Human input crucial to maximise effectiveness of AI, says new research

Research at Cambridge Judge Business School tackled the issue in a different way – and it provides valuable insight on how AI and humans can best work together creatively.

The findings? While collaboration between humans and generative AI does not automatically enhance their joint creativity – even after multiple rounds of co-creation sessions – joint creativity does improve over time if there are instructions and guidance on idea co-development. The research offers important practical insight for improving human-AI co-creation to achieve greater joint creativity.

It is co-authored by Luna Luan, a Lecturer in Management at the University of Queensland in Australia who recently earned her PhD at Cambridge Judge, Yeun Joon Kim of Cambridge Judge Business School and Jing Zhou, Professor of Management at Rice University in Texas.

One key aspect of the research is that it introduces a novel theory that develops a new conceptualisation of augmented learning - a concept first identified in a 1962 study as the phenomenon in which technology enhances human ability to acquire knowledge more effectively.

The research by Joon and colleagues shifts the focus of augmented learning from traditional human cognitive learning to collective learning between human and AI. In the context of human-AI co-creation, they define augmented learning as an evolutionary process in which humans and AI continuously adjust their levels of involvement across various co-creation activities (e.g., idea generation, feedback provision, combining ideas, refining existing ideas) in order to improve joint creativity over time.

While the original concept of augmented learning viewed technology as a supportive tool that boosts human cognitive learning by making existing data and other information more accessible, the advent of GenAI “has transformed technology from a passive assistant to a potential co-creator,” says the research.

It adds: “Although GenAI, by definition, must be initiated by humans, once collaboration begins, it can actively participate in tasks and decision-making. In this way, GenAI moves beyond the traditional role of technology as a mere tool for supporting human cognitive learning.

“Therefore, we propose that the focus of augmented learning between humans and GenAI should shift from an intrapersonal cognitive phenomenon to a collective, intersubjective process. In this reconceptualisation, augmented learning can be understood through collective learning between humans and GenAI.”

Although GenAI, by definition, must be initiated by humans, once collaboration begins, it can actively participate in tasks and decision-making. In this way, GenAI moves beyond the traditional role of technology as a mere tool for supporting human cognitive learning.

An example cited in the research as illustrative of effective human-AI pairing focuses on the creative production workflows at Netflix. Humans and AI collaborate at Netflix in script and content generation, but “instead of treating script development as a single task, Netflix breaks it down into sub-activities such as idea generation, evaluation, and selection, each involving different degrees of human and AI participation.”

So while human writers generate early drafts and storylines for Netflix productions, AI is used to assess character development, narrative pacing and patterns of audience demand to help refine scripts and position the production in a way that can be best marketed to audiences.

The research by Joon and colleagues traces previous studies in this area, highlighting that these studies (which have had positive, negative and mixed results) have focused on only a single collaboration between humans and GenAI on single-round creative tasks.

“These inconsistencies show that the existing approach to understanding the relationship between human-GenAI co-creation and joint creativity is limited, pointing to the need for a new theoretical lens,” the research says.

That new lens, outlined in a study published in Information Systems Research, is anchored in the concept of augmented learning in which humans and GenAI engage in an evolutionary process over time in order to improve creativity.

“This reconceptualisation reflects our new theoretical insights that the primary goal of human-GenAI co-creation is not limited to producing a single creative outcome but rather to achieving continuous improvement in their joint creativity via augmented learning,” say the authors.

As both humans and GenAI have limitations, this evolutionary process allows both humans and GenAI to continually adjust their roles in co-creation activities in a way that fosters joint creativity.

“We were surprised that human-AI pairs did not naturally improve through repeated collaboration,” says Joon. “Despite AI’s generative power, there was no spontaneous augmented learning: creativity simply did not increase over time. We found that improvement occurred only when we introduced a deliberate intervention.

“Specifically, instructing participants to engage in idea co-development – focusing on feedback and refinement rather than endlessly generating new ideas – was the key. Equally surprising was that idea generation itself had no relationship with joint creativity in human-AI collaboration, directly contradicting what we know from human-human creative teamwork.”

From a practical standpoint, the findings highlight that GenAI systems must be designed to support more than idea generation. Because the study shows that human-AI pairs failed to achieve improved joint creativity largely due to declining engagement in idea co-development over time (feedback exchange and iterative refinement), designers should build GenAI tools that actively prompt, scaffold and sustain co-development activities.

This includes features such as interactive feedback loops, mechanisms that encourage elaboration or critique of emerging ideas, and prompts that guide users toward deeper refinement rather than rapid idea production.

For organisations, the authors emphasise that effective human–AI co-creation requires deliberate structuring and guided learning, rather than assuming that creativity will improve automatically. The paper shows that augmented learning does not occur spontaneously and instead requires purposeful intervention – such as templates, instructions, or structured workflows – to help humans recognise when and how to engage in effective co-creation activities.

As a result, organisations should invest in training and development programmes that teach employees how to use GenAI not only as an idea generator but as a collaborative partner for refining and improving ideas. Such programmes should include hands-on exercises that cultivate skills in giving and receiving feedback with AI, integrating and reframing ideas, and iteratively improving concepts.

The study also urges a mindset shift in how organisations conceptualise human–AI collaboration. Instead of treating automation (AI doing the task) and augmentation (human–AI doing the task together) as a binary choice, the authors argue that tasks should be broken down into distinct co-creation activities with varying levels of human and AI involvement.

This expanded spatial scale allows organisations to match human and AI strengths more precisely for example – using AI for broad idea generation while positioning humans as critical evaluators and context experts during co-development. Such decomposition also supports augmented learning over time by enabling teams to iteratively adjust involvement patterns as they better understand the strengths and limitations of both humans and AI.

“Since the release of ChatGPT in 2022, many companies have rushed to adopt GenAI systems, driven by the belief that GenAI will empower employees to generate more creative ideas and thereby enhance overall firm performance,” the authors conclude.

“However, our research shows that the mere implementation of GenAI does not automatically offer these benefits. Particularly with GenAI, the effectiveness of its integration depends significantly on how well human users understand and interact effectively with it.”