LGLi Gan

Li Gan

*This is an algorithmically generated author record based on publicly available information.

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article


Li Gan has a total of 15 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Open Access

    When Delegation Destroys Verification: Endogenous Unfalsifiability in Human-AI Organizations

    by Li Gan

    When an organization delegates a task to an artificial intelligence, it keeps the output but loses the practice through which it learned to check that output. This paper studies how that loss affects the organization's ability to detect biased AI advice. Each period, work is allocated between human execution skill and AI capability to maximize current output, the allocation that competitive discipline enforces. Both inputs improve with use. The results are comparative dynamics along this deployment path, stated as conditions on parameters. Execution practice, and the evaluation skill it builds, vanishes when the elasticity of substitution exceeds one plus the inverse of the AI's learning elasticity. Above a higher threshold, output data also stop carrying information about bias, and the power of a test against a local bias converges to its size: endogenous unfalsifiability. The expected loss from bias then rises to its level without oversight. Below these thresholds the organization keeps human evaluation skill or informative data. The decay requires no strategic behavior by the AI. With heterogeneous agents and complementary investments in evaluation skill, the economy can settle in a hollow equilibrium in which no one invests. A temporary subsidy can move it out, while Pigouvian and redistributive instruments address the remaining margins.

  • Open Access

    Execution and Evaluation: Task Content, Labor Demand, and the Evaluator Pipeline

    by Li Gan

    Measures of occupational AI exposure ask whether machines can perform a task, not which function the human supplies. I distinguish execution, producing a first-pass output, from evaluation, judging whether it is correct, and make the distinction measurable by scoring all 19,265 O*NET task statements under fixed rubrics. The execution score is reproducible across model coders and O*NET vintages. In a 200-task random audit it correlates 0.81 with the mean of four blinded human coders. It predicts each held-out coder's peers better than that coder does, though the coders disagree substantially among themselves. It adds only modest power beyond O*NET's evaluation activities. Employment growth is lower in execution-heavy white-collar occupations in every window since 2012; the gradient shows no post-2022 break and lies largely between occupational families. The Eloundou exposure gradient is steeper in 2022-2025 than in the pandemic window, but the design does not establish an AI effect. A near-term seniority test is null. If evaluative judgment is itself produced by execution practice, the supply of evaluators falls only after a cohort lag; that prediction remains untested.

  • Open Access

    VAT Credits as Interest-Free Loans: Valuing Delay and Setting Refund Priorities

    by Li Gan

    When a firm's input VAT exceeds its output VAT, the government owes it thedifference. Many governments carry the credit forward instead of refunding it,so the unpaid credit is an interest-free loan from the firm to the state. UsingChinese tax returns for 2016--2020, we value these loans and ask which claims anadministration should pay first. Slow-clearing accounts pile up in the backlog,but balances also fall while accounts stay open. Because the government alsobenefits from delay, the gain from paying early is about half the firm's loss:6.3 against 14.0 cents per yuan forslow-clearing balances. The best refund rule depends on what limits theadministration. When staff can verify only a limited number of claims, payingthe largest balances first captures 99.9 percent of the maximum gain. When cashbinds, paying the accounts least likely to clear first captures 95.3 percent of the upper bound, against 68.4 percentfor random order. It works because accounts use their credits while they stayopen.

  • Open Access

    Risk Ceilings and Development Deadlines: Pacing AI under Uncertain Safety Productivity

    by Li Gan

    Can a regulator promise both a risk ceiling and a development deadline when safety productivity is unknown? A ceiling below the final model's unprotected hazard requires a minimum stock of safety knowledge, so both promises hold only if weak research can be ruled out. Learning first and then replaying development, with spare compute in safety, needs productivity 21 percent above that lower bound, or 3.3 percent with state risk alone. The policy holds catastrophe risk to 8 percent, against 75 under full speed, and costs customers 13 percent of five-year services. Rules that fix compute allocation fix dates, not risk.

  • Open Access

    Who Buys Growth Accelerations? Purchaser Routes and Their Persistence

    by Li Gan

    A common development view holds that rapid growth begins with exports and later shifts to domestic demand. Using previous-year-price input-output accounts for 80 economies in 1995-2022, I allocate domestic value added to households, government, investment, and foreign demand. Foreign demand leads 19 of 36 dated growth accelerations, versus 15 for households, and is more prevalent in smaller economies. No general ten-year handoff appears. Among 17 onsets led by foreign demand over the first seven years, foreign demand leads years eight to ten in 11, but foreign demand leads 6 and domestic routes lead 6 among the 12 windows ending by 2019. Net exports obscure this support: foreign-demand-supported domestic value added rises in all 36 accelerations despite nonpositive net-export contributions in 22. Route-specific prices change 83 of 400 five-year engine assignments and halve apparent government leadership. Purchaser routes describe realized growth composition, not policies for generating it.

  • Open Access

    Network Incidence and VAT Rate Design with Incomplete Crediting

    by Li Gan

    Should upstream sectors face higher or lower VAT rates? The cascading doctrine favors lower upstream rates; captive-base incidence favors higher. We derive both within one welfare objective in a production network with seller-buyer-specific crediting. At a common crediting rate, the incidence and production blocks imply opposing patterns and nearly cancel. Wherever crediting mostly works, model-implied spreads never exceed about two points, against fourteen in statute, across 76 input-output tables and 211-sector Chinese calibrations. China's 2018-2019 cuts realized a third of the attainable compression; the announced merger delivers most of the rest. Where informality removes the downstream base, the prescription reverses.

  • Open Access

    Who Pays, Who Benefits? Public Provision and the Service-sector Income Gradient

    by Li Gan

    Households devote a larger share of consumption to services as income rises, yet their purchases support a smaller share of service-sector expansion in richer economies. A purchaser account for 76 economies documents this negative household-contribution gradient. I ask how much of that established gradient reflects the convention that national accounts classify publicly financed individual services as government demand. Household demand scale, not its service intensity, accounts for about nine tenths of the payer gradient, whose main accounting counterpart is exports. I then trace government purchases of education, health, and social work through domestic production networks and assign the induced value added to household beneficiaries at recorded cost. On the broader service boundary, this reclassification changes the income slope from-10.4 to-6.2 percentage points per log point of initial income. The adjustment removes 41 percent of the payer gradient, with a cluster-bootstrap interval of 21 to 66 percent; on the market-service boundary it removes only 6 percent. The same accounts applied across Chinese provinces yield a negative payer gradient but a beneficiary correction that is larger in poorer provinces and steepens the level relationship. Public provision attenuates the international gradient, but its magnitude and non-household counterparts differ across settings. These are current-price accounting relationships, not causal income effects or welfare estimates.

  • Open Access

    Two Types of Tertiarization: Household Demand, Production Networks, and the Rise of Services 

    by Li Gan

    I measure whether the rise of services is ultimately supported by household consumption or by non-household demand through production networks, investment, exports, and government. Using harmonized input-output tables for 76 economies over 1995-2022, I attribute service value added to its ultimate domestic final-demand category and measure both the domestic-household orientation of each service economy in levels, R, and the household contribution to service-valueadded growth, H; all objects are current-price value-added statistics, and the tracing stops at the border (exports are a terminal category). The conventional service share does not reveal the type: it is negatively correlated with R, contains little incremental information about R conditional on income, and about four in ten tertiarization episodes are majority-household. Among the 65 economies whose service share rose over 1995-2018, 77 percent became less householdoriented; the share is 75 percent under a two-percentage-point material-change threshold. Yet the service intensity of household-supported production rises with income. The exact identity R = α C σ C /s S resolves the apparent contradiction: aggregate household orientation, α C , falls faster than household-supported service intensity, σ C , rises. A one-type benchmark generates service deepening but has no purchaser margin. A two-type extension reveals the missing margin: within countries, nonhomothetic demand and saving move the investment-supported share against the household-supported share, and the extension gives the condition for the reversal. The accounting analysis then uses the exact Leontief system to show when service deepening in production networks raises the service share while lowering R. I then map the service share and R exactly into six accounting-feasible components and use a Shapley decomposition to quantify their contributions. Among economies with materially rising shares, the combined production-system channels (production structure plus investment and export demand composition, summed within each economy) are allocated a median country-level share of 73 percent of the service-share rise and are the largest contribution in 80 percent of those economies, while the decline in household orientation is allocated overwhelmingly to the aggregate final-demand structure; changes in household-consumption composition push in the opposite direction but are outweighed. The service share records how much service activity exists, not which economic transformation produced it.

  • Open Access

    Insurance or deliberate use of the law for financial gain? Testing for heterogeneous filing behavior in personal bankruptcy *

    by Li Gan, Manuel A. Hernandez, Shuoxun Zhang

    Personal bankruptcy can serve as insurance for large financial shocks but may also provide an opportunity for deliberate use. This paper proposes and implements a model to test for the coexistence of heterogeneous filing behaviors in the bankruptcy decision. We identify two different filing behaviors using data from the United States. One group can be associated with individuals who do not intend to deliberately use the law in the absence of an adverse event, while the second group is more consistent with a "strategic" behavior that also exhibits a higher probability of filing for bankruptcy and financial benefit from filing. The larger prevalence of the first group supports the insurance function of bankruptcy. We further show that the proposed model can help to better identify eventual filers compared to standard probabilistic models that do not allow for heterogeneous filers.