Counsellor oversight makes AI career recommendations more useful and fair: students report better feasibility-adjusted fit and reduced socioeconomic gaps when human review accompanies algorithmic guidance. Simulations show accuracy- or parity-focused recommenders can steer low-SES students toward infeasible high-opportunity paths unless conversion factors or support are explicitly modelled.

Main Finding

Counsellor review or an AI that is informed about students’ practical constraints (financial, geographic, informational) produces higher feasibility-adjusted career fit (FAF) and far fewer infeasible recommendations than accuracy-optimised AI; AI-only guidance raises unadjusted fit but widens socioeconomic gaps, depresses student agency and increases algorithmic deference. Explanation partially improves calibration. Overall: evaluate career-guidance AI by feasibility, equity, calibrated uncertainty and preserved agency — not predictive accuracy alone.

Key Points

• Central construct: feasibility-adjusted career fit (FAF) = fit to interests/ability × feasibility given conversion factors (financial capacity, mobility, etc.). Grounded in Social Cognitive Career Theory and the Capability Approach.
• Study 1 (simulation, N=20,000):
  • Five policies compared: conventional counsellor heuristic, accuracy-optimised AI, demographic-parity-constrained AI, constraint-aware AI (AI given conversion factors, optimising fit×feasibility), and human-in-the-loop AI (counsellor reviews AI top-5).
  • Accuracy-optimised AI maximised unadjusted fit (0.92) but yielded lower mean simulated FAF (0.721) than conventional guidance (0.729).
  • Accuracy-optimised policy sharply widened SES gaps (Q5–Q1 FAF gap rose from 0.415 → 0.515) and assigned infeasible recommendations to ~51.2% of bottom-quintile students (vs 18.6% under conventional guidance).
  • Demographic-parity constraint equalised access to high-opportunity labels but did not improve FAF and increased infeasibility (bottom-quintile infeasible recommendations 53.6%).
  • Constraint-aware AI achieved highest mean simulated FAF (0.771) and the fewest infeasible recommendations for low-SES students (5.1% bottom quintile); human-in-the-loop produced similar outcomes (FAF 0.769; 7.7% infeasible bottom quintile), implying the value lies in access to conversion-factor information and feasibility objectives, not human judgement per se.
  • Results examined under baseline, technology-boom and downturn labour-market scenarios; patterns were robust.
• Study 2 (randomised experiment, N=987 Indian students across 14 institutions):
  • Four arms: counsellor-only, AI-only, explainable AI, AI + counsellor review.
  • Primary outcome: perceived FAF (self-reported fit and feasibility).
  • Counsellor-reviewed AI produced the highest perceived FAF (effect size d = 0.52 vs AI-only) and the smallest socioeconomic disparity.
  • AI-only guidance reduced reported agency and uncertainty awareness and greatly increased algorithmic deference (d = 1.20); students were more likely to accept recommendations without calibrated judgment.
  • Providing explanations partially restored calibration (reduced deference / increased uncertainty awareness relative to AI-only).
• Prespecified hypotheses largely supported: accuracy-focused AI improves apparent fit but undermines feasibility and equity; feasibility-aware selection or counsellor oversight yields better FAF and equity outcomes; explanation helps but is not a full substitute for feasibility information or human oversight.

Data & Methods

• Theoretical framing: Social Cognitive Career Theory + Capability Approach → focus on conversion factors and real freedom to pursue careers.
• Study 1: Transparent simulation
  • N = 20,000 synthetic student profiles with SES quintile, latent ability, interest vectors (10 sectors), financial capacity, mobility freedom; proxy achievement includes SES-correlated measurement advantage.
  • Careers: 10 sectors with documented opportunity, pathway cost, mobility requirement.
  • Outcome metrics: true fit (interest 0.6, ability–demand 0.4), feasibility (affordability × mobility), FAF = true fit × feasibility; infeasible if feasibility < 0.40.
  • Policies modelled: conventional counsellor heuristic, accuracy-optimised AI, demographic-parity AI constraint, constraint-aware AI, human-in-loop review.
  • Scenarios: baseline, technology-boom, downturn; 200 bootstrap replications for uncertainty.
• Study 2: Randomised field experiment
  • N = 987 secondary and university students in India, randomized to four arms (counsellor-only, AI-only, explainable AI, AI + counsellor review).
  • Outcomes: perceived FAF (primary), decision clarity, agency, trust, fairness perceptions, uncertainty awareness, algorithmic deference, recommendation acceptance.
  • Pre-specified analysis plan and ethics approval; perceived FAF used as primary experimental outcome; effect sizes reported for key contrasts.

Implications for AI Economics

• Evaluation metrics: Economic assessment of student-facing AI must go beyond predictive accuracy to welfare-relevant metrics (e.g., feasibility-adjusted fit, distributional welfare, and rates of infeasible allocations). Accuracy can create negative externalities by misallocating opportunity labels that are practically unattainable for disadvantaged groups.
• Distributional effects and inequality: Recommendation systems can amplify inequality at the critical entry point to the labour market. Economic models of AI adoption should incorporate conversion factors and heterogeneous constraints; welfare analyses should measure not only expected returns but accessibility-adjusted returns.
• Design incentives and market demand:
  • There is likely demand for feasibility-aware recommender systems (and for hybrid systems that combine AI efficiency with constraint information). Firms and public providers that embed conversion-factor data may deliver superior social welfare outcomes.
  • Markets that reward only predictive accuracy (e.g., adoption driven by headline accuracy metrics) risk deploying systems that widen socioeconomic gaps; procurement and procurement metrics should incorporate FAF-like criteria.
• Policy and regulation:
  • Regulatory frameworks (education, labour, high-risk AI rules) should require disclosure and mitigation of distributional harms; demographic-parity constraints can be misleading if not tied to capability/welfare outcomes.
  • Mandates for human oversight (as in some governance proposals) are supportable, but oversight must ensure access to conversion-factor information and an objective that accounts for feasibility — otherwise oversight may be performative.
  • Subsidies or integrated supports (financial aid, mobility assistance, informational campaigns) should accompany feasibility-aware recommendations to convert exposure into real opportunity—otherwise systems that avoid infeasible recommendations may also reduce poor students’ exposure to high-opportunity pathways.
• Cost–benefit and program evaluation:
  • Economic evaluations should compare the marginal cost of adding conversion-factor inputs, counsellor review, or supports versus the social benefits of increased FAF and reduced misallocation (long-term earnings, reduced dropout, social mobility).
  • Randomised and quasi-experimental impact evaluations should measure long-run outcomes (education choices, labour-market entry, earnings) and not rely solely on immediate acceptance or perceived fit.
• Measurement and modelling priorities:
  • Invest in measuring conversion factors at scale (household financial constraints, mobility limits, social capital) and incorporate them into recommender-system inputs and welfare models.
  • Move fairness criteria from label parity to welfare- or capability-based objectives; economic theorists should formalise conversion-factor–aware welfare functions for algorithmic allocation problems.
• Human capital and labour-market modelling:
  • When modelling supply-side responses to new information (recommendations), allow for endogenous agency and trust effects: AI-only advice can alter student agency and uncertainty perception, which in turn affect investment and search behaviour.
  • Calibration mechanisms (explanations, uncertainty communication, counselling) have economic value as they affect behaviour; quantify these effects in structural choice models.

Caveats and limitations to keep in mind - Simulation results depend on explicit modelling choices (distributions, thresholds); bootstrapped CIs are conditional on those assumptions. - Study 2 measures perceived FAF and short-term psychosocial outcomes; longer-run behavioural and labour-market impacts remain to be measured. - Context: India—large youth cohort and constrained counselling capacity—so external validity needs testing in other institutional settings. - Author conflict: the author is a practising counsellor with institutional affiliations disclosed.