From AI Credit Scoring to Accountable Lending

To what extent do Vietnamese lenders that use AI or automated credit scoring publicly explain how their decisions are made, and how a customer can question or challenge one?

Four smaller questions sit under it:

1. Which borrower safeguards get disclosed, and which almost never do?
2. Do commercial banks and consumer finance companies disclose different levels of protection?
3. Did disclosure change around the AI Law, where older archived versions of the same documents exist?
4. Is there a gap between how loudly a lender advertises its AI and how much it tells borrowers about their protections?

Vietnam's Law on Artificial Intelligence (No. 134/2025/QH15) passed on 10 December 2025 and took effect on 1 March 2026. It's the country's first standalone AI law, and it borrows a lot from the EU AI Act. Decree 142/2026, in force since 1 May 2026, fills in the detail: three risk tiers, conformity checks for high-risk systems, and duties around transparency, explanation, and human oversight. AI systems already running in finance have until 1 September 2027 to comply.

That makes right now the moment for a baseline, taken after the rules exist but before the finance deadline. Earlier work has shown algorithmic credit scoring operating in a kind of legal limbo in Vietnam, but no study has audited what borrowers are actually told in public.

I built a disclosure index. Every lender gets scored on 12 indicators, each 0 (not disclosed), 1 (mentioned in general terms), or 2 (disclosed specifically), for a maximum of 24. The indicators group into four pillars:

• Notice: whether they tell you a decision was automated.
• Explanation: what data they use, where it comes from, what it's for, and whether they explain a decision.
• Recourse: human oversight, how to appeal or complain, and how to get your data corrected.
• Governance: third-party sharing, fairness controls, model monitoring, and who is accountable.

A separate score from 0 to 2 rates how strongly each lender publicly claims to use AI, which is what answers the fourth question.

The documents come in three layers: what a lender legally promises (privacy policies and loan terms), what a borrower actually sees (product pages, FAQs, complaint pages, app listings), and what the lender claims about AI (annual reports, capital adequacy disclosures, press releases). I collect them in Vietnamese and English, save dated copies with archive links, and freeze the whole set on one snapshot date so the study can be reproduced even after websites change.

Every score has to trace back to a dated source and a quoted excerpt. If the evidence is ambiguous, I record the uncertainty instead of forcing a confident score. A second coder independently scores 20 to 25% of the sample and I report the agreement between us. I've also asked finance and law researchers to critique the categories and the scoring before the codebook is frozen.

Lenders qualify if they give retail, consumer, or SME credit in Vietnam, if an official source shows automated scoring in a credit decision, if there are at least two dated public documents to code, and if their credit AI can be told apart from chatbots and marketing. I screened 15 candidates across commercial banks and consumer finance companies, with a go or no-go gate of at least eight qualifying lenders. Lenders are studied, not interviewed: everything is public, and nobody has to respond.

Public transparency only. It doesn't open up anyone's scoring model, test lending outcomes, or decide whether a lender has complied with the law. A safeguard that isn't disclosed isn't proof the safeguard is missing.

The research window runs from September to the end of October 2026: legal mapping and a pilot of the codebook first, then the corpus freeze, then full coding with the second coder. Writing starts on 30 November 2026, aiming for 10,000 to 15,000 words. There are no results until it's written up.