This is a great New York Times piece on polling error and bias.

As you can see polling error can go both ways. Even in the same election, polls have underestimated both candidates depending on the state.

And as the piece notes, the polling error in 2016 and 2020 were not the same phenomenon. In 2016, polling was not weighting on education properly amidst a major realignment on education. In 2020, response bias may have been affected by the pandemic.

And pollsters have changed their methodologies. We won't know till the election if they over or under corrected. Or if they are accurate like they were in 2022. As Nate Cohn has noted, most pollsters are weighting on 2020 recall vote this cycle, which historically has biased polls towards the loser of that election- Trump.