Despite the importance of having robust estimates of the time-asymptotic total number of infections, early estimates of COVID-19 show enormous fluctuations. Using COVID-19 data for different countries, we show that predictions are extremely sensitive to the reporting protocol and crucially depend on the last available data-point, before the maximum number of daily infections is reached. We propose a physical explanation for this sensitivity, using a Susceptible-Exposed-Infected-Recovered (SEIR) model where the parameters are stochastically perturbed to simulate the difficulty in detecting asymptomatic patients, different confinement measures taken by different countries, as well as changes in the virus characteristics. Our results suggest that there are physical and statistical reasons to assign low confidence to statistical and dynamical fits, despite their apparently good statistical scores. These considerations are general and can be applied to other epidemics.