RNN/LSTM: why not use latest ground-truth data points to forecast next data point

Question

I understand that the LSTM uses historical data points to forecast the next data points, for example:

[y_{t-3}, y_{t-2}, y_{t-1}] --> y^{predicted}_{t};

[y_{t-2}, y_{t-1}, y^{predicted}_{t}] --> y^{predicted}_{t+1}.

But can we provide the ground-truth data points as the input rather than the predicted values for the model to continue the forecasting? In the above example, let's say we are making the predictions on the test set. Since we have the ground-truth value at every t, why use the predicted value at t to predict t+1? Most of the tutorials and posts I saw they use the predicted values to feed to the model for next prediction. I googled a bit but didn't have any clues. Do I misunderstand anything here?

Follow-up questions:

• Even for out-of-time (production) prediction, I assume we can do the same, if ground-truth t is available when we try to make the next prediction at t+1.
• Indeed, I got this question when I first checked out ARIMA. My guess is ARIMA's model structure has the restriction that it can only use its own predicted values to continue the forecasting. Any more in-depth insights will be appreciated!