanctions. For instance, Texas recorded a share of 18.4%, and Florida obtained a score of 14.3%.

This indicates that a higher number of incidents does not necessarily translate into a higher likelihood of severe penalties.

​Overall, the data highlights significant differences in enforcement. Some courts appear more inclined to impose serious consequences, but others lean towards corrective actions and opportunities to address the AI-related filing errors.

 

 

Who’s Behind the Most AI Filing Errors?

Most AI-related filing errors in this study do not come from lawyers. They come from pro se litigants, individuals who represent themselves in court without the help of a licensed attorney.

Self-represented parties account for 468 mentions, or 60.6% of all cases, compared to 288 mentions for lawyers at 37.3%. This gap is significant and shows who often rely on AI during filing processes.

Based on the data, we can assume that those without formal legal training use AI legal tools most heavily, and that is where most errors are surfacing.

For courts and legal professionals, this raises practical questions about guidance, oversight, and how to reduce preventable mistakes at the source.

 

 

AI Tools Most Linked to Hallucination Errors in Court Filings

The majority of cases do not specify an AI tool: 532 mentions (69.27%) are marked as implied use, and 150 do not identify the tool (19.53%). Because of this, it is harder to trace responsibility and know where breakdowns occurred.

Among the tools that are named, ChatGPT (OpenAI) is the most frequently cited, with 48 mentions, or 6.25% of the total. While that share is relatively small, it still places it well ahead of any other identified system.

It also appears most often in several high-incident states, including California, New York, Texas, Illinois, Michigan, Nevada, New Jersey, Oklahoma, Virginia, Utah, and Louisiana.

Other tools, including Microsoft Copilot, Claude, Google Gemini, and Perplexity, appear only occasionally in the data.

Many filings do not disclose which AI platform was used, which limits accountability. Where tools are named, most AI-related errors may not be confined to a single system, but occur across a range of systems used in legal work.

This highlights the importance of choosing the right AI tool for legal applications.

 

Tool	No of Mentions	% tool mentioned
Implied	532	69.27%
Unidentified/Blank	150	19.53%
ChatGPT / OpenAI	48	6.25%
Westlaw / CoCounsel / Legal AI	6	0.78%
Microsoft Copilot	9	1.17%
Claude	5	0.65%
Lexis AI	5	0.65%
Google Gemini / Bard	1	0.13%
Other tools	11	1.43%
Perplexity	1	0.13%

 

What Kinds of AI Mistakes Happen Most, and Where?

 

 

AI-related filing errors tend to fall into a few clear categories, with fabrication standing out as the most common. It accounts for 624 mentions, or 51.7% of all recorded mistakes.        

Fabrications come in the form of citations, cases, or legal authorities that do not exist but are presented as real in the document.

California, New York, and Texas have the most fabricated information in court filings, with 85, 50, and 40 instances of fabricated information, respectively.

Misrepresentation follows at 26%, with 314 mentions. In these instances, the source may exist, but the way it is described or applied is incorrect. This can include misstating legal holdings or stretching conclusions beyond what the source supports.

California and New York have the most reported cases, with the former recording 37 mentions, and the latter, 27.

False quotes make up 20.90% of cases, with 252 mentions. These involve attributing passages or quotations to sources that they never actually contained. These errors often draw on real cases, but the quotes are altered in a misleading way.

Again, California reports the most incidents, with 26 mentions, followed closely by New York with 22.  

Outdated advice appears less frequently, but it is still something to be aware of. This reflects reliance on information that is no longer current or applicable.

In general, fabrication appears consistently as the leading issue in the U.S., especially in high-incident jurisdictions.

 

 

Types of Documents With the Most AI-Related Errors

AI-related errors are heavily concentrated in one area: case law. The data shows that 81.66% (677 mentions) of hallucinations involve case-law materials, far outweighing any other document type.

Case laws refer to past court decisions used as precedent. Lawyers cite them to support arguments by showing how similar issues were decided before.

The main risk in these errors is not general drafting mistakes, but the use of inaccurate or nonexistent legal authority. When citations or precedents are wrong, the entire filing can lose credibility.

Other document types appear far less often. Legal norms, which are statutes, regulations, and formal legal rules that form the foundation of legal arguments, account for 7.72% (64) of mentions.

Exhibits and submissions, which are the supporting materials filed alongside a case, such as documents, contracts, motions, affidavits, and other relevant evidence, make up 5.91% (49 mentions) of the overall data.

Other document types, such as doctrinal work, overturned case law, and repealed law, appear only occasionally.

The data suggests the problem is concentrated in legal authority, where accuracy matters most.

 

 

Strengthening Legal Accuracy in the Age of AI

The most common AI errors are not minor slips. They go to the core of legal accuracy. Fabricated cases, misrepresented rulings, and false quotes directly weaken the credibility of a filing and can lead to real consequences in court.

When the legal authority itself is wrong, the entire argument is at risk. This makes verification non-negotiable.

Legal teams should therefore treat every AI-generated citation as unverified until checked, using a mix of trusted research tools and review processes before submission.

 

Methodology

This analysis is based on a cleaned version of the AI hallucination court-filings dataset. Records were systematically reviewed and aggregated to support state-level, temporal, sanction-related, user, tool, and document-type insights.

Source: https://www.damiencharlotin.com/hallucinations/

State-Level Analysis
The original court venue field was standardized into a Mapped State variable, consolidating district court labels, circuit abbreviations, state appellate designations, and venue-specific naming conventions into their corresponding parent state. Cases originating from federal, national, tribal, or administrative venues that could not be reliably attributed to a single state, including GAO proceedings, the U.S. Tax Court, D.C.-only filings, Puerto Rico, and Guam, were excluded from state-level analysis. This leaves 724 cases used in all geographic breakdowns.

Time-Trend Analysis
Trends were analyzed using the filing date field and summarized at both monthly and quarterly levels.

As the dataset includes partial data for 2026 (through March 17, 2026), the “sharpest rise” metric was calculated using full-year data from 2024 to 2025 only, to avoid distortions from incomplete reporting.

Sanctions Analysis
Monetary penalties were extracted from the text-based penalty field and converted into numeric values where feasible.

A severe sanction case was defined as any record where the Outcome / Sanction reflected a materially punitive consequence. This includes monetary fines, professional or disciplinary actions, case-dispositive sanctions, or other significant court-imposed penalties.

Non-punitive outcomes, such as warnings, corrective orders, or remedial directions, were excluded from this classification.

The severe sanction share was calculated as:
Severe sanction share = (Number of severe-sanction cases in a jurisdiction) / (Total incidents in that jurisdiction)

Data Coverage Note
The dataset’s “Confirmed vs. Suspected” field contains only “Confirmed” entries. As a result, the confirmed-only leaderboard is identical to the all-incidents leaderboard.