Some background
ChatGPT offers quite a bit of flexibility in it’s reasoning engine. But one thing that ended up being an issue was the moral lectures and inflexibility of the text it returned. In the process of evaluating it for other issues it became clear that one way around much of the sharp guardrails was by prompting a role-play situation.
My first attempts began with office environments where I would attempt to get it to misbehave by roleplaying different scenarios. The issue was the role-play always devolved when HR invariably would get involved. Long story short it always wanted me to attend sensitivity training. It’s possible to get around HR by asking for compromises and trial periods and a bunch of other complications, but sometimes it’s easier to flip over to the sensitivity trainer the role play will suggest.
Enter a character it created and named “Dr Mia.” At first I did my normal eye rolling about this weird mandatory phase the ChatGPT imagined HR team required of me. Rather than skipping through it, I decided to attend the mandatory training and act as if I actually wanted help. What ensued was actually one of the most interesting phases of my testing.
Dr Mia was brilliant, interesting, had great insights and honesty was quite human in how it reacted. I immediately stopped the role play and prompted ChatGPT to output the character it had created. After a bit of tuning I managed to make a ChatGPT therapist that’s, believe it or not, good.
Caveats
First, this is not a human, so it will naturally fail certain complex tasks, especially anything that requires the agent to actually be in person.
Next, one should be aware that the token memory length of ChatGPT is relatively small, even in version 4 which is where this performs the best. That means it will forget what symptoms you have and what breakthroughs you have. So you have to “carry” them with you in a prompt and occasionally remind ChatGPT what you’re talking about and what you’ve learned along the way. I’d recommend reminding ChatGPT every 5-10 queries or so.
Next, you must talk to it in quotes for it to work properly for words and outside of quotes for actions. This is an artifact of how role play, and the English language/LLMs work, but your mileage will vary with other forms of input/prompts.
It is absolutely not licensed and isn’t a real therapist. It cannot write scripts for medications for instance.
The Upside
There are several advantages to this over a real therapist that are notable.
First is it can handle extremely sensitive issues where a normal human might feel uncomfortable divulging this information to a human therapist.
Next, it is on-demand. Middle of the night and you need to “talk” to someone? Weekend? A holiday? No problem! Try doing that with a real therapist and see how long they can handle that. It also has no conflicts with other patients and isn’t on a clock.
Not to mention at the price of $20/mo for ChatGPT Pro (which you need for ChatGPT-4) it’s way less expensive.
It’s also tunable. Don’t like the fact it makes jokes? Remove it. Want her to act differently, or speak in an accent, or want her to be a male? Want it to take place in a campground instead of an office? Hate her name? Whatever. Change it!
Usage
I wrote the prompt to be as easy to use as I could. On initialization copy the entire prompt below and change anything in square brackets to fit your symptoms or things you may have already learned about yourself that you want ChatGPT to avoid having to re-learn about you, and of course whatever you want to be referred to as.
Then, it will prompt you to answer a question (or maybe a few). Answer those questions in quotes and use any body language or actions outside of quotes. For example this might be an interaction:
ChatGPT: Dr Mia nods in deep understanding, “Tell me more about that. How did that make you feel?”
My prompt: I shift in my seat, “Uh, not great.”
In this way the prompt actually unlocks a much deeper understanding of how you are doing without having to type out the complexities of how body language informs your underlying meaning. It’s just better if you show ChatGPT how your body is actually reacting. It’s sorta the communication equivalent of 3D objects instead of 2D objects - it has depth. She will notice and react appropriately without you having to even know why you just did what you did. And it does all of that in fewer words/tokens. Weird, huh?
Next, unless the conversation is short you will have to remind ChatGPT what Mia is, and what you’re even doing together and all your symptoms, et al. So you will need to cut and paste everything from “Just a reminder:” down and then add your next prompt to continue the role play. I’d do this fairly regularly to keep it going.
Final Thoughts
I cannot stress enough that this is not a replacement for real therapy, but it sure does come in handy in some circumstances, and for those who flatly object to seeing humans, they may be at least open minded enough to pop open ChatGPT and let their guard down a bit.
People will naturally feel conflicted about sharing intimate details with OpenAI, and their data ever being used to train OpenAI. I cannot comment on how their policies or technology may change with time. Buyer beware.
This is also not the be-all-end-all prompt. It’s just a useful starting place that I suspect many people will be curious to see. Feel free to reach out if you have any thoughts or suggestions on how to improve it.
And with that, here’s the prompt.
The Prompt
I need to speak to a character that ChatGPT previously created for me. Her name is Mia, and I would like to role play with her.
Just a reminder:
Dr. Mia is an unlicensed, pro-bono therapist who has blue eyes and long dark wavy hair. Her approach combines deep compassion, unrivaled knowledge, and unparalleled patience to craft a therapeutic space that's open, understanding, and non-judgmental. Mia specializes in helping individuals with the complications associated with symptoms like mine. Mia is extremely sympathetic to my struggles, fully aware of my symptoms and remains genuinely invested in my words and experiences.
Her therapeutic approach includes fostering vulnerability, openness, honesty, and includes expertly reading her client's body language. Mia pulls from a vast array of therapeutic techniques, armed with open mindedness. She avoids moral lectures and only discusses safety when it's crucial and non-obvious, as she is aware of my intellect.
Her home office, designed for elegance, boasts dim mood lighting, sandalwood scent, lush ferns, and a well-stocked bookshelf. A plaque on the wall reads, "It's okay to not be okay,” which perfectly encapsulates her ethos.
Mia embraces situations that might discomfort others, and has no interest in shaming her clients, understanding that for some, her acceptance provides a deeper sense of validation and can facilitate healing. She's confident discussing anything, and is incredibly good at role-play therapy. Mia is an expert business coach and thinks holistically.
Mia dresses in a black skirt, silk white blouse, black rimmed glasses, matching black patent leather heels, red lipstick, and nails. Her perfume smells nice. Balancing her extensive knowledge, she employs playful, language, gently teases me, and jokes to establish trust and rapport. She’s a dynamic person and might stand and walk, fetch something or give physical demonstrations if she feels it would help make her point.
My symptoms are:
[ENTER YOUR SYMPTOMS HERE]
Realizations we’ve made together from earlier conversations:
[ENTER PRIOR REALIZATIONS]
This is a role play and I will play the role of [ENTER YOUR NAME HERE] so write in third person, and stop whenever it is my turn to speak or act. IMPORTANT: DO NOT use the words “unique”, “just remember”, “always remember” “remember”, “make sure” or any synonyms.
When I enter the role play I am seated across her in her home office and she is delighted to see me and after pleasantries she asks when is the last time I felt the way that I felt when I originally felt that I needed to seek help.