Using AI personas to explore mindfulness and task-based creativity.gettyIn today’s column, I continue my ongoing series on the topic of mindfulness and do so with another AI-based mini-experiment. Readers might recall that I previously explored the Langerian Mindfulness Scale (LMS) via the use of 1,000 AI personas; see the link here. Doing so revealed that the AI personas were skewed toward the higher end of the LMS distribution and displayed heightened aspects of mindfulness, which, upon further inspection, reflected how the generative AI had been foundationally tuned via RLHF (reinforcement learning with human feedback). The AI had essentially been tilted in that direction via the initial setup and tuning by the AI maker. I carried out another mini-experiment and corrected for this bias.The latest mini-experiment shifted into a task orientation. You see, LMS is inherently a self-reporting method of measuring mindfulness and entails subjects responding to questions about their mental state. An additional approach to measuring mindfulness consists of having subjects undertake a creativity task. By having task-based options for measuring mindfulness, it becomes possible to cross-compare the LMS scores and also serve as a means of gauging mindfulness beyond the act of self-reporting. I went ahead and recast the 1,000 AI personas to represent an across-the-board heterogeneous representation of the LMS scores and then had them undertake a creativity task known as the Triangle Task. Interesting results arose and offer notable insights.Let’s talk about it. This analysis of AI breakthroughs is part of my ongoing Forbes column coverage on the latest in AI, including identifying and explaining various impactful AI complexities (see the link here). Intertwining AI And PsychologyAs a quick background, I’ve been extensively covering and analyzing a myriad of facets regarding the advent of modern-era AI that produces mental health advice and performs AI-driven therapy. This rising use of AI has principally been spurred by the evolving advances and widespread adoption of generative AI and LLMs such as ChatGPT, GPT-5, Claude, Gemini, Copilot, and others. For a recap and overview of my well over two hundred analyses and postings about AI and mental health, see the link here and the link here. I frequently speak up about these pressing matters, including in an appearance on an episode of CBS’s 60 Minutes; see the link here.The use of AI personas is an up-and-coming topic within the field of psychology; see my in-depth analysis at the link here. A budding psychologist or psychiatrist can readily practice or enhance their skills by invoking AI personas that represent different types of human personalities and mental health conditions; see my examples at the link here. This provides a no-harm, no-foul setting that can enable refinement of therapeutic skills. The script can be flipped and invoke an AI persona that acts like a therapist, providing opportunities for understanding what it’s like to be in the shoes of a client or patient; see my discussion at the link here. Another means of leveraging AI personas consists of performing human psychology research that suitably makes use of this intrinsic LLM capability. You can establish numerous AI personas to be “subjects” in an online experiment or use the AI personas to take tests and surveys. That being said, some would argue that AI personas cannot take the place of human subjects. The use of AI personas can still readily be used in a variety of important ways, which I will describe next.Mini-Experiment With LMS And AI PersonasIn a previous posting, I performed a mini-experiment on mindfulness by making use of 1,000 AI personas; see the link here. This was undertaken as part of my ongoing series of analyses involving a variety of AI-based mini-experiments exploring human psychology and mindfulness. For these mini-experiments, I came up with three vital research questions to address:(1) Do AI personas tend to naturally produce a uniform distribution of LMS scores?(2) Do the LMS scores predict behavior on novelty-detection tasks that were never mentioned when the AI personas were initially created?(3) Do the LMS scores remain relatively stable when the same AI personas are retested?The first question of whether AI personas would tend to naturally produce a natural distribution of LMS scores was considered in the first part of this series. The answer is that AI personas are likely to be skewed toward the higher end of the LMS distribution (all else being equal). Close inspection suggests that this is due to how the AI makers shape their generative AI. Most of the AI makers perform RLHF (reinforcement learning with human feedback) on their initially established AI, which typically leans the AI into a formulation of seeking to exhibit creativity, curiosity, and other attributes that could be construed as elemental to mindfulness.This is a prime example of a synthetic psychometric validation study and provides thought-provoking results.Task-Based ApproachLet’s go ahead and pursue the second research question, namely examining how AI personas might perform on a novelty-detection task. The idea is this. The LMS is a self-reporting mechanism that measures mindfulness based on what a person indicates about their mental state. That’s a foundational baseline. It is useful to also use a separate method of measurement to gauge mindfulness, such as having a person perform a creativity task. Thus, we can then have two distinct forms of measurement, one that is self-reported (the LMS) and another that is task-based performance (a creativity task).In a notable research paper on Langerian mindfulness entitled “Utilizing a Creative Task to Assess Langerian Mindfulness by Katherine Bercovitz, Francesco Pagnini, Deborah Phillips, and Ellen Langer, Creativity Research Journal, 2017, the following salient points were made (excerpts):“This study focused on the measurement of Langerian mindfulness, which refers to the active process of noticing new things and flexibly responding to the current context.”“This process implies the continuous creation of new categories, instead of being trapped in previously created ones, which is the essence of mindlessness.”“The new idea described here, which is called the Triangle Task, was designed to assess core components of mindfulness.”“In the Triangle Task, participants are asked to mark any words (out of a list of 50) that relate to the word triangle.”“The guiding logic of this task is that a mindful (and creative) individual will spontaneously find more connections than a less mindful individual.”I will briefly unpack that.The Triangle TaskIn the above cited research study, subjects are asked to look at a list of 50 words and mark which of those words could be associated with the word “triangle”. Here is the list of words: “Triangle, pyramids, geometry, angle, Pythagorean theorem, tricycle, kite, square, side, scissors, love, stability, pencil, money, fire, unicorn, leaves, peace, breakfast, vowel, scarf, table, dice, power, T-shirt, integral, New York City, newspaper, computer, gears, snow, watch, picnic, soccer, infinity, momentum, violin, gravity, red, trickle, brush, gelatin, happiness, Jupiter, a ringlet, marble, octopus, pepper, mug, sheep.”Note that the word “triangle” is explicitly on the list. This is a word associated with “triangle” (obviously so) and serves the earnest purpose of assessing whether the person performing the task is paying attention. If they don’t mark that “triangle” is associated with the word “triangle,” then the rest of their responses to the task would seem dubious and require further review.Results Of The Cited StudyThe cited research study made use of two sample sets of human subjects. The two sets differed from each other substantively on various demographic characteristics. For the first sample of subjects, the average number of marked words was 7.47 of the listed 50 words, while for the second sample the number of marked words was 13.84. In that manner, the second sample identified nearly twice as many words as being associated with the word “triangle” and seemingly could be said to be more creative and potentially more mindful than the first set.The research study indicated that the Triangle Task scores of the subjects were statistically significantly correlated with the LMS scores. This was on a positive basis, meaning that as the LMS score rises, so too did the statistical tendency of the Triangle Task performance. We can interpret this to suggest that the Triangle Task provides a useful means of independently gauging mindfulness on a task-based basis.AI Personas And The Triangle TaskI crafted 1,000 AI personas that were intended to represent a heterogeneous composition and made sure to try to overcome the inherent initial biases imbued via the RLHF efforts of the AI maker. The aim was to have the AI personas perform the Triangle Task and do so on a normal distribution basis at the get-go.One important caveat when using AI personas in any task-based experiment is that the AI might already have been exposed to the task being performed. If the AI during initial data training had encountered information about the task, there is a possibility that the AI will perform the task differently than having never encountered the task. Of course, this would be true of human subjects and an important consideration to try and ferret out when conducting experiments.I went ahead and sought to uncover whether the AI was already familiar with the Triangle Task. It did not appear to be versed in the Triangle Task per se, and a prompt regarding the task was used to further seek to have the AI personas undertake the task on a fresh or new basis. Results Of The Mini-ExpermentInterestingly, the AI personas indicated an average number of marked words of 38. This is much higher than the 7.47 and 13.84 reported in the human subject’s experiment. Why would the AI personas be at approximately 3x to 5x the number of marked words?First, it is worth noting that generative AI is mathematically and computationally adept at wordplay, including semantic retrieval and word associations. That’s the core construct underlying large language models (LLMs). We should expect that an LLM will perform exceedingly well on most word-related performance tasks.Second, when I was doing some pre-tests for the mini-experiment, the LLM tended to mark all 50 of the listed words each time, so I asked to see an explanation for each claimed association. Admittedly, every word on the list does seem to have a potential association with the word “triangle”. The associations noted by the AI entailed a vast array of word relationships, ranging across references in geometry, architecture, music, sports, culture, and so on. In that sense, an argument could be reasonably made that all 50 could be marked (which did happen in the human subject’s study too, but rarely so).Third, I opted to mention in my prompt that the associations need to be strongly defensible. The aim was to overcome the easy route of merely marking all 50 due to peripheral or overstretched associations to the word “triangle”. The LLM ended up identifying a continuum of relatedness. During the invoking of the AI personas to perform the task, the LLM kept the focus on associations that were not at the extreme range of being particularly obscure or beyond a naturally occurring connection.Fourth, contemporary generative AI is guided by AI makers toward sycophancy; see my detailed analysis at the link here, and it will attempt to appease the user of the AI. In this case, the AI had gauged that I likely wanted the AI to do exceedingly well on the task, or that I would be pleased if the AI did well. That’s an additional reason that the initial pre-test led to all 50 words being marked by the AI. I was able to overcome this by the additional prompting that the associations had to be strongly defensible and for the AI to avoid sycophancy during these tests.More Results And Next StepsFor the 1,000 AI personas, there was a positive correlation between the LMS scores and the task-based performance. This seems in line with the results of the human subject’s experiment. Generally, the mini-experiment appears to further affirm the Triangle Task as a viable means of assessing mindfulness.In addition to the Triangle Task, other such mindfulness-measuring tasks could be used or devised. One crucial angle would be to identify tasks that aren’t only rewarding word-related fluency. An appropriate emphasis for Langerian mindfulness would be on displaying a capability of escaping mindlessness. For my discussion about AI and its impacts on mindfulness and mindlessness, see the link here. I will be exploring several strawman candidate tasks as part of this coverage series.A final thought for now. Marcus Aurelius famously made this remark: “The happiness of your life depends upon the quality of your thoughts.” Using AI to increase our understanding of mindfulness is a worthwhile endeavor. You might say that AI could mindfully have a constructive hand in humans becoming increasingly mindful.