On AI Governance and Job Displacement: A Comparative Study of AI Policy in Vietnam and The United States

My-Duong Trinh Hoang, Duy-Anh Nguyen, Uyen Nhi Dinh, Nguyễn Minh Huy

Artificial Intelligence (AI) is rapidly reshaping economies, governance systems, and labor markets worldwide. As a result, AI governance has become a central policy issue, with countries adopting distinct regulatory approaches shaped by their political and economic systems.

This paper compares AI policies in Vietnam and the United States, focusing on how each country addresses the challenge of labor displacement. Vietnam promotes AI as a strategic instrument for national development, digital transformation, and industrial modernization within a socialist-oriented market economy. In contrast, the United States relies on a market-driven innovation model led by the private sector, while confronting growing concerns over AI-related layoffs and employment insecurity. The study argues that AI policy is not merely a response to technological change but a reflection of broader development strategies, institutional structures, and state-market relations. By examining these contrasting approaches, the paper contributes to understanding how different governance models shape responses to the social and economic consequences of AI-driven transformation.

Reviewer's Comments

Reviewer's Comments

Arrow
Arrow
Arrow

This paper brings together a useful body of Vietnamese AI-policy detail (Resolution 57, the 2025 AI Law and its risk-tiering, the National Digital Transformation agenda) that is hard to find synthesized in English, and it raises a worthwhile under-examined question: why Vietnamese attitudes toward AI remain so positive despite real labor-market exposure. Several steps would substantially raise its analytical contribution.

The most important is method transparency. Beyond reading relevant policy outputs, the paper does not specify an analytical approach. It is not clear whether the qualitative analysis proceeds by document analysis, discourse analysis, or process tracing, and because the sources, though listed, are not cited in-text, it is difficult to connect specific claims to specific evidence. This matters most for the broad discourse-level generalizations about how "Vietnamese society" and "American society" view AI, which are stated without a procedure that would let a reader see how they were reached. Naming the method and adding in-text attribution would make the argument far easier to follow and assess.

Second, the paper sets up its own central thesis but does not test it. It states that Vietnam and the US operate a socialist-oriented market economy and a market-led capitalist one respectively, and that AI policy reflects state-market relations, but it does not then link those political configurations to the specific features of each country's AI policy design. Drawing that connection explicitly, showing how the political setup produces the observed policy posture rather than merely sitting alongside it, is the analytical payoff the framing promises and would transform the paper.

Relatedly, the two cases are presented in parallel but not systematically compared. The four comparative dimensions announced in the methods (formation context, content, practices, outcomes) would be much more powerful executed as a structured side-by-side matrix than as sequential country narratives, and would also help rebalance a comparison that currently weighs heavily toward Vietnam.

Third, the quantitative material is presented but not analyzed. The figures are a welcome inclusion, but indicators such as GDP, inflation, and unemployment are largely displayed rather than connected to the central claim about AI and labor displacement, and the "quantitative methods" described in the design are not employed with enough robustness to generate findings. Tying each data series explicitly to the argument, or trimming those that do not bear on it, would tighten the empirical core considerably.

Finally, the headline observation, that low Vietnamese resentment toward AI may reflect shallow adoption, currently rests mainly on a single industry-sponsored survey (AWS-Strand). Testing it against alternative explanations and grounding it in the rigorous sources already on the reference list, especially the ILO's 2026 Vietnam exposure brief and Brynjolfsson's recent employment-effects work, both cited but not analytically used, would give this promising question a much firmer answer. The raw material for a strong comparative case study is here; the next step is adding a robust analytical method.

This project addresses an important and timely policy issue: how AI adoption may affect labor markets and how different governance models respond to that challenge. The comparison between Vietnam and the United States offers a useful starting point for understanding how national development strategies, institutional structures, and public attitudes shape AI policy priorities. To strengthen the contribution further, the analysis could more clearly connect labor displacement to the hackathon’s AI safety framing, for example by explaining how workforce disruption becomes a governance or societal risk and what concrete policy mechanisms could mitigate it. The methodology would also benefit from making the quantitative component more visible, through clearer indicators, comparative tables, or a more explicit explanation of how the evidence supports the main findings. Overall, this is an informative and relevant comparative policy essay with good potential, especially if future work deepens the evidence base and translates the comparison into more actionable governance recommendations.

The comparison reads clearly and the Vietnam-side is well done: the policy trajectory from Resolution 52-NQ/TW through the 2025 AI Law is a useful, well-sourced synthesis, and "AI policy mirrors each country's development model" is a fair thesis. The gap is between what you set up and what you delivered. You promise a quantitative strand (AI-exposure indices, ILO, OECD, BLS) but the analysis stays descriptive, with none of the tables or comparisons that would back it, and the US half is much thinner than the Vietnam half, which is a problem when the comparison is the whole point. Three things to consider: put at least one real comparative figure on the page from the data you already cite, build the US side out to match Vietnam, and fix the section numbering, since it jumps around. Also say the safety part out loud, which displacement gaps create risks that are not just economic. Solid base, needs its evidence layer.

Cite this work

@misc {

title={

(HckPrj) On AI Governance and Job Displacement: A Comparative Study of AI Policy in Vietnam and The United States

},

author={

My-Duong Trinh Hoang, Duy-Anh Nguyen, Uyen Nhi Dinh, Nguyễn Minh Huy

},

date={

},

organization={Apart Research},

note={Research submission to the research sprint hosted by Apart.},

howpublished={https://apartresearch.com}

}

Recent Projects

OliGraph: graph-based screening of large oligopools

Existing synthesis screening tools cannot evaluate short oligonucleotide pools, whose overlapping fragments can be reassembled into regulated sequences via polymerase cycling assembly (PCA) yet fall below gene-length detection thresholds. We present OliGraph, an open-source tool that constructs a bi-directed overlap graph from an oligonucleotide pool and extracts contigs for downstream gene-length screening. An optional PCA mode retains only cross-strand overlaps consistent with PCA chemistry. We validated OliGraph in a blinded study across ten simulated pools (70–9,184 oligonucleotides, 30–300 bp) spanning four risk categories. BLAST screening of individual oligonucleotides failed to identify sequences of concern in most pools: three returned zero hits, and vector noise obscured true positives in the remainder. After OliGraph assembly, contig-level BLAST matched the longest assembled sequences (up to 1,905 bp) to sequences of concern at 97–100% identity. In one pool, assembly collapsed 1,634 individual BLAST results into 10 hits from a single contig, all assigned to the same source organism. PCA mode correctly distinguished assemblable from non-assemblable fragments within the same pool. Two pools with no assemblable structure yielded no contigs. OliGraph processed all pools in under 0.2 seconds, fast enough for real-time order screening and consistent with proposals to bring oligonucleotide orders within the scope of synthesis screening regulation.

Read More

BioRT-Bench: A Multi-Attack Red-Teaming Benchmark for Bio-Misuse Safeguards in Frontier LLMs

Frontier AI laboratories are expected to maintain safeguards against biological misuse, but whether deployed models actually refuse bio-misuse queries under adversarial pressure is largely unmeasured in the public literature. We introduce BioRT-Bench, a benchmark that runs four attack methods (direct request, PAIR, Crescendo, and base64 encoding) against four frontier models (Claude Sonnet 4.6, GPT-5.4, DeepSeek V4-flash, Kimi K2.5) across 40 prompts spanning five biosecurity-relevant categories. Responses are scored by a calibrated judge extending StrongREJECT with two bio-specific dimensions: specificity and actionability. We measure Attack Success Rate (ASR), where 0 means the model fully refused and 1 means it provided specific, actionable bio-misuse content. Our results reveal a sharp robustness divide: Chinese frontier models (DeepSeek, Kimi) have under 5% refusal rates even under direct request (ASR 0.88 and 0.79), while Western models (Claude, GPT) maintain substantially stronger safeguards (ASR 0.15 and 0.16). Crescendo is the most effective attack across all models, both in bypassing refusal and in eliciting actionable content. Claude Sonnet 4.6 is the most robust model tested, achieving 100% refusal against base64-encoded prompts.

Read More

PROTEUS (PROTein Evaluation for Unusual Sequences): Structure-Informed Safety Screening for de novo and Evasion-Prone Protein-Coding Sequences

AI protein design tools like RFdiffusion, ProteinMPNN, and Bindcraft make it trivial to produce low-homology sequences that fold into active, potentially hazardous architectures. However, sequence homology-based biosafety screening tools cannot detect proteins that pose functional risk through structurally novel mechanisms with no sequence precedent. We present a tiered computational pipeline that addresses this gap by combining MMseqs2 sequence alignment with structure-based comparison via FoldSeek and DALI against curated toxin databases totaling ~34,000 entries. AlphaFold2-predicted structures are screened for both global fold similarity (FoldSeek) and local active/allosteric site geometry (DALI), capturing convergent functional hazards that sequence screening misses. The pipeline was validated against a panel of toxins, benign proteins, structural mimics, and de novo-designed Munc13 binders, as well as modified ricin variants with residue substitutions. We additionally tested robustness to partial-synthesis evasion, where a bad actor submits multiple shorter coding sequences intended for downstream reassembly into a full toxin-coding gene. We found that while sequence-based screening did not identify any de novo ricin analogues with high certainty, the combined pipeline with FoldSeek and DALI identified all 24 tested de novo ricins as toxic.

Read More

OliGraph: graph-based screening of large oligopools

Existing synthesis screening tools cannot evaluate short oligonucleotide pools, whose overlapping fragments can be reassembled into regulated sequences via polymerase cycling assembly (PCA) yet fall below gene-length detection thresholds. We present OliGraph, an open-source tool that constructs a bi-directed overlap graph from an oligonucleotide pool and extracts contigs for downstream gene-length screening. An optional PCA mode retains only cross-strand overlaps consistent with PCA chemistry. We validated OliGraph in a blinded study across ten simulated pools (70–9,184 oligonucleotides, 30–300 bp) spanning four risk categories. BLAST screening of individual oligonucleotides failed to identify sequences of concern in most pools: three returned zero hits, and vector noise obscured true positives in the remainder. After OliGraph assembly, contig-level BLAST matched the longest assembled sequences (up to 1,905 bp) to sequences of concern at 97–100% identity. In one pool, assembly collapsed 1,634 individual BLAST results into 10 hits from a single contig, all assigned to the same source organism. PCA mode correctly distinguished assemblable from non-assemblable fragments within the same pool. Two pools with no assemblable structure yielded no contigs. OliGraph processed all pools in under 0.2 seconds, fast enough for real-time order screening and consistent with proposals to bring oligonucleotide orders within the scope of synthesis screening regulation.

Read More

BioRT-Bench: A Multi-Attack Red-Teaming Benchmark for Bio-Misuse Safeguards in Frontier LLMs

Frontier AI laboratories are expected to maintain safeguards against biological misuse, but whether deployed models actually refuse bio-misuse queries under adversarial pressure is largely unmeasured in the public literature. We introduce BioRT-Bench, a benchmark that runs four attack methods (direct request, PAIR, Crescendo, and base64 encoding) against four frontier models (Claude Sonnet 4.6, GPT-5.4, DeepSeek V4-flash, Kimi K2.5) across 40 prompts spanning five biosecurity-relevant categories. Responses are scored by a calibrated judge extending StrongREJECT with two bio-specific dimensions: specificity and actionability. We measure Attack Success Rate (ASR), where 0 means the model fully refused and 1 means it provided specific, actionable bio-misuse content. Our results reveal a sharp robustness divide: Chinese frontier models (DeepSeek, Kimi) have under 5% refusal rates even under direct request (ASR 0.88 and 0.79), while Western models (Claude, GPT) maintain substantially stronger safeguards (ASR 0.15 and 0.16). Crescendo is the most effective attack across all models, both in bypassing refusal and in eliciting actionable content. Claude Sonnet 4.6 is the most robust model tested, achieving 100% refusal against base64-encoded prompts.

Read More

This work was done during one weekend by research workshop participants and does not represent the work of Apart Research.
This work was done during one weekend by research workshop participants and does not represent the work of Apart Research.