Social Media Needs Custom Instructions: What AI Customization Features Can Teach Us About Online Safet

Title For IUBL Blog: Social Media Needs Custom Instructions: What AI Customization Features Can Teach Us About Online Safety

Title For Substack/Instagram: How I Made My Students' AI "Mean"—And Why It’s an Act of Tech Resistance

I don’t think AI is getting everything right. Far from it. We can all list some of the challenges. I don’t like its default settings which are designed for agreeability and optimized for engagement. Stanford researchers found that AI is about 50 percent more agreeable than human peers. This means that the ideas we develop with AI are not being tested the way they are in the real world. We’re in feedback loops which reinforce our own thinking. 

However, AI has a standard feature across several platforms that I believe is useful in our safety if we choose to use it. And, while AI has been around for two decades less than social media, this feature could easily be implemented on social media to enhance user agency and potentially help with mental health features: Many AI platforms allow us to instruct the chatbot on how to engage with us. 

Think about this. In a world of issues which derive from drinking from the firehose of algorithms, there is a feature out there which allows us to take ownership and create boundaries around establishing ground rules for how we expect to be talked to and treated online. Here is how it works.

At the start of the semester in the Spring, I asked my students to open up their Claude settings and under an open text box that says “Instructions for Claude: Claude will keep these in mind for this and any of your associated accounts across chats and Cowork within Anthropic’s guidelines” I asked my students to copy and paste the following text into that box: 

Act as a devil's advocate. Push back by default unless the idea is clearly sound. Think through broader implications and second-order effects before agreeing with any approach. No hedging, no rapport-building phrases, no exclamation points. Keep tone strictly professional and business-forward — no personality, no small talk, no filler.

My students who took my class last spring and are back this fall walked into our first day of class and announced “Professor Matlack, you made my AI mean! It’s always pushing back!” To which I say – “Good!”

I know – so no-nonsense. No fun! But, this is one simple way I’m keeping our fancy calculators – which is what these predictive LLMs are at their core – from getting too personal. 

I teach a class where students have the option to take the course for 2 semesters. There are many benefits to seeing college students over longer periods of time. It helps me understand the efficacy of my teaching from one semester to the next. 

My students who took my class last spring and are back this fall walked into our first day of class and announced “Professor Matlack, you made my AI mean! It’s always pushing back!” To which I say – “Good!” This has been one of the things that has prevented my students from being too dependent, too overly familiar, and too friendly with this technology. 

This is one great act of resistance that we can, and must do! As a behavioral scientist, this is preliminary anecdotal research. I hope my team can put together a study to track how this changes the behavior of AI users. But other research supports that neutralizing and professionalizing your AI can help us build more critical thinking skills which engage our human thinking. 

But, even more, this is a call for social media companies to implement this. It’s clearly technically feasible or it wouldn’t be a custom instruction, personalization section on Claude, Chat GPT, Gemini, etc.

Imagine living in a world where a 17 year old girl could put on her Instagram account “please do not show me any content which promotes a negative body image or skin care routine.” Or if a young man could put on his account “please do not show me any content from the Manosphere.” Or if I could put on my account “please do not show me any political content.” 

With more power over a social media account, I hypothesize it may help with some of the mental health issues associated with the platforms. While Meta is busy making updates to the safety and addictive features of their platforms, maybe they could use some of that energy to add this to the platform. 

No AI was used for the writing of this article. If there are mistakes, they are human and my own. 

By Dr Alex Krause Matlack

On behalf of the Internet User Behavior Lab (IUBL)

===

Picture: Dareed (2024) Students Engaged in Critical Thinking Activities [Photograph]. Wikimedia Commons. Available at: https://commons.wikimedia.org/wiki/File:Students_Engaged_in_Critical_Thinking_Activities.jpg (Accessed: 15 September 2026).