Is it ever wrong to use heteroskedastic robust standard errors? What happens if the data is in fact homskedastic?

Thanks in advance,
Richard