An AI Legal Research System That Doesn't Hallucinate
Andrew Eichen
An AI-powered legal workflow for statutory analysis, designed by a practicing attorney to enforce interpretive discipline in AI tools.
AI Gets the Law Wrong in Predictable Ways
These are the most common failures identified while supervising AI-assisted legal analysis.
"Recalls from Training or Reads Summaries"
AI either recalls imprecisely from its training or uses general summaries from web search.
Shallow Reading
Applies a general rule without checking for the exemptions or parsing the specific elements a violation actually requires.
Misreading Standards
Reads "knows" and silently applies "should have known," creating obligations from thin air.
"It's Trained to Help, Not Hedge"
AI fills gaps instead of flagging silence.
Sycophancy
When challenged on one point, abandons the entire analysis instead of defending what was correct.
Fabrication
Fills gaps in a statute's silence with common-sense reasoning that has no statutory basis.
"It Assumes Every Law Applies"
AI skips the threshold question a lawyer asks first.
Overbroad Conclusions
Assesses scope so broad it would sweep in every company in the industry.
Skips Applicability
Determines what a statute requires without first checking whether it reaches this entity at all.
The Fix Isn't Smarter AI.
It's Encoded Discipline.
Hard rules enforced for every agent. Legal reasoning discipline built into architecture, derived from cataloging and correcting real errors across dozens of engagements.
From Question to Verified Analysis
Purpose-built AI workflows with built-in quality controls.
Interpretation Notes
Metadata attached to every law in the library
Principle
"Knows" means actual knowledge. The statute does not say "knows or has reason to know."
Consideration
State-level trigger language varies by jurisdiction; compliance timelines are unresolved.
The requirement to activate protections for minors applies only when the operator has actual knowledge (not constructive) that a specific user is a minor. This requires the operator to know, not merely have a reason to know. The statute is silent on whether a provider is considered to know if the user mentions their age to the chatbot.
One sentence for clear-cut conclusions. Full discussion for ambiguous applications. The depth of treatment is proportional to the analytical uncertainty.
The System Doesn't Look at What It Has.
It Looks at What the Question Requires.
Issue spotting in isolation prevents the library from biasing what the system looks for.
The Difference Discipline Makes
"The client is developing a resume screening model that a downstream employer will use in hiring. If the model produces disparate outcomes for a protected class, the client may be held liable under Title VII, the ADA, and related federal anti-discrimination statutes."
"Title VII imposes obligations on employers, not on tool vendors. Whether a developer that licenses a screening model is itself a covered entity under Title VII is far from settled, and would depend on whether they could be considered an employment agency, indirect employer, or agent of the employer. The more concrete exposure here is FTC UDAP enforcement, and state or local statutes that expressly reach vendors and service providers."
Covered entity verifiedThe first analysis extends a statute to a party it does not clearly reach. The second identifies who the law actually regulates and where the real exposure sits.
Built for Legal Complexity
Statutes tracked and analyzed across AI, privacy, and biometric law
Active AI litigation cases tracked and synthesized
Parallel research tracks covering statutes, case law, and terms simultaneously
Independent quality checks on every analysis
Connected to live legal databases across federal, state, and EU jurisdictions
Every design decision was made by a lawyer solving a problem encountered in client work.