This paper investigates how the psychological bias in online product reviews affects the social learning of customers. When customers write product reviews after purchase, they might give extremely negative or positive reviews because of the asymmetric perception of the product quality compared to the quality they ex ante expected. An important feature of our problem is that customers who arrive later cannot precisely estimate the very reference point on which each review depends. The lack of information does not allow simple Bayesian learning, so we devised a rational non-Bayesian belief updating mechanism. Our study first verifies the resulting bias originating from the biased method of review generation, and then proposes a simple, well-operating bias correction rule that can be implemented by customers. To analyze the limiting behavior of the quality beliefs, we utilized a fluid-approximation that makes the stochastic sequential process into a continuous deterministic model that solves a differential equation corresponding to the original recursive relation of development. Based on this analysis, we show that the proposed bias correction rule induces successful learning; that is, customers can become aware of the true quality in the long run if all of them use the bias correction rule. However, when customers are heterogeneous in their review writing process, the rule that is needed to fix the bias is more complicated than intuitive. Even this rule requires the portion of customer type in the whole population to be common knowledge. Otherwise, bias cannot be successfully corrected with any customer-side correction rule.