A confident answer is not a correct one
Ask an AI a legal question and you'll often get a fluent, well-structured answer in seconds — complete with case names, G.R. numbers, and quoted provisions. It reads exactly like something a lawyer would write. That fluency is precisely what makes a wrong answer so dangerous.
A general chatbot doesn't retrieve the law. It predicts the most likely next words based on patterns in its training data. When the most likely-sounding answer includes a citation, the model produces one — whether or not that case actually exists. The result is a fabricated authority that looks indistinguishable from a real one.
This isn't hypothetical. Courts around the world have already sanctioned lawyers for filing briefs built on AI-invented cases. In the Philippines, where citing a fabricated authority can expose a lawyer to administrative liability, the stakes are immediate — and they fall on you, not the tool.
The hidden danger
The risk isn't a tool that's unsure. It's a tool that's confidently wrong.
The moment any AI generates an answer on top of what it retrieved, it can over-reach, misstate a holding, or stitch together a citation that was never in the source — and say it with total confidence. That's true of a general chatbot and of a purpose-built legal tool alike.
It can't tell fact from fiction
A fabricated case and a real one read identically. A model that generates text has no internal way to flag which citation it invented and which it actually read.
Opinions hide the gaps
A confident legal conclusion masks the missing context — the exception, the later amendment, the superseding rule — that a careful researcher needed to see.
You carry the risk
If you file on a wrong AI answer, the sanction, the dismissed motion, and the harm to your client land on you — not the chatbot.
What the research actually found
A peer-reviewed study from Stanford researchers — published in the Journal of Empirical Legal Studies — tested leading purpose-built legal AI research tools, the ones marketed to law firms as grounded and reliable. These are not generic chatbots; they use retrieval-augmented generation over real legal databases.
They still hallucinated. The most accurate tool produced fabricated or misgrounded information on roughly 1 in 6 queries; another did so on close to 1 in 3 — despite marketing that promised "hallucination-free" citations.
The tools tested were U.S. products. The reason the finding still matters here is architectural: any tool that retrieves from a database and then generates an answer over it shares the same failure point. The database can be real and the generated answer can still be a prediction. That's the part to watch, wherever the tool is built.
Even the best
1 in 6
queries to the most accurate purpose-built legal AI tool returned fabricated or misgrounded information — and weaker tools failed close to 1 in 3 times.
The lesson isn't "grounding doesn't work." It's that grounding alone isn't enough. The moment an AI generates an answer on top of the source, it can drift from it — and the more confident it sounds, the harder that drift is to catch.
Where this leaves Philippine practitioners
A general chatbot isn't built for Philippine law at all. Its training data is dominated by foreign material, so when you ask a Philippine question it does the one thing it always does — generates something plausible. For Philippine statutes, Supreme Court decisions, and the Rules of Court, "plausible" is not the same as "real."
The encouraging news is that Philippine-built legal AI has closed much of that coverage gap. Several local tools are grounded in real Philippine jurisprudence and take care to cite their sources. The question that remains — for any of them, and for BatasDB — is what the tool does after it retrieves: does it hand you the law, or does it hand you a generated conclusion you now have to fact-check?
BatasDB's bet is to cover Philippine law from the same authoritative primary sources and then stop at reporting it — the actual Batas Pambansa, the actual Republic Act, the actual G.R. number, ranked and quoted — and let you draw the conclusion. Where it adds any context beyond what it retrieved, it says so plainly.
What to ask any legal AI before you trust it
This isn't a scorecard for one tool against another — it's a checklist you can run against whatever AI you're considering, ours included. The more "yes" answers, the safer the tool is to put near a pleading. Here's how BatasDB answers each.
What to check
How BatasDB answers
Is it built on Philippine law — Supreme Court decisions, Republic Acts, the Rules of Court — rather than foreign material?
BatasDB indexes 200,000+ documents drawn directly from authoritative Philippine primary sources, including every available Supreme Court decision since 1946.
Does every result link to a real source you can open and read yourself?
Every citation points to a real document. Nothing is asserted that you cannot click through and verify.
Does it show you the actual provision or ruling text — not just a paraphrase of it?
Results are quoted from the source and split along the law’s own structure — section, article, paragraph — so each is a real, citable unit.
Does it tell you when the answer isn’t in its library, instead of generating one anyway?
If the law isn’t in the corpus, BatasDB says so. A known gap is far safer than a confident fabrication.
Does it summarise only what it actually retrieved — and visibly flag anything that came from outside the sources?
When BatasDB adds context beyond the retrieved documents, it labels it as such. You always know which sentence rests on a citation and which does not.
Does it rank Philippine binding authority first — Constitution, then SC en banc, down to circulars?
Results are ordered by how much authority a document carries, so the binding answer surfaces first.
These questions describe what to look for in any AI legal-research tool; the answers describe how BatasDB is built, not how any other product works internally. Run the same questions against any tool you're evaluating, including this one.
What's under the hood
"Reporting instead of opining" isn't a slogan — it's an engineering choice that runs through how BatasDB retrieves the law. The points below describe how our system is built.
Two searches at once
A meaning-based search and an exact-text search run in parallel, then merge — so "getting scammed by a car dealer" and "Article 315 RPC" both land on the right provision.
Split where the law splits
The corpus is broken along the law's own structure — section, article, paragraph — so every result is a real, citable unit you can quote, not a random fixed-size slice.
Ranked by legal authority
Results are ordered by how much authority a document carries — the Constitution and an SC en banc ruling outrank a circular on the same question, so the binding answer surfaces first.
Want the full walkthrough? See how our search works and how we rank by authority.
The bottom line: don't ask an AI what the law is — ask it to show you the law. BatasDB shows you the source and lets you verify every word.
Sources
- Magesh et al., "Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools" — Stanford RegLab / HAI; published in the Journal of Empirical Legal Studies, vol. 22, pp. 216–242 (2025).
- Stanford HAI, "AI on Trial: Legal Models Hallucinate in 1 out of 6 (or More) Benchmarking Queries" — summary of the study from the Stanford Institute for Human-Centered AI.