Every AI answer needs a receipt. A plain-English guide for the people who’ll be asked to explain one.

Picture this

(A made-up example ofcourse but a familiar one)

A customer asks your bank’s AI assistant whether there’s a fee for paying off their financing early. The assistant says, politely and confidently, that the fee is waived.

It isn’t.

The customer complains. Then someone senior asks the question that matters:

“Where did it get that from?”

Which document? Which system pulled it out? Which model wrote the sentence? Was it an old rate sheet, a bad search result, or did the model simply make it up?

Most teams can only answer: “the AI said it.” They have logs of what happened, but no record of who handed what to whom, and no way to prove that record wasn’t changed afterwards.

Why this keeps getting harder

A modern AI answer rarely comes from one place. It passes through several hands:

  • a search over your documents,
  • a language model that writes the answer,
  • sometimes other AI “agents” and tools in between.

Any one of those hands can drop a detail, bend it, or invent it. The final answer still sounds just as confident. Nothing in the sentence tells you which step went wrong.

A 1,200-year-old answer to a very new problem

This isn’t a new problem. It’s an old one in new clothes.

Classical hadith scholars spent centuries on exactly this question: when a saying reaches you through a chain of people, how much should you rely on it? Their method was simple and strict:

  1. Every report carries its full chain: who heard it from whom, in order.
  2. Every person in the chain has a track record, built up over many reports.
  3. A chain is only as strong as its weakest link.

ISNAD is that method, rebuilt for AI systems. The “people in the chain” are now your documents, your search system, your models and your tools.

What ISNAD does, in four steps

  1. Records the chain behind every answer: which document, which system, which model, in order.
  2. Grades each step from its track record, so you know which sources and models have earned your reliance and which haven’t.
  3. Lets the weakest step decide what happens: send the answer, send it with a warning, send it to a human, or block it.
  4. Writes a signed receipt for every answer. If anyone edits that receipt later, the edit shows.
Want to see this on one of your own AI pipelines? see isnad.islamandai.com

Does it actually work?

I publish the weak results next to the strong ones, because you should be able to check everything. Here are three tests in plain terms. All the numbers are public in the project’s GitHub repo.

Test 1: the original method. I ran ISNAD’s weakest-link rule on 575,060 hadith chains that scholars had already graded. It matched the scholars’ final verdicts very closely: a Cohen’s κ of 0.871, for the technical readers. That shows the method works on its home ground.

Test 2: a cheap shortcut that failed. My first attempt at spotting made-up AI answers used a quick text-similarity trick. It did no better than a coin flip. I published that too.

Test 3: real AI answers. I used a public research dataset (RAGTruth) of answers from six well-known AI models, where humans had already marked which answers contained made-up content. I added a strict AI “checker” to ISNAD and compared.

  • It caught almost every made-up answer (97.6%).
  • But about 1 in 4 of its alarms were false alarms, and a quarter of the answers couldn’t be scored at all. That still needs fixing.
  • It rated every model as worse than it really was. Yet when it ranked the six models from most to least reliable, it got the order nearly right (4 of 6 exactly, with the only swap between two models that were almost tied).

That last point is the whole idea in one result. A single check on a single answer is noisy. A track record built from many checks is sturdier. That’s why ISNAD grades your sources and models over time, instead of only judging answers one by one.

What it does not do

You’ll hear these limits from me before you hear them from your security team:

  • It doesn’t tell you an answer is correct. It tells you where the answer came from and how reliable each step has been. A reliable-looking chain can still carry a wrong document.
  • It doesn’t make you “compliant” with anything. It produces evidence, and your lawyers decide what that evidence covers.
  • The grades start with you. You decide how much to rely on a source at the beginning, and ISNAD then updates that grade based on what actually happens.
  • The receipts have limits. Signing has to be switched on with your own key. And someone who holds the keys and rebuilds the entire log from scratch isn’t caught yet. That needs an outside anchor I haven’t shipped.

Why this matters now

Three kinds of people are starting to ask “where did that come from?”:

  • Customers and complaint teams, when an AI answer costs someone money.
  • Auditors and boards, who want evidence rather than screenshots. If you work in Islamic finance, this won’t feel new: showing the basis for a position is already how Shariah review works.
  • Regulators. Under the EU AI Act, AI used in areas like credit decisions and education carries record-keeping and human-oversight duties. For those “high-risk” uses, the deadline is 2 December 2027.

Who this is for

  • Banks and fintechs, including Islamic finance, with AI assistants that talk to customers.
  • EdTech companies whose AI tutors answer students.
  • Any team whose AI answers get checked by an auditor, a regulator or an unhappy customer.

How to try it

  • Free: the core is open source (Apache-2.0), so you’re never locked in. pip install isnad
  • Hosted: ISNAD Cloud (isnad.islamandai.com) runs it for you, including signing keys, storage and one-click PDF evidence reports.
  • Free look: I’ll walk through one of your AI pipelines and show you what its receipts would look like.
  • Trust Assessment ($5,000, two weeks): on one of your pipelines, you get a report on which sources and models are least reliable, signed receipts you can check yourself, and a map for your lawyers showing what the evidence does and doesn’t cover. The full $5,000 counts toward a yearly plan if you continue within 90 days.

Email alizahidrajaa@gmail.com to start with the free call

ISNAD Cloud is not a compliance or certification product; it provides no legal advice and does not make any AI system “compliant.” Under the Digital Omnibus on AI (Regulation (EU) 2026/1744), EU AI Act high-risk obligations apply from 2 December 2027 for Annex III systems and 2 August 2028 for Annex I systems; whether either applies to you is a question for your counsel. A grade reflects the methodological reproducibility of a provenance chain, not the truth of any claim.

Links: isnad.islamandai.com · github.com/alizahidraja/isnad · arxiv.org/abs/2607.24117