Why Most Legal AI Gets the Law Wrong (And How to Fix It)

Ask ChatGPT or Claude to calculate child support in California. You'll get an answer. It will sound confident. It might even cite a statute.

But here's what you won't know until it's too late: Is that statute current? Is the formula actually from California, or did the model mix in Texas law? Does that case it cited actually exist?

This isn't a hypothetical. It's the daily reality of legal AI built on general training data.

The Foundation Problem

Most legal AI applications are built the same way: take a powerful language model (GPT-4, Claude, Gemini), add some legal-sounding prompts, maybe fine-tune on legal documents, and ship it.

The problem isn't the AI's intelligence. It's the foundation.

General training data is a snapshot of the internet—outdated court opinions, conflicting blog posts, legal information written for consumers (not practitioners), and content from jurisdictions you're not practicing in. The model doesn't know what's current. It doesn't know what's authoritative. It doesn't know what jurisdiction you need.

It just knows what sounds right.

What "Sounds Right" Actually Looks Like

When I started building Divorce.law, I ran a simple test. I asked a general-purpose AI to calculate child support for a California case with specific income figures and parenting time.

The response was confident. Professional tone. Cited what looked like relevant authority.

Three problems:

Wrong formula - It used an income shares approach when California uses a specific guideline formula under Cal. Fam. Code § 4055

Wrong inputs - It didn't account for California's mandatory deductions or the parenting time adjustment

Wrong number - The final figure was off by over $400/month

A $400/month error. On something as basic as guideline child support. With a confident, authoritative tone that would make most people trust it.

That's not a bug. That's what happens when AI "knows" law the way it knows everything else—from patterns in training data, not from actual legal authority.

Why Family Law Is Especially Hard

Family law compounds this problem in ways that don't affect other practice areas:

1. Jurisdiction Matters More Than Anywhere Else

In corporate law, Delaware law dominates. In IP, federal law controls. But family law is radically local. Each state has:

Different child support formulas (income shares vs. percentage of income vs. hybrid)

Different custody presumptions (joint vs. primary caretaker vs. approximation)

Different property division rules (community property vs. equitable distribution)

Different spousal support factors (some states have 15+ factors, others have 5)

California alone has unique rules: the DissoMaster calculation, the Gavron warning requirement, the Family Code § 2640 reimbursement rules. None of this exists in Texas. Texas has its own universe.

A model trained on "legal text" doesn't distinguish California Family Code from Texas Family Code. It sees patterns. It generates plausible text. It doesn't know which law applies to your client.

2. The Law Changes Constantly

Child support guidelines get updated. Case law evolves. Legislative sessions happen. COVID-era temporary rules expired. Tax law changes affect support calculations.

Training data has a cutoff date. The model doesn't know what changed after that date. It doesn't know if the statute it's citing was amended, repealed, or superseded.

When a client's financial future depends on accurate numbers, "probably current" isn't good enough.

3. Formulas Require Precision

Family law involves actual math. Child support isn't a range or an estimate—it's a specific dollar amount calculated from specific inputs using a specific formula mandated by statute.

Language models are trained to predict the next word, not to execute precise calculations. They can describe how child support works. They struggle to actually calculate it correctly.

The Solution: Structured Legal Knowledge

When I realized general-purpose AI couldn't reliably handle family law, I had a choice: accept the limitations or build something different.

I built something different.

Victoria AI doesn't rely on what the model "knows" from training data. She has access to a structured legal foundation—jurisdiction-specific knowledge that tells her exactly what the law is in each of the 64 family law jurisdictions across the US and Canada.

How It Works

When Victoria calculates support in California:

She knows Cal. Fam. Code § 4055 is the guideline formula

She knows the specific income adjustments required

She knows how parenting time credits apply under the statutory scheme

She pulls from verified, current legal authority—not internet patterns

When the same user switches to New Jersey:

She knows it's NJ Court Rule 5:6-2 that governs

She knows what a Case Information Statement requires

She knows the 12 mandatory disclosure documents for that jurisdiction

She applies the correct income shares formula for NJ

This isn't prompt engineering. It's not asking the model to "act like a California family lawyer." It's providing the model with actual legal knowledge, structured in a way that ensures accuracy.

The Difference in Practice

General AI approach:

"Based on my training data, California child support is typically calculated using the income shares model, considering both parents' incomes..."

[Wrong. California uses a specific guideline formula, not income shares.]

Victoria's approach:

"California child support under Family Code § 4055 uses the formula: CS = K[HN - (H%)(TN)]. With your inputs: high earner net $12,400, low earner net $4,200, and 20% timeshare, the guideline amount is $1,847/month."

[Correct formula. Correct inputs. Correct result. Citable authority.]

What to Ask Before Trusting Legal AI

If you're evaluating legal AI tools, one question cuts through the marketing:

"Where does your legal knowledge come from?"

Listen carefully to the answer:

"The model's training data" - Be careful. You're relying on internet patterns.

"We fine-tuned on legal documents" - Better, but still no guarantee of currency or jurisdiction accuracy.

"We use RAG with legal databases" - Getting closer, but retrieval quality varies wildly.

"We built jurisdiction-specific knowledge structures" - This is what actual accuracy requires.

Red Flags

The tool doesn't ask which jurisdiction you're in

It gives confident answers without citing specific authority

It can't show you the source of its legal conclusions

It treats all states the same for state-specific questions

It doesn't know when statutes were last updated

Green Flags

Jurisdiction selection is required before legal questions

Answers cite specific statutes, rules, or cases

The tool acknowledges when something is outside its knowledge

Calculations show their work (formula, inputs, result)

Knowledge sources are transparent and auditable

The Real Cost of Wrong Answers

Legal AI errors aren't abstract. They affect real people.

A $400/month child support error over 18 years is $86,400. A custody recommendation based on the wrong state's presumptions could mean years of litigation to correct. A property division calculation that ignores a state's reimbursement rules could cost a client their separate property.

And unlike other AI applications where errors are inconvenient, legal errors can be malpractice.

The bar for legal AI should be higher than "sounds convincing." It should be "actually correct."

Building for Accuracy

It took me over a year to build Victoria's legal foundation. Not because the technology was hard—because the law is hard.

Every jurisdiction required:

Reading the actual statutes (not summaries)

Understanding the case law interpreting those statutes

Mapping the calculation formulas exactly as courts apply them

Identifying the mandatory disclosures and procedural requirements

Building validation to catch edge cases and exceptions

Then auditing it. Then auditing it again. Then testing against real calculations from actual cases.

This isn't the kind of work that makes for good demos. You can't see it in a screenshot. But it's the difference between an AI that sounds like a lawyer and an AI that can actually help you practice law.

The Future of Legal AI

The legal AI market is going to split into two categories:

Demo tools - Impressive interfaces, confident answers, built on general training data. Great for marketing. Dangerous for practice.

Practice tools - Less flashy, more rigorous, built on verified legal knowledge. Actually useful for client work.

Right now, most of what's available is in the first category. That will change as the market matures and lawyers start asking harder questions about where the answers come from.

The winners will be the tools that got the foundation right from the beginning.

Conclusion

AI is transforming legal practice. That transformation can either make lawyers more accurate and efficient, or it can introduce new categories of errors at scale.

The difference comes down to foundation. General training data produces general answers. Jurisdiction-specific, verified legal knowledge produces accurate answers.

Before you trust any AI with your client's case, ask the question: Where does your legal knowledge come from?

If you don't like the answer, find a different tool.

Antonio Jimenez is the founder of Divorce.law and creator of Victoria AI. He built Victoria's jurisdiction-specific legal foundation after discovering that general-purpose AI couldn't reliably handle family law calculations for his own cases.