Outliner detection with LSTM autoencoder

I am learning about autoencoders for outlier detection. I have searched enough and internet suggest to use LSTM autoencoders for outlier detection from multivariant time series data. I have watched this video and other tutorials. I can not understand how can I find outliers from it?

I know autoencoder 1st deconstruct and then reconstruct the input, through this process it finds outliers. But when it decides there is an outlier. What should be the difference?

for example, in this video, the author tested the pre-trained model with [[1,2,2]] then the coded value was close to input. When/ how it will be decided that there is an outlier?

I will really appreciate your help.