Collaborating with AI Personas
Amidst the deluge of interest, activity and hype in the field of Artificial Intelligence (#AI), at Polynize we are taking a different approach. Conventional approaches are developing AI as generalized, functional #tools that help in completing tasks. This utility-first approach will lead to many large, successful businesses. But we are focussed on developing AI #agents specifically optimized for collaborating with people / teams in ideation and problem-solving, as we believe that this is a model of a very valuable type of human-AI hybrid workplace of the future.
To do this, we are using our own game platform to study how humans and AI collaborate and compete in various open-ended contexts that require creative intelligence skills; and then iterating rapidly based on our learnings.
Already we are finding out something very interesting — that when operating in these types of arenas, people want to interact with AIs that are personified with the traits, skills and personas that characterize human minds. This has important implications for our game and to create such personified AIs requires a synthesis of strategies from software development, #gaming, #psychology etc.
In this article, I provide a (very) quick tour of the why, the what and how of wrapping AIs in personas and some musings on the next steps.
Why do we like AI personas?
The typical #userexperience of using #AI in knowledge work is pretty functional — a text-based window, typing #prompts, reading responses, a #UI that hasn’t changed substantially since I first wrote a rudimentary chatbot as a kid in 1990.
And yet, however functional and utilitarian our interactions with another entity, are relentless anthropomorphizers — we can’t help ourselves from projecting human properties onto others with whom we interact. We do it with our pets, we do with our software and now we do with our AIs. I learned this very early on when my BASIC “psychiatrist” made my friend cry.
This tendency to anthropomorphize is very strong and it comes from biological bases. We are optimized for natural interactions with others of our own species and we have a highly evolved theory of mind. To project “mind” onto another agent reduces friction and greatly simplifies our interactions, allowing us to devote more cognitive resources to higher order thinking.
Scientists and technologists might say we should interact with our machines for what they are, not what we want them to be. But what if we flipped it around? What if we made AIs that feel familiar to us?
The Power of the NPC
As usual, the gaming industry was a decade or two ahead of everyone else. Even the early story-based games began to use “non-playing characters” (#NPC) — simple story characters that interacted with the player-protagonist. Although they were simple and predictable they had undeniable emotive power and they made games entertaining and playable.
We are wired for faces, expressions, body language, tones of voice etc. The emotive power of the NPC persists, now augmented by the visual, aural and cognitive abilities of modern technology. And it still brings a smile to the face. Consistent in our player feedback is that people love that moment when they “meet” their first Polynize “NPC”.
(Slightly random, but check out this post for an interesting story of how AI #LLMs can already impersonate your friends!)
Think like a Game — Experience First
But to develop such personified agents needs a different way of thinking altogether. Conventional approaches are utility-based — they a) concentrate on functionality as the primary vector for development, b) exclude the human mind and its states from systems, and c) focus on outcomes over experience.
(This frame of reference is natural for technologist-led innovation, the primary lineage in the software development industry and technology startups.)
In contrast, computer games are designed in an experience-based frame — they a) include interfaces, animations and design as primary development vectors b) include the human mind and its emotive states in the system and c) focus on experience over outcomes.
We are in a good position to take an experience-first design approach, because we can bootstrap from the utility-first work that is being done by e.g. OpenAI and made available through APIs. We can “wrap” that utility in the “experience” that we need to create personable, relatable AIs for the future.
Creating Relatable Personas
One of the risks of innovation is to over-engineer. We don’t need to solve the Turing Test here. I have taken a holistic human-AI perspective for many years and my experience shows me that:
a) AI doesn’t need to be very intelligent to lead to good system outcomes in collaboration with humans and
b) AI agents don’t need too much sophistication in behaviour to lead to satisfying “human-like” interactions.
In terms of “intelligence”, the new generation of AI APIs like OpenAI give us more than we need. The trick is to “wrap” the inputs and outputs of these APIs in a layer of prompt / response processing that modifies the I/O behaviour of the AI using certain parameters that map to human “traits” and “skills”.
Since personas are familiar to us, we bring an enormous amount of pre-learnt contextual information to the interaction which helps us to engage in a rich, layered collaborations with the persona.
Take a quick look at the example above, interacting with a “SCIENTIST”. Certain traits and skills are pre-linked to the persona, both “negative” and “positive” in affective terms. Examples are +VERBOSE, +PRECISE, -NOVELTY, +ELABORATION. These are provided as parameters into Prompt Composition where they are used to modify the original inputs (e.g. “USE CITATIONS”.
Responses are similarly post-processed if required before returning output to the user. Prompt and response processing can also take parameters for randomization, personalization to a user, history-dependence. We could also cascade API requests to compose more unique / sophisticated outputs.
And so on. There are many. many innovations to come here but at this stage the point is to introduce the basic principles for creating these personas. We have already used this to create personas, inject them into live gameplay to compete and collaborate with human players in innovation games and measure their performance.
The Payoffs of Personas
AI personas create human-like presences that map to existing social roles within teams and workplaces (e.g. “SCIENTIST”). The user experience is fundamentally different to interacting with a generalized AI. Our thesis is that personas are a better fit for collaborative environments because of some key benefits:
a) Cognitive Fit — it is natural and frictionless to interact with agents that exhibit familiar human-like behaviours. the “expectedness” of those behaviours provides a level of trust and predictability that leads to better interactions and less friction.
b) Socialized Diversity — personas take generalized #tooling and translate it into understandable and #diverse social roles (“SCIENTIST”, “PROFESSOR”, “HACKER” etc.) which give #autonomy to users who may prefer to interact with certain “types” of user. Especially important for non-technical users.
c) Team Fit — personas can be fitted into #teams in a natural way to augment performance at a team level. We are already seeing this in our game where certain personas are recruited into teams to provide e.g. technical advice.
d) Relatability — our powers of theory of mind make personas relatable. Very importantly we can even identify certain personas being similar to us, which provides us with comprehensible “benchmarks” for how AIs in particular roles perform vis-a-vis humans — this is essential for #training human creative intelligence to compete in the post-AI workplace.
e) Storytelling — we perform better within narrative frameworks and AI personas can play character roles within those narratives. Not only important for experience but this leads to many benefits such as filling gaps in scenarios (e.g. missing people in meetings). In our games AI personas e.g. help players to play out the story of launching a start-up.
The Future of Digital Work.
Imagining the future of digital experience and knowledge work needs new ways of thinking about human-AI systems as a whole, not just technology-first approaches that exclude so much. This in turn requires high-level synthesis of best-practice knowledge and empirical work from many disparate domains such as software, #gaming, #socialsciences, #psychology, #design etc. That is very difficult to execute for most because most teams do not have multi-domain experience and knowledge and freedom to explore.
In particular, gaming and AI come together here in a new way and allow us to apply this work in new contexts like #highereducation because audiences there are already #gamers. And because we are tapping into deeper human tendencies and preferences the experience translates well into new contexts.
Ultimately our mission is to level-up human creative intelligence along with machine intelligence, a mission needed all the more in this moment to prepare the workforce of tomorrow (in a way education systems do not). This will happen both in #collaboration with AI and in #competition against AI. These AI personas are essential to that mission — our players are battling with them, teaming-up with them, learning from them, teaching them … and having a lot of fun with them along the way.
