I’m building an app focused on mental health, where users can create journals, set goals, attend classes, and more. I plan to incorporate AI assistance based on user-selected focuses, such as meditation, business, sleep, and others (up to 50 keywords). Users can select 5 to 10 focuses, and based on these, I’ll develop various AI-powered features within the app.
Could you recommend which OpenAI text model would be best suited for this application? I’d appreciate any suggestions based on your experience.
My biggest recommendation is to build it so that you can switch easily because models keep changing. What is best today will be different tomorrow, for your use case or your wallet.
You can use the playground as well to try out and compare results.
I’ve been using GPTs and other software for a similar purpose daily for over two years, including, as you mentioned: journaling, meditation, mindfulness, skills training, daily life counselling and advice.
Here’s what I feel I need to say in the strongest possible terms:
You cannot and must not use any language model for this purpose in a generalised way, built into any application used by other people, unless each user is supervised by a health professional.
The compatibility of any model with any given person is currently unclear. It doesn’t matter whether it’s GPT-4, 4o, Sonnet, Mistral, or even something designed to generate images or music - what seems like a stone cold machine to one person may feel like the warmest, most affirming companion to another.
There are no consistent guidelines yet, so it would be irresponsible to use models for mental health purposes unless you are your own test subject and patient.
Perhaps this will change in the future, but for now it would be unethical and dangerous to use these tools in a generalised way for mental health.
And to pre-empt the argument “but you’re using it yourself, so my point is moot” - yes, I am. But I also have a clinical psychologist who supports me, who I can reach daily.
I don’t want to discourage you or tell you to give up, because who knows what the future will bring, but right now it’s important to prioritise user safety and professional supervision above all else when developing something as sensitive as this.
They dishonestly tune the model for the tests.
It’s no more a reflection of actual outcome than public school testing.
They both “teach to the test”.
But, of course, the whole point of 4o is to cut corners in costs, it’s “lighter”, and they hope tuning it to pander fools users.
The idea that adults need to be “supervised” in their personal activities is authoritarian horror.
Adulthood is about personal competence and responsibility. Anyone who can’t handle being self-regulating in their own personal life is not an adult, no matter what age they are.
Actual adults must be free to choose, and held responsible for their choices.
.
