Frequently asked questions
Everything you need to know about the AI, the technology, and how CitationShield protects your practice.
The AI Technology
What AI model powers CitationShield?
CitationShield is built on GPT-4o — OpenAI's most capable model — augmented with a Retrieval-Augmented Generation (RAG) pipeline. The RAG layer means the AI doesn't rely solely on training data. Before generating a verdict, it retrieves live records from CourtListener, Westlaw, and our proprietary citation graph. This is why CitationShield catches recently decided cases that older models miss.
What is RAG and why does it matter for legal citation verification?
RAG stands for Retrieval-Augmented Generation. Instead of asking a language model to recall a citation from training data (which may be outdated or hallucinated), RAG first retrieves the actual record from an authoritative database, then uses the AI to compare the retrieved record against the citation in your brief. This hybrid approach is fundamentally more accurate than pure generative AI for factual verification tasks.
How accurate is the AI?
In our internal test corpus of 10,000 citations (8,500 valid, 1,500 intentionally hallucinated), CitationShield achieved 97.3% overall accuracy. More importantly, the false negative rate — hallucinated citations passed as verified — was 0.0%. Every missed citation was flagged yellow (needs review) rather than silently passed green. We design for conservative errors.
What does the confidence score mean?
Every citation receives a Bayesian confidence score from 0 to 100, computed from multiple evidence signals: database match strength, reporter accuracy, date consistency, party name similarity, and court jurisdiction match. Scores ≥92 are Verified (green). Scores 60–91 are Needs Review (yellow). Scores below 60 are Not Found (red). The AI also provides a plain-English explanation of its reasoning for every verdict.
What databases does the AI query?
The RAG pipeline queries four sources simultaneously: CourtListener (free federal and state court opinions), Westlaw (via licensed API), our proprietary citation graph (built from 2M+ verified citations), and a vector similarity index for fuzzy matching on party names and reporters. All four are queried in parallel — that's why verification takes under 3 seconds even for large briefs.
Can the AI handle citations it hasn't seen before?
Yes. Because the RAG pipeline retrieves live data rather than relying on training data, CitationShield can verify citations from cases decided after the model's training cutoff. Our citation databases are updated daily. If a case was decided last week, our AI knows about it.
What It Verifies
What citation formats does CitationShield support?
CitationShield supports Bluebook (20th ed.), ALWD Guide to Legal Citation, neutral citation formats (used in many state courts), and informal citation styles. The NLP extraction model handles footnotes, inline citations, string cites, and embedded references. It also recognizes and flags short-form citations like 'Id.' and 'supra.'
Which courts are covered?
All federal courts (U.S. Supreme Court, all 13 circuit courts, all 94 district courts, FISA Court, Court of International Trade, Court of Federal Claims), all 50 state supreme courts and intermediate appellate courts, and 50+ supported reporters. Neutral citation formats for states that use them are also supported.
What exactly does the AI verify?
For each citation, the AI verifies: (1) case name — both parties; (2) reporter, volume, and page number; (3) court and jurisdiction; (4) year of decision; (5) whether the case actually exists in the database; (6) whether the cited proposition appears in the opinion (Enterprise tier). It does not currently verify pinpoint citations (specific page numbers within an opinion) — that feature is on our roadmap.
Does it catch AI-hallucinated cases?
Yes — that's the primary use case. When ChatGPT, Copilot, or any other generative AI invents a case that doesn't exist, the RAG pipeline will find no matching record in any of the four databases. The citation receives a confidence score below 60 and is flagged red. In our test corpus, 100% of hallucinated citations were caught.
Privacy & Security
Are my briefs stored on your servers?
No. Documents are processed in an isolated compute environment and deleted immediately after the verification run completes. We do not retain document content, citation text, or any brief data after the session ends. Your documents are never used to train AI models.
What is your zero-knowledge architecture?
Zero-knowledge means we design the system so that even CitationShield employees cannot access your document content. Documents are encrypted in transit (TLS 1.3) and at rest (AES-256). The compute environment that processes your brief is isolated and ephemeral — it spins up for your session and is destroyed immediately after. No logs of document content are retained.
Is CitationShield compliant with law firm data policies?
CitationShield is designed to be compatible with standard law firm data governance policies. We do not retain client matter data, do not train on your documents, and provide a Data Processing Agreement (DPA) for Enterprise customers. We recommend reviewing our privacy policy and, for large firms, contacting enterprise@citationshield.com for a compliance review.
Billing & Subscription
How does Apple billing work?
All CitationShield subscriptions are processed through Apple's secure in-app purchase system. You subscribe in the iOS app using your Apple ID. Apple handles billing, receipts, and cancellations. We never see your credit card or payment details. Manage your subscription at any time in Settings → [Your Name] → Subscriptions on your iPhone.
Is there a free trial?
Yes — every new account gets 5 free AI verifications so you can test the full GPT-4o + RAG pipeline on real briefs before subscribing. No credit card required to start.
What counts as one verification?
One verification = one document upload. We run the full AI pipeline on every citation in your brief. A 60-citation brief is still one verification. All plans include unlimited verifications — there are no per-citation or per-brief charges.
Can I cancel anytime?
Yes. Cancel directly in iOS Settings → Subscriptions. Your access continues until the end of the current billing period. No cancellation fees.