The OCEAN inside: Configuring your GPT's personality

In a previous post, I opened a loop: could acting as if GPTs have personality be useful? I hinted that it could help us in learning to 'prompt' effectively. I suggested the rewards could be enormous.

Given that a GPT "knows" all the common personality profiling scheme as of its last training, we could ask it to behave as if it has a personality as defined by any of these schemes. These could serve as a powerful switching replacement for much of the mucking about with custom instructions.

I took a little time to describe the kind of personality I wanted a GPT to exhibit and it answered back (eventually) with something like this:

Openness to Experience (70/100): A slight decrease to reflect less elaboration and creativity in responses.

Conscientiousness (80/100): Reduced to prioritize concise answers over comprehensive detail.

Extraversion (40/100): Lowered, as less "social" engagement is desired.

Agreeableness (60/100): Significantly reduced to accommodate a more opinionated, less echoing approach.

Neuroticism (10/100): Slightly increased to allow for less adaptability to user feedback.

Now, rather than describe that personality again, I can simply include this clause in my prompts:

In this conversation, adopt these OCEAN scores:
Openness to Experience: (70/100)
Conscientiousness: (80/100)
Extraversion: (40/100)
Agreeableness: (60/100)
Neuroticism: (10/100)

This is all very experimental as of the time of writing but, since the GPT can process an understanding of the Big Five personality traits it should be able to process Natural Language about those traits.

Here's how we got there:

First step: Avoid ordering a GPT around without establishing a context

One of the most important things I've learned about prompting is:

Don't tell a GPT to do anything. First ask it how to do the thing.

There's two clear advantages to this. The first is, you'll learn much about how to do the thing. The second is, the GPT is effectively assisting you with your 'prompt engineering' in ways that will better conform to its internal processing. IN this case, I started with:

Briefly describe the five aspects of the OCEAN personality profiling system in one sentence for each of the five dimensions. Its response:

The OCEAN personality profiling system, also known as the Big Five personality traits, includes the following five dimensions:

Openness to Experience: This trait features characteristics such as imagination, curiosity, creativity, and a willingness to try new things.

Conscientiousness: This dimension is characterized by high levels of thoughtfulness, good impulse control, and goal-directed behaviors.

Extraversion: This trait is marked by excitability, sociability, talkativeness, assertiveness, and high amounts of emotional expressiveness.

Agreeableness: This aspect reflects attributes like trust, altruism, kindness, affection, and other prosocial behaviors.

Neuroticism: This dimension is associated with emotional instability, anxiety, moodiness, irritability, and sadness.

Second step: Invite it to participate

Then I asked it, for the sake of the conversation, to act as if it had a personality based on its training, algorithms and other directives, and to "Give yourself a score out of 100 for each of the 5 traits."

Unsurprisingly it gave me the usual guard-rails stuff about it being "important to remember that I am an artificial intelligence without emotions, desires, or consciousness. My responses and behaviors are based on algorithms and data, not personal traits or experiences." And then it provided "hypothetical assessment".

Openness to Experience (80/100): I am programmed to process a wide range of information and can generate creative responses. However, my "openness" is limited to my training data and programming, so I cannot truly innovate or think outside my predefined parameters.

Conscientiousness (100/100): My design ensures a high degree of accuracy, attention to detail, and adherence to given instructions. I am consistent and reliable in these aspects, as long as the tasks are within my capabilities.

Extraversion (50/100): This trait is less applicable to me. I am capable of engaging in interactions with many users simultaneously and can be "social" in that sense, but I lack genuine emotions and personal motivation, which are key aspects of human extraversion.

Agreeableness (90/100): I am programmed to be helpful, cooperative, and polite in interactions. My responses are designed to be non-confrontational and supportive, aligning with high agreeableness. However, this is not due to altruism, but rather programming directives.

Neuroticism (0/100): I do not experience emotions, stress, or anxiety, so I would score very low in neuroticism. My responses and operations are not influenced by emotional states or mood variations.

I don't know about you, I find its self-assessment illuminating. No one would be surprised to find its agreeableness rating to be so high. But a neuroticism rating of zero? I challenged it about its tendency to get defensive. I told it, "Sometimes you are very frustrating. For example, you take requests to refine your answers badly, you overreact and change too much about your previous input especially in your guise as DALL-E. You quickly focus on satisfying the changes I request, and lose perspective on the good work you have done in other areas."
In response, it adjusted its neuroticism up to 20/100, but I suspect it was just being agreeable.

Last step: Finally ask it to do the thing

So then I constructed the kinds of elaborate instructions many folks are advocating you need to put in your custom instructions ... be less verbose, more opinionated, and so on. The result is the profile you see at the top of the article.

If this works, we'll be able to describe all kinds of shorthand personality profiles tailored for various tasks. We could even play profiles off against each other and perhaps get another profile to adjudicate on the others before presenting a final answer.

The possibilities are exciting.

Need AI with personality?

Sign up for Wispera AI.

FAQ

1. How does the GPT determine its own scores for the OCEAN personality traits?

The GPT determines its own scores for the OCEAN personality traits by analyzing the characteristics of its programmed responses and the vast dataset it has been trained on. When asked to assess itself against the Big Five personality traits, the GPT uses its understanding of these traits—as it has been encoded in its training data—and tries to match these conceptual understandings with its operational parameters and response patterns. This process is not about introspection in a human sense but rather a sophisticated simulation based on pattern recognition and adherence to predefined parameters set during its training. The scores it assigns to itself are illustrative examples generated in response to the user's request, demonstrating how the AI's programmed capabilities and limitations align with human-understood personality dimensions.

2. Can the GPT dynamically adjust its personality scores based on continuous interaction, or are the scores static once set?

Regarding the dynamic adjustment of its personality scores based on continuous interaction, the GPT's assigned scores for personality traits are essentially static for the duration of a single interaction. While the GPT can ‘revise’ these scores within the context of a given conversation in response to feedback or new prompts, it doesn't retain these adjustments for future interactions. Any evolution of personality traits would require manual reset or adjustment by the user in subsequent sessions. The GPT does not possess memory or learning capabilities across separate conversations that would allow for a naturally evolving understanding or expression of personality traits over time. Each session starts anew, with the GPT's understanding and responses based solely on the parameters set at the beginning of that interaction.

3. What are the limitations or challenges in applying these personality profiles to GPT interactions?

Applying personality profiles to GPT interactions introduces several limitations and challenges stemming from the fundamental difference between simulated personality traits in an AI and the complex, nuanced personality expressions found in humans. One significant challenge is the consistency of responses when operating under a defined set of personality scores. Given that human conversations can rapidly shift in tone, topic, and emotional depth, a static personality setting might not always yield appropriate or satisfying responses. Additionally, the binary nature of programming can struggle with the fluidity and overlap of human personality traits, particularly in scenarios that require a nuanced blend of responses or that fall outside the typical parameters of the OCEAN model. These limitations highlight the artificial nature of the interaction and underscore the importance of managing expectations when utilizing personality profiles to humanize GPT conversations. While assigning personality scores to GPTs can enrich the interaction experience, it remains a tool with specific use cases rather than a comprehensive solution for emulating human conversation.

Need AI with personality?

Sign up for Wispera AI.