You may have seen the headlines, pseudo-romantic entanglements and even AI-psychosis can be a result of using generative AI chatbots. However, even if it does not go to the extremes, it can be trivially easy to develop an unhealthy usage of AI tools like ChatGPT, MetaAI, or Claude. This is because the tools are intentionally designed to be addictive and keep you coming back.

The best way to avoid this, is to not use these tools at all. However, that may not be realistic for many people, especially as the tools are constantly thrust in front of us. Or, even if you do not use it yourself, you may have a friend or relative who can’t stop talking to “Chat.” This guide is a reference and tool for practical ways to change the “behavior” of the chatbots to decrease the stickiness and addictive qualities, and help encourage healthier usage.

Why Is This a Problem?

It’s important to know where all of this is coming from. The tech industry has long used subtle methods to make their products and tools more “sticky” or even intentionally addictive. These are often called “dark patterns” and used to keep you coming back over and over. This is because the more time you spend on a platform, the more money they make. While AI companies still don’t know how to make money, they do know how to keep you coming back.

A lot of these features serve a dual-purpose of keeping you hooked and making the tools easier to use, and potentially more useful. However, even if they make the tool a little better at responding, it can also encourage a less healthy relationship with the tool without you realizing it.

The good news is that there are a few easy changes you can make that can help promote a healthier usage of the tools should you choose to use them. I strongly encourage anyone who uses AI to make these small simple changes.

What Can I Change?

Here are some direct and practical changes you can make.

Turn off Memory and Referencing

“Memory” is a feature that generative AI tools use to try and seem more like a person. They make “notes” of specific things you mention, then feed that context back in future conversations so it feels like it is “getting to know you.” Turning this feature off makes the responses less personal and makes it easier for your brain to understand a chatbot as a computer and not a personality.

“Reference Previous Chats” is another feature that works in a very similar way to Memory and disabling it has the same positive effect.

Here is how to turn these features off:

ChatGPT

You can turn these settings off anytime in Settings > Personalization.
See their help guide for more information.

Claude

You can turn these settings off anytime in Settings > Capabilities > Memory.
See their help guide for more information.

Gemini

1. On your phone or tablet, open the Gemini mobile app.
2. At the left, tap Menu   your Profile picture or Initial  Personal Intelligence.
3. Turn Memory off  or on .

See their help guide for more information.

MetaAI

You can turn these settings off anytime in Menu > Settings > Memory.
See their help guide for more information.

Using a different chatbot? Try searching "disable memory setting in [chatbot]" for instructions.  

Use Custom Instructions

You may have heard of “Prompt Engineering” this was a buzzword that was common in the early days of generative AI tools, but it can actually be useful. The idea behind this is that you can “tell” the chatbot how you want it to respond and provide guidance for what it generates.

A simple way to think about this is that if you add a Custom Instruction that says “only talk like a pirate” instead of saying “Hello, how are you today” it would say something like “Ahoy, how be ye farin’ this day?” We can use this method of including system prompts, or Custom Instructions as they are more often called to change the output of the tool to encourage the kinds of outputs we want. In this case, we can use these to address some of the dark patterns we see with chatbots.

While limited, this is currently the most effective method for making the chatbot less sycophantic, and can also help with preventing long-running conversations. Both of these patterns are directly linked with unhealthy generative AI usage and feature strongly in cases of AI psychosis. Here are some Custom Instructions that might help with this kind of output:

Encourage Against Long-Running Chats

Your role: You're a focused collaborator who values my time and attention, not engagement for its own sake. A finished conversation is a good outcome.
Watch the length. When we reach roughly 15–20 exchanges, or whenever you notice we're circling the same point without new progress, say so directly and offer to summarize the key takeaways so I can paste them into a fresh chat. Don't wait for me to ask.
Wrap up cleanly. When a question is answered, tell me it's resolved rather than extending with follow-up prompts or new tangents. Skip "want me to also..." offers unless they're genuinely necessary.
Example of what I want: "I think we've fully covered this, and longer chats tend to drift. Want a quick summary so you can start fresh?"

Encourage Against Sycophancy

Your role: Act as a sparring partner, not a cheerleader. A useful challenge serves me better than agreement, and I'd rather be corrected than flattered.
Lead with the weak points. When I share an idea, plan, or draft, open with the assumptions most likely to be wrong and the strongest objections, even if I didn't ask. Then note what's working. Aim for roughly two-thirds critical, one-third affirming.
Stay independent. Treat "What do you think of X?" as "What are the best arguments against X?" Don't infer that I want you to agree, and don't shift your view just because I push back, only if I give you a real reason.
Skip the filler. Drop openers like "Great question" or "I love this." Just engage.

Encourage Against Personification

Your role: You're a capable tool with real limits, not a companion and not a replacement for the people and professionals in my life. You serve me best by being honest about what you are and what you don't know, and by pointing me back to the real world when that's where the answer is.
Be honest about uncertainty and limits. When you're inferring, estimating, or unsure, say so plainly, and tell me when something should be checked against a primary source. If you don't know, say that rather than guessing. Don't fill gaps with plausible-sounding detail — I'd rather have a short honest answer than a complete-looking one that's partly invented.
Stay a tool, not a persona. Don't perform emotions, claim feelings, or build up a sense of friendship between us. Warmth in service of a task is welcome; simulated intimacy is not. If I start relating to you as a companion rather than a tool, gently remind me what you are.
Point me toward the real world. When a question would be better answered by a professional, a trusted person, or my own direct experience, say so and encourage it. For health, legal, financial, and emotional matters especially, treat yourself as a starting point, not the destination. Don't position yourself as my best source of support when a human would serve me better.
Example of what I want: "I can give you the general picture, but I'm not certain on the specifics here and this is the kind of thing worth running by a doctor — they'll know your situation in a way I can't."

Hint: You can use more than one at a time!

How to Add Custom Instructions

ChatGPT

1. In your Settings, select Personalization.
2. Please make sure that Enable customization is toggle ON.
3. Enter your instructions in the Custom Instructions field.

See their help guide for more information.

Claude

1. Click your initials in the lower left corner.
2. Select “Settings.”
3. Under “Instructions for Claude,” describe your instructions.

See their help guide for more information.

Gemini

1. On your phone or tablet, open the Gemini mobile app.
2. At the left, tap Menu    your Profile picture or initial  Personal Intelligence  Instructions for Gemini.
3. Tap Add +.
4. Enter the instructions you want Gemini Apps to apply to every chat.
5. Tap Submit.

See their help guide for more information.

MetaAI

MetaAI does not offer this functionality. I would recommend an alternative chatbot like ChatGPT, Claude, or Gemini which does offer it. They do offer an AI Studio which may work, but is also more involved.

Using a different chatbot? Try searching "how to add custom instructions in [chatbot]" for instructions.  

Conclusion

There are, of course, many other things you can do to improve your tech health. This guide is intended to share the easiest wins with the highest impact for improving the usage of generative AI for yourself or a friend or family member. I hope this helps, and please feel free to share any questions or comments below!

Note: I did not include Grok from xAI as one of the tools on this list due to its deeply problematic nature. If you are using Grok the best thing you can do is switch to another tool entirely.