Vietnam’s AI rules rely on human judgement that AI sways in lab tests

Under Vietnam’s rules on artificial intelligence, a system that only analyses, forecasts or recommends “for reference”, and whose output may not be the sole basis for the final decision, is not to be proposed for the high-risk list. In a laboratory preprint testing reasoning rather than regulatory safeguards, participants who consulted a deliberately wrong AI followed it about 4 times in 5.

What the rules say

The Law on Artificial Intelligence, No. 134/2025/QH15, took effect on 1 March 2026 and sorts systems into three risk tiers. High-risk systems must be designed for human oversight. Subject to statutory exceptions, deployers compensate harm from high-risk systems operated in compliance with the law; reimbursement from providers, developers or other parties depends on agreement. Systems in finance that were operating before 1 March 2026 have until 1 September 2027 to comply.

Decree 142/2026/NĐ-CP, in force since 1 May 2026, sets criteria under which a system that meets the risk tests is not proposed for the high-risk list. One covers systems with technical mechanisms and operating procedures that ensure substantive human oversight, so that an authorised person can independently review, intervene in, refuse or change the system’s decision before it takes effect. The other covers systems that only analyse, forecast, assess or recommend for reference, where the provider and the deployer may not use the output as the sole basis for the final decision. The Prime Minister may still list such systems, stating grounds, where measurable technical or practical evidence shows serious risks or vulnerabilities capable of directly harming defence, security, public order or social safety.

Decision 33/2026/QĐ-TTg, effective 15 August 2026, sets the high-risk list. The Government’s published summary names two banking systems, those that automatically execute electronic banking transactions and those that automatically decide on credit, and no securities or investment-advice system. The signed annex is a scanned image and was not read in full.

What the experiments found

Steven Shaw and Gideon Nave of the Wharton School call the pattern “cognitive surrender”: adopting AI output with minimal scrutiny. Their preprint, published in January 2026 and not yet peer-reviewed, reports three preregistered experiments with 1,372 participants, recruited in a university laboratory and online, on an adapted Cognitive Reflection Test, a set of questions with an intuitive but wrong answer. The researchers secretly varied whether the AI assistant answered correctly.

In the first study, participants answered 45.8% of questions correctly without AI, 71.0% when the AI was right and 31.5% when it was wrong. On trials where the AI was wrong and they consulted it, they followed it roughly 4 times in 5. Separately, participants with access to AI reported confidence 11.7 percentage points higher than those without, although about half the AI’s answers were wrong. In the third study, payment plus feedback raised rejection of consulted faulty AI answers from 20.0% to 42.3%; among participants consulting AI on at least two trials, the intervention group’s accuracy was 84.8% on AI-accurate trials and 40.6% on AI-faulty trials. The authors note that the task is narrow and that field tests in financial applications are needed.

A second paper, by Grace Liu and colleagues at Carnegie Mellon, Oxford, MIT and UCLA, posted on arXiv and marked as a COLM 2026 conference paper, recruited 1,222 participants for randomised trials; 1,060 remained after exclusions. After about 10 minutes of AI help on fractions, participants in the first experiment solved on average 57% of test problems without AI, against 73% for a control group. They were also more likely to give up, although that effect was not significant in the second experiment. The effects were measured immediately, and the authors say their durability is unknown.

A third line of research concerns models that agree with their users. Researchers at Anthropic found that five 2023-generation AI assistants tilted answers towards the user’s stated view. A study published in Science in March 2026 found that 11 models affirmed users’ actions 49% more often than humans did, and that users trusted and preferred the flattering models; that study concerned personal advice, not finance.

Counter-evidence and another limitation

A preprint by Conlon and Schwardmann, published in July 2026 with 1,500 participants across 30 decision tasks, found that AI advice on average moved people away from their initial leanings, even though the model was measurably sycophantic. More sycophancy weakened that effect, and participants did not prefer the more sycophantic advice.

A separate limit concerns the advice itself. A preprint by Ross and Lo of MIT found that the allocations language models recommend are largely determined by the risk tolerance the user reports.

Where investing comes in

The studies reviewed do not measure the effect of AI chatbots on real trading or investment returns, in Vietnam or elsewhere. Brad Barber and Terrance Odean studied 66,465 US households at a discount broker from 1991 to 1996 and found that those who traded most earned 11.4% a year after transaction costs while the market returned 17.9%; they proposed overconfidence as an explanation. That AI-raised confidence would lead to more trading and lower returns is a hypothesis the evidence reviewed does not test.

The Vietnam Securities Depository and Clearing Corporation counted 13,814,693 domestic trading accounts held by individuals at 31 August 2026; one person can hold several. A draft decree reported by the Ministry of Finance’s magazine on 1 September would let securities firms test AI investment advice and AI trading in a sandbox for professional investors only; AI trading would be capped at 1,000 investors per solution and limited to shares outside the VN All Share basket. It remains a draft, and its text was not located.

The warnings reviewed concern unlicensed advice and fraud. The State Securities Commission warned on 17 April 2026 that unlicensed buy, sell or hold recommendations on social media breach the Securities Law, and the Ministry of Public Security lists scams using AI, deepfakes and cloned voices among 25 fraud scenarios for 2026.