AI in Thailand: Will Easier Work Weaken the Incentive to Learn?

In Thailand, this is becoming a practical question. ETDA's 2024 assessment with NSTDA received responses from 580 public and private organisations across ten target sectors, from 3,758 contacted. Among respondents, 17.8% reported using AI and 73.3% planned to do so. Improving production or service efficiency was among the main adoption objectives, while skills, data quality and funding were identified as barriers to generative AI use. The survey offers a dated snapshot rather than a current census, but it shows organisations pursuing efficiency while already recognising constraints around capability (ETDA, 2024).

Those ambitions create a measurement problem. Consider a hypothetical reservations team in Phuket using AI to draft replies in several languages. Faster responses could improve service and free staff to handle complicated requests. Yet the same output could conceal gaps in understanding cancellation conditions, recognising a misleading translation or deciding when a request needs escalation. The employee can look more capable without becoming more capable, while the business may still be better off. That is what makes the problem interesting.

International evidence shows why immediate performance and learning need separate assessment. Bastani and colleagues studied nearly 1,000 high-school mathematics students in Turkey. Two forms of AI assistance improved practice performance, but students using the less constrained GPT Base subsequently scored 17% lower relative to the control group when assistance was removed. A structured GPT Tutor largely avoided that disadvantage, without establishing a positive unaided-exam gain (Bastani et al., 2025).

The important finding is not simply that one AI group performed worse. The two tools produced different later outcomes despite both improving assisted practice. Guardrails helped avoid the measured learning loss associated with the less constrained tool. At the same time, the experiment changed the assistance available rather than the reward for learning, so it does not confirm the incentive hypothesis. Students may have reduced their effort, used fewer productive mental steps or misjudged how much they understood. The result establishes a risk that completed work can conceal, but it does not identify one explanation for it.

Evidence of learning gains prevents the argument from becoming one-sided. Kestin and colleagues compared a structured AI tutor with active learning in university physics. Students using the tutor achieved higher post-test scores, with a median learning time of 49 minutes against an assumed 60 minutes of classroom learning. The intervention combined prepared solutions, structured activities and personalised feedback (Kestin et al., 2025).

The comparison requires care because one time figure was measured and the other assumed, while the assessments were short term and do not establish lasting mastery. Even so, the result demonstrates that time saved can be compatible with learning more. Confusion, delayed feedback and unsuitable pacing consume time without necessarily providing useful practice. If AI removes those costs, a capability previously considered too difficult or expensive to acquire could become worth pursuing.

Together, the education studies lead to a more useful conclusion than the claim that comfort weakens development. Difficulty has value when overcoming it develops something useful. What matters is whether the design of assistance preserves the thinking, practice and feedback needed to apply knowledge beyond the immediate task.

Workplace research adds a different perspective. Brynjolfsson, Li and Raymond studied the staggered introduction of AI assistance among 5,172 customer-support agents. Their final published analysis reports a 15% average increase in issues resolved per hour, with larger gains among less-experienced and lower-performing workers. The most experienced and highest-performing workers gained less, with small declines in some quality measures. The authors also report evidence of durable learning: productivity gains persisted during outages, when AI recommendations were unavailable, and agents with two months of tenure performed as well as agents with six months of tenure once they had access to the tool (Brynjolfsson, Li and Raymond, 2025).

This study still measures productivity and inferred learning rather than directly testing reward-driven underinvestment, and its quasi-experimental interpretation depends on assumptions about the rollout. Even so, it complicates a simple reading of the incentive hypothesis. Where Bastani's unguided tool left students worse off once assistance was removed, this workplace tool appears to have built durable capability alongside the immediate output gain. AI can raise the performance of less-experienced workers towards the existing standard, making expertise harder to infer from routine output. A smaller measured gain among experts does not make their knowledge unimportant. Its value may appear in unusual cases, error detection, customer judgement or the knowledge embedded in the AI system itself.

The underlying concern predates generative AI. Bainbridge observed that automation can leave people responsible for unusual failures while reducing the routine practice that prepares them to respond (Bainbridge, 1983). That is a loss of practice rather than a reduced incentive to learn, but the two effects could reinforce each other.

My assessment is that AI creates a conditional risk of underinvestment in learning alongside a real opportunity to make learning more accessible. The evidence shows that assisted performance can conceal weaker later results, that structured support can improve learning and that less-experienced workers can obtain large immediate productivity gains. It does not show that AI is already weakening incentives to learn in Thailand.

The risk appears where an organisation rewards the output AI has made easy while continuing to depend on a capability it no longer observes, practises or rewards.

The defensible conclusion is more specific. The risk appears where an organisation rewards the output AI has made easy while continuing to depend on a capability it no longer observes, practises or rewards. Thailand's adoption ambitions make that question relevant now, while the limited local evidence gives businesses, educators and researchers good reason to examine it more closely.

The next question is who benefits when progress occurs. If AI makes expertise cheaper to acquire and services easier to produce, the gains could appear as higher margins, better pay, lower prices, improved quality or more free time. How they are distributed will depend on competition and bargaining power as well as technology. That is the next issue in examining AI's economic and social consequences for Thailand.