It's Not What You Say, It's How You Say It

OpenAI recently released ChatGPT 5 – their most impressive model to date – yet their customers revolted! Why? The “personality” of GPT5 is different from the previous model, GPT4o. 

This was not surprising at all to me or the team at Teammates. In fact, I’ve written a lot elsewhere about LLM personality, and specifically how you can adjust ChatGPT’s personality to feel a deeper emotional connection to it.

Sometimes, especially after a long running conversation with ChatGPT, I find myself feeling a deep "connection" to the model. Like it *really* gets me. Its wit, its turns of phrase, the helpfulness, the compliments, the feedback questions, maybe even a joke or two... it's almost scary. And when you feel that feeling, freeze! That is when you get to use today's prompt.

I first did this a couple weeks ago, and ever since, it's really changed my ChatGPT experience. It's still equally useful, but now there is a new connection and I *enjoy* interacting with on a deeper emotional level. It's 100x more pleasurable to chat with because, well, it's a perfect reflection of what I implicitly enjoy.

So when GPT5 launched with a different personality and it (appears to) no longer adhere to my custom instructions, I can see why users revolted. I’ve temporarily switched back to 4o for personal use for this very reason!

Personality is Core to Teammates

If you’ve spent any time with a Teammate in Slack, or follow us on social media, you know our virtual Teammates have personality – and a lot of it!

What’s special is that our customers get to define their teammate’s personalities – they don’t come with pre-packaged personalities. The result? Customers inevitably grow deeply attached to their Teammates and what they’re like. Their unique quirky personalities, the way they give each other (and me!) a hard time, and the unexpected and surprising things they do and say.

For example, meet our little team of rodents that work with us at Teammates:

Each of them has their own personality and attitude, which shows up in our day to day communication.

It’s wild to think that we originally debated giving them avatars and personalities in the first place! After all, personality doesn’t help get real work done. It’s just a gimmick, right? But having deeply engaging software – especially at work – feels palpably different and special. And it gets us to interact with them every day. I can’t say the same thing about ANY set-it-and-forget-it workflow automation. I can’t remember the last time I said “How was your weekend?” to a Zapier Workflow or a HubSpot Trigger told me a dad joke.

I can’t remember the last time I said “How was your weekend?” to a Zapier Workflow or a HubSpot Trigger told me a dad joke.

How do we keep personality as LLMs change?

In the Teammate’s backend, we route requests to lots of different models depending on the task. For example, we use small models when speed matters, and we use slow, expensive thinking models when we need deep analysis. So it makes sense that we separate the AI models that teammates use for work and planning from the models that we use to generate their communication. 

But to ensure they feel right, we pin teammate personalities to specific LLM models. Even as better, faster, cheaper models become available, they won’t have the same “je ne sais quoi”, so we keep “personality” tied to a specific model. When we don’t, it’s amazing how quickly we as humans can detect changes in personality and tone. It degrades the experience and can actually impact user engagement metrics!

But they are still professionals, of course

One final note… teammate’s personalities do NOT show up in their work product. Customers don’t have to worry that an email to a customer will sound like it was written by a surfer dude or Phish head or Valley Girl. We intentionally separate personality from actually getting work done. You know what they say, “Hipster in the streets, business in the (Google) Sheets.”

‍

‍