With that context, I can't see a way in which ERP therapy with an LLM (at least, the consumer system prompts we are used to and often think of) would not backfire and be actively harmful.
There are two reasons, and both relate to the people-pleasing tuning that ChatGPT etc have:
1. ERP is actively uncomfortable, that is the point. But there is a delicate balance to the level of discomfort.
Too comfortable and you make no progress. Too uncomfortable and you can set yourself back enormously. We aren't talking one step forward and two back. I got overly confident two months in and put myself in a worse state than I had been on entering therapy (though not my worst state overall over the 20 years). I haven't seen many people in support groups do very successful self-administration of ERP for this reason. My therapist had to do a lot of non-verbal reading of my reaction to gauge if I was at an appropriate level of discomfort, as it was not something I was able to verbalise myself without a lot of practice and learning - and that learning and practice came from guidance by a human who didn't need me to be able to articulate it already.
2. Reassurance seeking is a compulsion, and it is one of the most common and difficult to stamp out.
I can't see an LLM not providing reassurance when asked for it, and so I can't see anyone using an LLM for therapy as making any progress at all without a level of discipline and awareness they would have had to obtain from in person ERP anyway.