This paper critically examines the 'rough volatility' hypothesis, which claims that log-volatility follows a fractional Brownian motion with Hurst parameter H<0.5. The authors first demonstrate via simulations that the popular scaling method of Gatheral et al. (2014) systematically produces H≈0.1 even when the true volatility is not rough (e.g., an Ornstein-Uhlenbeck process), because the method ignores the estimation error in realized volatility proxies. To address this, they develop a novel quasi-likelihood estimator for the parameters (H, η) of a fractional stochastic volatility model. The estimator incorporates the measurement error of realized variance via a central limit theorem approximation (i.i.d. normal with variance 2/m), and uses a Whittle-type likelihood for the approximate Gaussian increments of log-volatility. The authors prove consistency under high-frequency asymptotics where both the sampling interval δ→0 and the number of intraday observations m→∞, with conditions ensuring that the measurement error vanishes sufficiently fast. Simulation studies confirm good finite-sample performance. Applying the estimator to daily realized volatility data from the Oxford-Man Realized Library for major stock indices (e.g., S&P 500) yields H estimates around 0.08–0.09, even smaller than the previously reported 0.1, thus providing stronger evidence for the rough volatility hypothesis. The paper highlights the importance of accounting for measurement error in volatility proxy data when estimating the smoothness of latent volatility.