Both things can be true, but for what you're describing — CPU jumps just from being logged into the UI, console closed — it's almost certainly just the web interface polling overhead.
When you're signed into Proxmox, the backend keeps hitting every VM's QEMU process via QMP to pull status, metrics, config info, all that. All that CPU time from the qemu process gets lumped into the VM's usage graph. So even a totally idle guest will show higher numbers as long as the UI is open somewhere. It's real overhead from monitoring, not actual load inside the VM, and it drops right back when you log out.
That said, there's a second reason the numbers never line up perfectly: the Proxmox figure also includes time the vCPU threads spend waiting on the host scheduler. If the node itself is busy, an idle guest can look busier than it is. And with Windows guests? Forget about it — Task Manager counts completely differently, they'll never match.
Easy way to confirm: log all the way out of the UI, close the tabs, then watch the qemu-system process for that VM in top on the host. If it drops immediately, it's just UI polling — nothing broken. If it stays high, then dig into the vCPU threads and check inside the guest.