Synthesis, Frontier Questions, And Student Research Designs

Learning Objectives¶

By the end of Week 6, students should be able to:

1. convert a broad technology topic into a labor-economics research question with a clear outcome, mechanism, comparison, and welfare object;

2. distinguish technology invention, adoption, use, exposure, diffusion, and organizational implementation as separate empirical margins;

3. choose an appropriate unit of analysis for technology-and-labor evidence: worker, occupation, firm, establishment, commuting zone, sector, platform, or country;

4. evaluate what the existing literature measures well and where measurement remains weak;

5. identify credible frontier projects on AI, firm organization, worker adjustment, market power, institutions, global service trade, infrastructure, and welfare incidence;

6. diagnose common failure modes in this literature before investing in a full research design.

The capstone question is this: what makes a technology-and-labor project recognizably a labor-economics paper rather than a generic technology, business, macro, industrial-organization, or management study?

Opening Orientation¶

The course began with tasks and adjustment, then moved through labor demand, worker adaptation, firm adoption, and market-level incidence. Week 6 turns that sequence into a research architecture. The aim is practical: students should leave with a disciplined way to launch projects, evaluate papers, and decide whether a technology question belongs in labor economics.

The key shift is to stop asking only whether “technology affects jobs.” A credible project must specify the technology margin, the labor outcome, the worker or firm adjustment channel, the relevant market, the counterfactual, and the welfare or distributional object. The same AI tool can be a worker-use measure in one design, a firm-adoption treatment in another, a task-exposure proxy in a third, and a platform-governance institution in a fourth. Those are not interchangeable treatments.

Week 6 is therefore a capstone rather than a recap. It synthesizes the course into a design workflow for research on labor demand, worker adjustment, firms, market structure, institutions, inequality, and welfare.

Bridge¶

Weeks 1 through 5 can be summarized as six design objects:

1. Technology margin: what changed technologically, and is the observed measure invention, adoption, worker use, exposure, diffusion, vacancy demand, or organizational implementation?

2. Affected labor outcome: what moved: employment, wages, earnings, hours, vacancies, task content, training, internal mobility, promotion, authority, job quality, labor share, bargaining power, or welfare?

3. Worker adjustment channel: did workers reskill, switch tasks, move occupations, change firms, retire, retrain, change hours, or bear risk?

4. Firm response channel: did firms substitute, augment, scale output, hire around incumbents, train incumbents, redesign jobs, change supervision, alter promotion ladders, or reorganize production?

5. Market or equilibrium incidence: how did effects aggregate through local labor markets, sectors, platforms, employer concentration, product-market competition, outsourcing, or institutions?

6. Welfare or distributional object: who gained, who lost, who captured rents, who bore adjustment costs, and what changed beyond observed wages or headcounts?

This bridge disciplines the capstone. A paper that only names a technology has not yet stated a labor question. A paper that only estimates an average wage or employment effect has not yet explained the mechanism or welfare object.

Field Core¶

The Technology-To-Labor Research Architecture¶

A durable project usually moves through the following sequence.

Step 1: formulate the labor question. Start with the labor-market object, not the technology label. “Does generative AI matter?” is too broad. “Does worker use of generative AI compress performance differences among customer-support agents, and does the firm convert those gains into wages, promotions, or monitoring?” is a labor question.

Step 2: choose the technology measure. The measure should match the economic object. Robot installations are close to embodied automation adoption. Patent text measures innovation direction. Job postings capture intended demand. Firm surveys capture reported adoption. Worker-use data capture realized use. Platform logs can capture assignment, monitoring, and governance. These measures should not be treated as interchangeable Acemoglu & Restrepo, 2020Webb, 2020Kogan et al., 2023Bick et al., 2024.

Step 3: choose the unit of analysis. Worker-level designs are strong for adjustment, heterogeneity, and welfare. Firm-level designs are strong for adoption, organization, hiring, training, and composition. Occupation-level designs are strong for task exposure but weaker on actual adoption. Establishment and team designs can identify implementation. Local labor-market designs capture broader incidence but require careful exposure construction. Country or cross-establishment designs are valuable for diffusion, absorptive capacity, informality, and global service trade.

Step 4: specify timing and comparison. Adoption is selected, usage can be endogenous, and worker adjustment can be delayed. A persuasive design explains why the comparison is informative: exposed versus less exposed places, earlier versus later adopters, treated versus untreated teams, workers near versus far from an internal rollout, or a policy/institutional shift that changes technology incentives.

Step 5: name mechanisms and welfare. Displacement, augmentation, scale, training, internal reallocation, platform control, rent capture, and institutional mediation imply different welfare interpretations. Wages and employment are often necessary but insufficient. Job quality, schedule risk, autonomy, bargaining power, promotion prospects, mobility costs, and rent sharing may be the real labor objects.

Research Design Template¶

Where The Literature Is¶

The literature is strong where technology measures, labor outcomes, and adjustment units line up well.

Task-based frameworks are the core vocabulary. Autor, Levy, and Murnane show how computerization changes routine and nonroutine task demand Autor et al., 2003. Acemoglu and Restrepo formalize automation, displacement, productivity, and new-task creation Acemoglu & Restrepo, 2019. These papers remain essential because they keep the field from treating occupations as indivisible or technology as a single aggregate shock.

Automation and robot evidence is one of the best-developed empirical areas. Local labor-market exposure designs show how industry composition can translate robot adoption into wages, employment, and labor-force participation across places Acemoglu & Restrepo, 2020Dauth et al., 2021. The strength is clear exposure construction and market-level incidence. The limitation is that robot adoption is narrower than software, AI, platform governance, or generative-AI use.

Worker-adjustment evidence has become richer. Research on skill obsolescence, fields of study, occupational mobility, training, and displacement emphasizes that aggregate adaptation can coexist with individual welfare losses Gathmann & Schoenberg, 2010Deming & Noray, 2020Katz et al., 2022. The literature measures earnings and mobility better than it measures information frictions, perceived risk, job amenities, family constraints, and the subjective cost of retraining.

Firm-adoption and organization evidence is expanding quickly. The IT literature already showed that productivity effects depend on skills, management, and organizational complements Bresnahan et al., 2002Bloom et al., 2012. AI research now measures firm investment, AI-skilled hiring, worker use, use cases, training, manager beliefs, and algorithm aversion Babina et al., 2023Aghion et al., 2025McElheran et al., 2025Dargnies et al., 2026.

Inequality, rents, market structure, and institutions are also central. Superstar-firm and labor-share evidence links technology, market shares, concentration, and rent capture Autor et al., 2020. Patent-rent and firm-wage-premium evidence shows that innovation gains can reach workers, but unevenly across firms and worker groups Kline et al., 2019Card et al., 2018. Labor-market concentration and institutions matter because technology changes outside options, bargaining, platform rules, monitoring, and worker voice Azar et al., 2022Benmelech et al., 2022.

Where the field is / Where it is going¶

What Is Measured Well And Poorly¶

The field measures some objects relatively well:

• embodied automation in manufacturing through robot data;

• task exposure through occupation and task databases;

• job-posting demand for skills and tools;

• firm AI investment and AI-skilled hiring in some linked datasets;

• local labor-market outcomes such as employment, wages, and labor-force participation;

• short-run worker productivity in some randomized or rollout settings.

Other objects remain weakly measured:

• actual worker use rather than access or exposure;

• organizational implementation after adoption;

• internal task reassignment and promotion effects;

• incumbent training versus external hiring around incumbents;

• algorithmic management, monitoring intensity, and worker autonomy;

• job quality, risk, dignity, and voice;

• distribution of rents within firms and platforms;

• general-equilibrium spillovers across places, sectors, and countries;

• compute, electricity, and data-center constraints as labor-market forces.

This measurement gap is not a nuisance. It is one of the highest-return research opportunities in the field.

Where The Literature Is Going¶

The first frontier is generative AI use versus firm adoption. Worker-use measures show who experiments with tools and how usage diffuses Bick et al., 2024. Field evidence shows large short-run productivity effects in some tasks, but the long-run labor question is whether firms convert those gains into wages, promotions, monitoring, fewer workers, or redesigned jobs Brynjolfsson et al., 2025Dell'Acqua et al., 2026Humlum & Vestergaard, 2025.

The second frontier is complementary organizational change. AI adoption may require workflow redesign, data systems, training, manager trust, legal routines, and new evaluation standards. Designs that observe only tool access will miss whether the firm reorganized around the technology.

The third frontier is energy, electricity, and compute infrastructure. Data centers, grid capacity, electricity prices, cloud dependence, and compute bottlenecks may shape where AI can be adopted and which local labor markets receive construction, operations, technical, and service demand International Energy Agency, 2026. The labor object is spatial incidence and adoption capacity, not generic energy commentary.

The fourth frontier is outsourcing, global service trade, and the new geography of work. AI can automate some traded service tasks, augment others, and create demand for oversight, localization, data work, and infrastructure. The same technology may reduce one outsourcing channel while expanding another Stiglitz et al., 2021Liu & others, 2025Rodrik et al., 2024.

The fifth frontier is measurement of technology and innovation. Patents, text embeddings, vacancies, surveys, worker-use logs, vendor records, and administrative data identify different margins. Papers that validate one measure against another can be major contributions, especially when they clarify which labor outcome each measure should predict.

The sixth frontier is labor-market power, platform governance, and worker voice. Technology can change wage setting, monitoring, scheduling, dispute resolution, transparency, and the right to contest algorithmic decisions. These are labor institutions, even when implemented in software.

Research Opportunities¶

The best student projects do not need to settle the entire technology debate. They need a credible margin, a labor object, and a mechanism that matters beyond the setting.

Frontier Project Opportunities Map¶

Some areas are relatively saturated as broad questions but still open in narrower designs. “Do robots reduce manufacturing employment?” is heavily studied in high-income settings. A stronger entry point might ask how robot exposure interacts with retraining, worker age, firm wage premia, or institutions. “Does generative AI raise productivity in writing tasks?” already has strong short-run evidence. A better project might ask whether those gains persist after organizational redesign, whether lower-performing workers capture durable wage gains, or whether firms use AI to change monitoring and promotion.

Other areas remain open. Worker-level AI use linked to wages, promotion, and mobility is still thin. Firm adoption versus actual usage is not yet well mapped. Energy and compute infrastructure are barely integrated into labor-market incidence. Platform governance and algorithmic worker voice need designs that connect software rules to bargaining, risk, and dispute outcomes. Global service trade after AI is wide open if the paper can measure task tradability, outsourcing exposure, and destination-sector job quality.

Measurement contributions are especially valuable. A paper that validates a technology measure, separates adoption from use, or shows which exposure proxy predicts which labor outcome can reshape later causal work.

Research Lab¶

The Week 6 research lab follows Reproduce -> Diagnose -> Transfer.

The primary anchor is Acemoglu and Restrepo’s local robot-exposure design Acemoglu & Restrepo, 2020. It is central to the field, bounded enough for a teaching reproduction, and useful for showing how a technology measure becomes a local labor-market exposure object.

The challenge and transfer anchors are newer AI-use and firm-adoption designs. Brynjolfsson, Li, and Raymond provide a worker-use and productivity benchmark for generative AI Brynjolfsson et al., 2025. Aghion, Bunel, Jaravel, Mikaelsen, Roulet, and Sogaard push toward firm-level AI adoption by use case Aghion et al., 2025. These anchors help students see why old exposure designs do not mechanically answer newer questions about actual use, organization, and incidence.

The lab does not claim to replicate official estimates. No official robot, Census, firm-adoption, worker telemetry, or confidential linked employer-employee data are bundled in this repository. The teaching path uses deterministic synthetic data to practice the design logic.

Reproduce. Students construct a local exposure measure using baseline industry employment shares and technology adoption intensity. They estimate a simple reduced-form relationship between exposure and labor-market outcomes, then interpret it as design logic rather than published evidence.

Diagnose. Students classify what the exposure measure captures and misses. They ask whether it measures adoption or exposure, whether the unit is the right unit, whether timing is plausible, whether firm responses are hidden, and whether wages and employment are sufficient welfare objects.

Transfer. Students redesign the exposure logic for generative AI. They move from local robot exposure to worker use and firm adoption, then evaluate whether the new design should use workers, firms, occupations, teams, platforms, or places as the primary unit.

Methods Box¶

Common Failure Modes¶

Several mistakes recur in technology-and-labor projects.

1. Technology-first framing. The project studies AI, robots, platforms, or patents because they are salient, but never states the labor mechanism.

2. Proxy overclaiming. A patent, vacancy, exposure score, or survey response is treated as the technology itself.

3. Exposure-adoption confusion. A design infers firm adoption or worker use from exposure measures that only capture potential task susceptibility.

4. Selected adoption without a strategy. Early adopters differ in size, management, skills, productivity, and growth trajectories.

5. Partial-equilibrium interpretation of equilibrium evidence, or the reverse. Worker, firm, and local-market designs identify different objects.

6. Timing compression. Short-run productivity, medium-run hiring, and long-run worker welfare are collapsed into one coefficient.

7. Ignoring firms. A worker-exposure paper misses hiring, training, monitoring, and internal reallocation.

8. Ignoring workers. A firm-adoption paper treats workforce composition as if it were welfare.

9. Weak external-validity claim. A sector, country, platform, or firm is described as interesting without explaining what mechanism travels.

10. Wage-only welfare. Wages are interpreted without hours, risk, autonomy, schedules, bargaining power, job quality, and mobility costs.

Portability and Relevance Map¶

Reading Ladder And References¶

Start with the conceptual frame: task content, automation, augmentation, and new tasks Autor et al., 2003Autor, 2015Acemoglu & Restrepo, 2019. Then read local labor-market automation evidence to see how exposure designs work Acemoglu & Restrepo, 2020Dauth et al., 2021.

Next, add worker adjustment and career risk Gathmann & Schoenberg, 2010Deming & Noray, 2020Katz et al., 2022. Then move inside firms through IT complements, AI investment, AI adoption by use case, usage, and algorithm aversion Bresnahan et al., 2002Babina et al., 2023Aghion et al., 2025Brynjolfsson et al., 2025Dargnies et al., 2026.

Finally, read market-level and institutional work on labor share, rents, employer concentration, global service transitions, and AI infrastructure Autor et al., 2020Kline et al., 2019Benmelech et al., 2022Stiglitz et al., 2021International Energy Agency, 2026.

A productive reading path for student projects is:

1. one foundational task or automation framework;

2. one canonical exposure or adoption design;

3. one worker-adjustment paper;

4. one firm-organization or worker-use paper;

5. one market-structure, institution, or welfare paper;

6. one frontier paper closest to the student’s proposed setting.

Exercises And Discussion Prompts¶

1. Take a technology topic you care about and rewrite it as a labor question. Name the labor outcome, adjustment channel, unit, comparison, and welfare object.

2. Compare occupation exposure, firm adoption, and worker-use measures for generative AI. What does each measure identify, and what does each miss?

3. Design a study that separates incumbent retraining from firms hiring around incumbents after AI adoption.

4. Give an example where an employment increase could still reduce worker welfare.

5. Explain how a local exposure design and a firm-adoption design can both be correct but answer different questions.

6. Pick a country, sector, platform, or occupation. What mechanism would make evidence from that setting travel to labor economists elsewhere?

7. State one saturated question in technology-and-labor research and one still-open version of that question.

8. Propose a measurement contribution that would help future causal work.

9. How would you study worker voice or bargaining power during a technology transition?

10. How could compute infrastructure or electricity constraints become a labor-market research design?

Reproducibility And Code Lab Note¶

The Week 6 code lab lives at labs/06-synthesis-frontier-questions-and-student-research-designs/. It is a bounded synthetic teaching path that implements the capstone Reproduce -> Diagnose -> Transfer workflow. The smoke test constructs a local exposure design, writes a simple reduced-form incidence table, diagnoses whether the measure captures adoption, use, or exposure, and transfers the design to generative-AI worker use, firm adoption, organization, and market-incidence questions.

The lab requires no proprietary data, official replication packages, robot installation files, linked firm-worker microdata, platform logs, or external downloads. It is student-facing by design: the purpose is to practice how research design choices map into labor questions.

Slide Companion Note¶

The Week 6 slide deck lives at slides/week6/06-synthesis-frontier-questions-and-student-research-designs.tex. The deck does not duplicate the chapter. It defines the capstone question, lays out the technology-to-labor research architecture, maps where the literature is, identifies frontier directions, isolates research-design choices and common failure modes, and closes with the Reproduce -> Diagnose -> Transfer lab workflow.

Bridge Forward¶

The course ends by turning technology into a disciplined labor research agenda. The durable lesson is that frontier projects are not the ones with the newest label. They are the ones that make technology empirically legible as a question about labor demand, worker adjustment, firms, institutions, inequality, market incidence, and welfare.