A new generation of AI serving human creativity
Tensions are opportunities that need creative solutions. Great entrepreneurs hone their skills at turning observations of points of tension into new business opportunities. For example, Amazon recognized early on the tension of waiting for products to arrive while simultaneously paying high fees for shipping. Now we have next-day free delivery.
The front line of an organization knows whether something is working or not. They sense a problem or an opportunity before the rest of the organization is aware or able to respond. For today’s organizations to stay competitive, they must increase excellence in rapid co-creation. They need to turn tensions in the business into creative solutions like Amazon.
However, translating observed tensions in a business or government organization into innovative solutions is a highly dynamic and complex process that requires a new approach.
Many innovation initiatives within corporations need help determining what innovations are worthy of significant funding. How do we pick the innovations bound for success? Startups are maverick organizations at the very center of innovative disruption. Yet most fail. Attempts to apply AI to find nascent winners tend to fail because data on emerging disruptive innovations are scarce and subject to rapid change.
The common theme for successful accelerators is the right collection of smart people with relevant experience and perspective — collective human intelligence is central to success in innovation.
Automating the acquisition of collective knowledge
Creativity and imagination for novel and disruptive innovations remain in the realm of human co-creativity. For that reason, we set out to build a new approach to AI that closely resembles the MIT Center for Collective Intelligence’s mission to create technology that combines human and machine intelligence in a system that collectively acts more intelligently than any person, group, or computer. A hybrid form of AI and human intelligence promises a more predictive path in a business environment under pressure to become more decentralized and adaptive.
Automating the collective knowledge acquisition of groups is key to a new generation of AI. The cost of manually acquiring knowledge from humans led to the collapse of AI 1.0 — knowledge-based expert systems. I was there. I was the co-creator of the Knowledge Engineering Environment, a software system for programming knowledge models. AI 1.0 delivered hand-crafted expert systems that transformed many areas of business, from inventory management (airline seat pricing was automated during the first wave of AI) to the configuration of complex systems (from machinery and trucks to computer systems). Handcrafting knowledge models were costly, biased, and often isolated to the expertise of one individual. We will return to the subject of knowledge acquisition shortly, but first, a bit on the second wave of AI.
The onset of the internet led to sufficient data to learn models from data with neural nets and other machine learning techniques, leading to AI 2.0 and the success of deep learning models. AI systems generated purely learning from data removed humans from the loop. The results of deep learning models continue to astound — large language models being one example. There are limits, as noted by Yoshua Bengio, one of the founders of deep learning. He acknowledges the limitations of our current models:
“In terms of how much progress we’ve made in this work over the last two decades: I don’t think we’re anywhere close today to the level of intelligence of a 2-year-old child. But maybe we have algorithms that are equivalent to lower animals for perception.”¹
Creativity and innovation require a different kind of intelligence. In his book Why, Judea Pearl describes the need for thinking that goes beyond sensory learning from patterns of data — understanding the why behind a prediction — a process of causal reasoning. He states: You are smarter than your data. Data do not understand causes and effects; humans do… The surest kind of knowledge is what you construct yourself.²
A new path forward crystalizes when we learn the important pieces that need to be integrated by learning from the symbolic era of AI. In AI 1.0, the hand-crafting process involved interviewing experts to link observable conditions to underlying causes based on the expert’s accumulated practical knowledge. Patterns learned from experience are represented as cause-and-effect rules. A physician links signs and symptoms to possible causes. Those relationships are captured in rules. Diagnosticians work with probabilities and gather more evidence through tests to narrow down the possible causes. Pearl realized that process could be captured in a Bayesian Belief Network. Beliefs (hypotheses) connect causes with observable conditions through probable relationships.
Scalable acquisition of knowledge from groups of humans solving a problem together is possible but requires bringing together three areas of science and technology. For this discussion, I will provide a high-level view of why it is relevant to scalable, collective knowledge acquisition and co-creation. The areas are collective intelligence, cognitive science, linguistics/deep language models, and complex adaptive systems,
- Recent advances in collective intelligence show that collectively, we are more accurate in predictions and decision-making as a cognitively diverse collective.³ A cognitively diverse group always outperforms an individual expert.⁴ Collective intelligence is fundamental to human knowledge development, as evidenced by the scientific method. We build lasting knowledge through collaboration that centers on peer-reviewed and supporting evidence. Human collective intelligence is believed to underlie the remarkable success of human society Scalable knowledge acquisition requires instrumenting the process of diverse groups of humans working together to co-create solutions.⁵ For collective intelligence to work, we must pay attention to the rule: cognitive diversity in the participants increases the predictive accuracy of the results.
- Cognitive science, linguistics, and representations of human knowledge provide the interface between humans and machines. For humans to understand the recommendations and actions of machine intelligence, we need explanatory models in natural language. Deep learning transformer models popularized as “large language models” provide a present and practical bridge between humans and machines. We use these technologies to capture the reasoning and results of people working together to solve a problem. From their discussions and deliberations, we create models of their collective knowledge. Deep learning transformer language models provide the means to extract knowledge from collaborative work.
- Complex adaptive systems are extremely hard to predict. An AI that learns patterns of knowledge from human interactions is grounded in the mathematics of detecting emergent patterns. The new tools require technologies that enable navigation in a complex adaptive environment. Google search is an example of applying the mathematics of complex adaptive systems to a practical problem. Early attempts to organize web searches by creating directories (Yahoo) gave way to PageRank, based on a complex adaptive system model used in discovering emergent patterns (Markov modeling). To the user, it simply means speed and accuracy when you type into the box. Harnessing the power of a complex adaptive algorithm led to a simple interface with a complex dynamic system that is massive and forever changing. Co-creation is a complex adaptive emergent process that yields answers to questions we did not think to ask. To the users, it simply means making decisions faster with less endless arguing and greater confidence.
How do we create a scalable, decentralized, automated system for acquiring knowledge? How do we harness the collective intelligence of a diverse group of experts in a rapidly changing environment to achieve a predictable outcome? An analogy that helps bring it all together is the facilitator of an interactive process that balances the generation of ideas with learning relevance and prioritization. Thinking of this new AI as a super-intelligent, adaptive learning facilitator/orchestrator of a process provides a way to bring the various technologies together — for simplicity, we will call it the Collective Reasoning Facilitator(CRF).
The CRF is guided by a solution template system that gathers each participant’s inputs and gathers information about their reasoning — a process called collective reasoning. The CRF is driven purely by the script provided by the solution template. No human intervention is required. The collective reasoning process produces a highly predictive group probability estimate (following the principles of collective intelligence) and a causal model of their reasoning represented as a probabilistic graphical model (BBN).
The CRF guides a problem-solving process, thus the article’s title. Given a framework for solving a problem, the task for the CRF is to engage with the relevant participants anywhere, anytime. Co-creation of solutions requires a focus on evidence-based reasoning — learning what knowledge we tap into when forming an opinion about a decision. The CRF wants to know what you know, not who you are, so the CRF makes certain ideas stand independently, separate from the contributor’s identity or position. The CRF is blind. It doesn’t know if the great idea came from the c-suite or the mailroom.
The CRF builds a collective knowledge model, a picture of the collective mind, if you will, of the group’s collective reasoning about a problem or decision. The CRF also learns whose knowledge the group values, which varies with the particular engagement. The system treats participating members as “authors” in various co-creating, problem-solving engagements guided by a solution template. Each engagement results in a collective mind map of ideas considered and applied to solving a problem or creating a new widget. And all in real-time. Each engagement results in a way to attribute value to the participant’s contribution to a result. By this approach, we bring order to a knowledge economy of co-creating groups of any size.
Startups constitute an excellent testing ground for guided problem-solving and co-creation because they represent the core problem innovation — taking the risk to fund a project without a track record of performance.
The solution template is typically created from an existing heuristic or rubric to guide a decision. In the case of startup investing, the template derives from best practice research in angel investing. The CRF uses it as a guide to engaging with participants on the specified decision criteria. A solution template can be as simple as a single qualitative question. For a predictive model, a set of quantitative questions with sub-questions guides collective reasoning about what thinking motivated the score. The image below illustrates how it works:
- Decide the key questions that guide the discussion and deliberation process (solution template)
- Invite as many participants as needed to tap into their knowledge
- Participate in an asynchronous, single-blind engagement
- Get real-time results
Consider the case of an angel investment community that wants to co-create curated investment opportunities. They agree on a set of four criteria for the predictive decisions model:
- Quality of the business opportunity (Is it a compelling opportunity)
- Fitness of the team to the business needs
- The networking power of initial investors and advisors
- Likely to invest, which covers conviction about the investment opportunity (deal terms etc.)
P(Outcome)= Market_Opportunity + Team + Investor/Advisor Network + Likely_To_Invest
Each feature is viewed as a probability estimate (e.g., there is an 80% chance that this is the right team for the business). The system guides the participants through each question asking reasons for the score. Each time a participant asserts their reason(s) for a score, they are presented with a dynamically generated list of what others think. They can indicate alignment with others by a prioritization process and submit new thoughts stimulated by seeing others’ thinking. The system is constantly evolving, learning, and looking for emergent alignment patterns (or not) from the participating group.
The process results in a score representing the probability of a successful investment and an explanation of the reasoning process. In addition, the system captures all of the thinking in a live, simulatable knowledge model — a probabilistic graphical network that links collective reasoning to scores.
The participating angel investors see their work result in a predictive score expressed as a probability — the group believes this startup has a 77% chance of success. The group would also see all of their work integrated into a knowledge model where they can explore the reasoning behind their score.
The model shown is a graphical representation of all the collecting reasoning of the group in creating the prediction. Reasoning influences score value through the probability of relevance of evidence. Clusters of reasons are grouped into themes. For example, the chart illustrates how the “defensibility of technology” is a concern reducing the overall estimate of success.
The system produces a high-level abstraction of the reasoning process while retaining the “voice ” (original wording) of the participant.
The probability of importance of reasons given for a score is a vs. the breadth of support. This display shows that team execution is a top priority reason for most participants.
We can also look at how categories of reasons align with quantitative scores.
Traction and management are themes associated with high scores while IP defensibility is associated with low-scoring patterns.
The system builds a model of the collective mind. It maps the results of the co-creation process into a visualizable space where it is possible to get an in-depth understanding of alignment, outlying ideas, and compelling market opportunities.
The semantic model of the collective mind is mapped into a 3D space with rich tagging information that includes the original text, automated labeling into thematic categories, and probability of relevance, creating a deep understanding of the participants’ knowledge.
The CRF learns the participants’ role in producing a result, creating a score called InfluenceRank. InfluenceRank allows participants who evaluated a successful startup to receive credit for the value of their relative contribution.
We found this approach to discovering disruptive innovation works on multiple levels. First, it was remarkably accurate (>80%) in picking follow-on funding. We learned this by applying this to a group of private/angel investors evaluating ~150 companies. Second, the process reduced bias: >40% of the founders were female, even though nothing was done to curate companies or participating investors.
Participants in the process were angel investors, technologists, and private investors. A unique, cognitively diverse team was created for each startup. The CRF system builds a model of the participant’s contribution level using a measure called InfluenceRank. Each startup evaluation project resulted in a ranking of contributors based on the single-blind peer review process.
Why is this relevant to a new process of guided problem-solving and co-creation?
All innovations face three questions similar to the startup case: 1. Will it work, 2. Is it worthy of funding and resourcing 3. Can we measure its impact? These types of questions apply to all innovative projects.
Business execution is measured by the speed and accuracy of applying capital to the right products, projects, and initiatives that will increase competitive advantage. As in the case of startup investing, there is generally insufficient data to make decisions with clear confidence. Business operates in a complex adaptive world challenged by a high degree of connectivity, complexity, and change. The tools we relied on for so many years are failing us because they were designed for a world that no longer exists. Technology that enabled global connectivity made an irreversible change to business and society. Decentralization of the workforce creates new challenges for management and corporate productivity. We are ”not in Kansas any longer.” We need new tools for organizations and business enterprises in a complex adaptive world.
In their book Design Unbound: Designing for Emergence in a White Water World, Ann Pendelton-Jullien and John Seely Brown state that organizations need to design for emergence rather than for absolute outcomes. The old concept of a strategic planning process producing a static plan must be replaced with a situation-response style strategic war room equipped with the best of an organization’s talent and technology to respond to emergent risks and opportunities. We must find ways to resonate with emerging movements and capture the opportunity with speed and accuracy.
Innovation is a key driver of transformation in today’s climate. Innovation drives the evolution (and survival) of all organizations. Co-creation underpins sustainable innovation.
It is critical to recognize the value contribution of each participant. Each participant's contribution's exact working (voice) is retained in all language analysis and reporting. Collective Voice Analysis builds a narrative of the collective mind while retaining the integrity of each contributing participant’s voice.⁶
The CRF scales to groups of any size, and the process is asynchronous, supporting the growing hybrid work environment. In addition, these characteristics make it well-suited for Web3 implementations.
We are on the cutting edge of this new era of AI-guided co-creation. So far, we have seen applications in the co-creation of strategic solutions for improving organic revenue growth, decision-making processes for global organizations, co-creation of consumer products for brands, co-creating solutions for the future of work, and co-creating strategies for business transformation.
Co-creation is critical for our global future. We must return to the fundamental value of human collective intelligence and its power to transform. Recently, we started an experiment in inclusive transformation and co-creation of ideas to reduce violence in our cities.
One simple way to get started with AI co-creation is in your next Zoom video conference call. For groups larger than 10, it becomes increasingly difficult to give everyone a voice and to drive the meeting to a set of priorities for a conclusive outcome. CrowdSmart Lite is an app in the Zoom app store. You might ask, “What can we do to make everyone in our video conference calls feel more engaged?”
If you have existing projects in innovation or co-creating with internal teams or customers, contact us. We can import your data into the system, allowing you to see your project through the lens of co-creation with the CRF.
If you have a Slack discussion and want to focus it on an action and decision forecasting project, let us know we will translate existing communication streams into focused co-creation projects.
What ideas do you have? Contact me at firstname.lastname@example.org
 MIT Technology interview
 Judea Pearl interview The Book of Why Basic Books 2018
 Tetlock, P. E., & Gardner, D. (2015). Superforecasting: The art and science of prediction.
Page, Scott E., (2018). The Model Thinker
 Scott Page is a scientific advisor to CrowdSmart and played a key role in the discussions in creating the collective reasoning facilitation process.
 Collective Voice Analysis data unveils a fundamental property of human co-creation activity: the results all appear to follow a power law distribution. Specifically, the ideas that align people follow the same mathematics as wealth distribution, population of cities, and trading volumes on the stock market. This topic will be covered in a separate paper.