Blogpost written by Chennampalli Bhanu Prakash Reddy and Mohammed Imthiyaz Basha
Key Points
- Human-AI collaboration exceeds in performance either of them individually.
- Humans often lack metaknowledge, which is the ability to assess their own capabilities, in making apt delegation of tasks.
- AI outperforms humans in delegating tasks.
In many professions, repetitive tasks dominate the workday, leaving little room for more difficult and meaningful tasks that require true expertise. Consider doctors who spend hours looking at medical images, processing lab results, or dealing with routine diagnoses that do not require their full attention. Every day, customer service representatives answer routine inquiries. These tasks, while important, are monotonous and lead to decreased job satisfaction. And that’s where artificial intelligence comes in, providing an alternative by taking on these routine tasks. By automating these tasks, AI allows professionals to focus on what is truly important—whether that is solving complex medical cases or assisting customers in need.
However, as AI is increasingly integrated into our day-to-day work, a very important question arises: how do we balance human expertise with the capabilities of AI? Is it up to the humans to determine which tasks to outsource to AI, or can AI learn when to ask for human intervention? In healthcare, for example, it could be that AI handles the routine cases and flags complicated ones for the doctors. Similarly, in customer support, AI can handle repetitive inquiries while sensitive issues will need to be handled by human agents.
When human intelligence complements the work of AI or vice versa, the outcomes become powerful, enhancing efficiency and effectiveness. This balance between technology and human skills needs careful thought to make sure the partnership leads to the best possible results while trust and control are maintained.
The promise of Human-Al Collaboration
The idea of human and AI collaboration is a simple one: humans possess tactical knowledge due to their experiences and AI processes vast amounts of data and identifies patterns. Human thoughts are the most diverse that are partially visible in their actions. So humans may have knowledge that complements AI algorithms. Collaboration of work between humans and AI can leverage these complementarities and their joint performance may exceed the performance of either of them individually.
In the study by Fügener et al. (2022), human participants performed image classification tasks in collaboration with an AI system. The research was conducted in a controlled experimental setting to analyze decision-making processes in human-AI collaboration. Participants were required to decide when to delegate tasks to the AI and how to integrate its recommendations into their decision-making.
The experiments revealed that task delegation between humans and AI led to improved outcomes. When humans delegate tasks to AI (delegation), results surpass those of humans working alone. Likewise, when AI delegates tasks to humans (inversion), performance exceeds that of AI acting independently.
But here’s the twist from Fügener et al (2022): When humans delegate tasks to AI the performance does not quite improve. However, when AI delegates tasks to humans accuracy has improved significantly in image classification.
The Delegation Dilemma: Why Humans Struggle
Why do you think this disparity is prevalent? The key seems to lie in something known as meta knowledge. Meta knowledge basically is the ability to assess one’s capabilities about knowing what you know and crucially what you don’t know. This ties closely to Polanyi’s Paradox, which says that humans know more than they can explicitly articulate. In other words, we can perform complex tasks intuitively, but we often struggle to explain how we do them. Before humans delegate to AI they need to be able to judge whether they can solve a problem or not.
The research reveals that a significant meta knowledge gap exists, humans are surprisingly bad at assessing their own abilities especially with the difficult tasks. This leads to inappropriate delegations which results in poor performance of Human-AI collaboration in predictions.
The AI Advantage
- Accurate Self assessment: In contrast to humans the AI is pretty good at assessing itself; it knows when it is likely to get an answer wrong and can then delegate to humans to attempt the difficult tasks. For instance, an AI can easily delegate tasks to humans when the prediction confidence falls below a certain threshold; otherwise, it continues handling the predictions independently.
- Consistent Delegation: Sometimes humans fail in delegation of tasks because of their overconfidence, uncertainty and mistrust of AI algorithms. But AI does not suffer from this problem, since AI can accurately assess its own performance, delegation strategies (delegate when the threshold value is not met) and not be hampered by cognitive biases like humans.
- Reliability: AI’s capacity to handle enormous amounts of data with consistency and precision makes it extremely trustworthy. Unlike humans, AI is unaffected by fatigue or emotional influences, resulting in consistent performance over time. Its ability to learn from massive volumes of data also allows it to develop and adapt to new difficulties, assuring long-term performance.
Can delegation strategies reinforce humans in collaboration?
The study by Fügener et al (2022), highlights how AI improves by taking help from humans. When the confidence level of AI falls below a certain threshold, it knows that it requires humans to complete the task. On the other hand, Humans often make poor choices in delegating tasks to AI. They tend to keep challenging tasks with them and assign easier ones to AI. This is not due to mistrust of AI, but because they fail to understand their own abilities of handling different tasks. By figuring out better ways of delegating tasks, humans could work more effectively with AI.
To address this issue, researchers examined two groups of participants. One group was simply explained delegation strategies, while the other was instructed to follow them. Both groups showed some improvements in both accuracy and delegation rates. Accuracy increased slightly, while delegation rates saw a significant rise. Interestingly, the delegation rates in the enforced strategy condition (second group) were similar to those in the explained strategy condition (first group), indicating that participants were effectively following the suggested delegation rules and were not reluctant to use AI in their work areas. This suggests that clear guidance, along with enforced practices, helps humans make better decisions when collaborating with AI.
Creating a Collaborative Future: Humans and AI in Sync
To truly unlock the potential of Human-AI collaboration, we must first address the gap in meta-knowledge. This involves developing AI systems that provide tailored feedback, helping humans better understand their strengths and weaknesses. Additionally, designing systems that enable AI to proactively delegate tasks to humans, based on its assessment of both its own capabilities and the human’s, will improve efficiency and collaboration.
Whether in healthcare, business, or other fields, the question is not whether humans or AI perform better individually, but how they can work together to achieve better results. The key is to strike the right balance, combining AI precision and human intuition to create a partnership that truly outperforms its individual parts.
Investigate further
Andreas Fügener, Jörn Grahl, Alok Gupta, Wolfgang Ketter (2022) Cognitive Challenges in Human–Artificial Intelligence Collaboration: Investigating the Path Toward Productive Delegation. Information Systems Research 33(2):678-696. https://doi.org/10.1287/isre.2021.1079