Definition and Principles

Human-in-the-Loop (HITL) systems represent a powerful framework that allows for the seamless collaboration between human intelligence and machine learning processes. The essence of HITL lies not just in improving machine learning performance, but also in enhancing human efficiency by enabling dynamic, real-time interaction with AI models throughout the learning process. By embedding human expertise, intuition, and contextual knowledge into various stages of data processing, HITL systems unlock the potential for a more symbiotic relationship between human oversight and automation (Monarch, Reference Monarch2021).

At its core, HITL systems engage humans in tasks like data preprocessing, annotation, and interactive labeling, where human insight is crucial for optimizing the quality and context of data that machines learn from. This direct involvement ensures that the strengths of human cognition – such as pattern recognition, domain expertise, and creative problem-solving – complement the raw processing power of machine learning algorithms (X. Wu et al., Reference Wu, Xiao, Sun, Zhang, Ma and He2022).

In domains like Natural Language Processing (NLP) and Computer Vision (CV), where training data often requires nuanced interpretation, the HITL paradigm can significantly enhance both the accuracy and efficiency of AI models. By iteratively incorporating human feedback into the learning cycle, these systems mitigate the limitations posed by sparse, ambiguous, or noisy data, fostering the development of AI that is not only technically proficient but also adaptable to complex real-world scenarios (Mosqueira-Rey et al., Reference Mosqueira-Rey, Hernández-Pereira, Alonso-Ríos, Bobes-Bascarán and Fernández-Leal2023). The ultimate goal of HITL is to integrate the strengths of both humans and machines, leveraging the precision of machines and the adaptability of humans to accelerate AI innovation and improve decision-making across various domains.

The foundational principles of Human-in-the-Loop Machine Learning (HITL-ML) are as follows (Monarch, Reference Monarch2021):

• Improving Model Accuracy: One of the core objectives of HITL-ML is to enhance the accuracy of machine learning models by incorporating human feedback during key stages of the learning process. Human insight, particularly in tasks like annotation and validation, adds critical value to data quality, thereby refining model predictions.

• Accelerating Target Accuracy: HITL systems aim to expedite the process of reaching desired accuracy levels for machine learning models. By involving humans in the iterative learning loop, the model can achieve optimal performance more quickly through targeted corrections and guidance.

• Maximizing Combined Intelligence: HITL synergistically combines human intuition and machine precision to maximize accuracy. By leveraging the strengths of both human decision-making and automated learning, these systems can overcome limitations inherent in either component alone.

• Enhancing Human Efficiency with AI Assistance: Another fundamental principle of HITL is to increase human efficiency by automating routine and labor-intensive tasks. Machine learning algorithms assist in handling large-scale data and repetitive tasks, freeing human experts to focus on higher-level decision-making and strategic oversight.

Benefits and Limitations

As with any system, HITL systems come with their inherent advantages and challenges.

The benefits of HITL systems include:

1. (1) Improved Accuracy in Sparse Data Contexts: One of the most significant benefits of HITL systems is their ability to excel in situations where data is scarce. By integrating human a priori knowledge, especially in data-limited fields such as clinical diagnosis, HITL systems can enhance model accuracy. In cases where high-quality labeled data is not readily available, human expertise fills the gap, ensuring that models perform optimally despite data constraints (X. Wu et al., Reference Wu, Xiao, Sun, Zhang, Ma and He2022).

2. (2) Enhanced Model Performance through Minimal Feedback: Another key benefit of HITL is its ability to improve model performance with minimal human input. Cognitive science research shows that even small datasets of human feedback can lead to substantial improvements in machine learning outcomes. HITL systems leverage this principle, allowing better training outcomes even with limited human engagement, thus increasing efficiency (X. Wu et al., Reference Wu, Xiao, Sun, Zhang, Ma and He2022).

3. (3) Increased Interpretability and Usability: Traditional machine learning models can often be difficult to interpret, leading to hesitancy in their adoption. HITL systems address this issue by incorporating human input into the learning process, which not only improves the interpretability of the results but also makes the models more usable in real-world applications. This human–AI interaction builds trust and allows practitioners to better understand and utilize AI-generated insights.

4. (4) Tailored Learning through Human Preferences: HITL systems offer the unique advantage of aligning AI models more closely with human preferences. By incorporating datasets reflecting human preferences, such as preferred summaries, HITL systems employ supervised learning to develop reward models that predict outcomes aligned with human choices. Reinforcement learning can then be applied to maximize these predictions, ensuring models meet user expectations (Ziegler et al., Reference Ziegler, Stiennon, Wu, Brown, Radford, Amodei, Christiano and Irving2020).

While HITL systems offer numerous advantages, several challenges remain. These potential constraints highlight areas where further research and refinement are necessary.

1. (1) One of the key challenges in HITL systems is how to efficiently and effectively embed human expertise and knowledge into machine learning models. Current approaches often rely heavily on manual data annotation (X. Wu et al., Reference Wu, Xiao, Sun, Zhang, Ma and He2022), but this limits the system’s ability to learn from human experience on a broader scale. Additionally, maintaining system robustness while integrating human input is a critical issue.

2. (2) Another significant limitation is the absence of standardized evaluation benchmarks for HITL systems (X. Wu et al., Reference Wu, Xiao, Sun, Zhang, Ma and He2022). Determining which key samples to use and how to construct appropriate evaluation metrics remains unclear. This becomes particularly important in scenarios such as active learning, where learners request human labeling. Ensuring the questions are presented in a clear and understandable manner is crucial, making the usability features – such as clarity, consistency, and efficiency – highly relevant in these systems (Mosqueira-Rey et al., Reference Mosqueira-Rey, Hernández-Pereira, Alonso-Ríos, Bobes-Bascarán and Fernández-Leal2023).

3. (3) As HITL systems increasingly involve human input, ethical considerations such as fairness, transparency, and interpretability become more pressing. While HITL aims to make machine learning more understandable and reliable, integrating these human aspects into AI systems requires careful thought and ongoing development to ensure equitable and transparent outcomes (Mosqueira-Rey et al., Reference Mosqueira-Rey, Hernández-Pereira, Alonso-Ríos, Bobes-Bascarán and Fernández-Leal2023).

Definition and Principles

Active Learning (AL) is a machine learning approach where the learner plays an active role in selecting data for human annotation (Monarch, Reference Monarch2021). The learner requests an oracle, typically a human expert, to label ambiguous or informative examples that will contribute significantly to improving the model (Mosqueira-Rey et al., Reference Mosqueira-Rey, Hernández-Pereira, Alonso-Ríos, Bobes-Bascarán and Fernández-Leal2023). This method is designed to address the challenge of the labeling bottleneck by iteratively presenting unlabeled instances to be annotated by the oracle (Burr, Reference Burr2009; Schröder & Niekler, Reference Schröder and Niekler2020). Active Learning aims to reduce the amount of data that requires human labeling while maintaining or even enhancing the overall model performance (Schröder & Niekler, Reference Schröder and Niekler2020). Through selective data sampling, AL allows for more efficient learning, leveraging both labeled and unlabeled data in the process (P. Liu et al., Reference Liu, Wang, Ranjan, He and Zhao2022; Monarch, Reference Monarch2021).

Active learning has already been widely applied in various fields, including collaborative filtering recommender systems, supervised remote sensing image classification, text classification using deep neural networks, and medical image analysis (P. Liu et al., Reference Liu, Wang, Ranjan, He and Zhao2022).

The following core principles guide the operation of Active Learning systems:

• Learner Control Over Data: In Active Learning, the learner actively controls which data to query. The learner can request a knowledgeable entity, often a human expert, to annotate selected unlabeled examples, thereby focusing on the most informative data (Mosqueira-Rey et al., Reference Mosqueira-Rey, Hernández-Pereira, Alonso-Ríos, Bobes-Bascarán and Fernández-Leal2023).

• Iterative and Incremental Annotation Process: Active Learning follows an iterative, cyclic process in which the learner employs a query strategy to request labels for specific examples. These labeled examples are then used to refine and improve the model incrementally (Mosqueira-Rey et al., Reference Mosqueira-Rey, Hernández-Pereira, Alonso-Ríos, Bobes-Bascarán and Fernández-Leal2023).

• Combination of Labeled and Unlabeled Data: AL effectively operates as a form of semi-supervised learning, combining both labeled and unlabeled data. This approach is particularly valuable when labeled data is costly or scarce, enabling efficient use of available resources (Burr, Reference Burr2009).

• Query Strategy for Effective Sampling: A critical aspect of AL is the query strategy (X. Wu et al., Reference Wu, Xiao, Sun, Zhang, Ma and He2022), which determines which examples the learner requests for labeling. Common strategies include:

  1. (a) Membership Query Synthesis: The learner may generate and request labels for any instance, even synthetic ones.

  2. (b) Stream-based Selective Sampling: The learner evaluates instances one at a time, choosing whether to query or discard each.

  3. (c) Pool-based Sampling: The learner evaluates a pool of data to select the most informative examples for labeling.

• Focus on Informative Instances: The ultimate goal of Active Learning is to select the most informative or uncertain examples for annotation. This approach maximizes the efficiency of the labeling process, leading to faster model improvement and reduced reliance on large labeled datasets.

Quality Control and Cost Consideration

Active learning’s quality control is centered on selecting the most valuable samples for labeling, assuming that human experts provide accurate annotations. The design of the selection process, whether model-driven or data-driven, plays a critical role. Key factors influencing quality include (P. Liu et al., Reference Liu, Wang, Ranjan, He and Zhao2022):

1. (1) Data: Representativeness, diversity, and other characteristics influence the quality of samples. Sample distribution is an inherent property of data, and there are at least three primary methods to leverage it: representativeness, diversity, and core-set selection. These methods connect sample selection to distribution from different angles, as a metric is still required to make choices, even when the distribution is well understood.

   1. (a) Representativeness: Selecting samples that effectively represent the overall data distribution, ensuring that the chosen samples reflect the key characteristics of the entire dataset.

   2. (b) Diversity: Selecting samples that exhibit a broad range of variability within the sample space, capturing more distinctions and reducing model bias.

   3. (c) Core-set: Identifying a set of influential samples that can maximally summarize the information contained in other samples, thus simplifying the learning process.

2. (2) Model Attributes: Metrics such as gradient length and Fisher information are used to assess how much influence unlabeled data can have on the model (Chaudhuri et al., Reference Chaudhuri, Kakade, Netrapalli and Sanghavi2015; Sourati et al., Reference Sourati, Akcakaya, Leen, Erdogmus and Dy2017).

3. (3) Metrics: Includes adversarial metrics and uncertainty measures to guide sample selection.

   1. (a) Best-versus-Second Best (BvSB): Compares the probabilities of the top two predictions to identify the most uncertain samples (Joshi et al., Reference Joshi, Porikli and Papanikolopoulos2009).

   2. (b) Multiple Peak Entropy (MPE): Measures the entropy in the model’s predictions, selecting samples that reflect multiple uncertainty modes (B. Liu & Ferrari, Reference Liu and Ferrari2017).

   3. (c) Query by Committee (QBC): Uses a committee of models to assess disagreement among them, selecting samples where the models exhibit the most uncertainty (Kee et al., Reference Kee, Del Castillo and Runger2018).

Cost considerations are an essential part of active learning. While annotating data is typically expensive, active learning helps mitigate these costs by minimizing the number of labeled samples required for effective learning. Some costs that may need to be considered in research and practice include:

• Decision Cost: In active learning, decision cost considers both the expense of labeling data and the potential cost of future misclassification. This requires balancing the cost of annotation with the risk of errors if the instance is used in training. For example, researchers suggest summing labeling costs, assumed to be proportional to instance length, with expected misclassification costs (Kapoor et al., Reference Kapoor, Horvitz and Basu2007). While practical, this approach requires both costs to be expressed in the same currency, which can be challenging for some applications.

• Model Training Costs: In active learning, sample selection can be computationally expensive, as each iteration typically requires re-evaluating the informativeness of every data point. This iterative process can be time-consuming.

• Expert Availability: In active learning, expert availability is a critical cost, particularly in specialized fields like medical image analysis, where labeling requires a high level of expertise (Budd et al., Reference Budd, Robinson and Kainz2021). The limited availability of qualified experts can slow down the annotation process and increase costs.

• Annotation Time: In active learning, annotation time could be a key cost, as the time spent labeling data can significantly impact the overall efficiency of the process. The design of the user interface and the size of the labeled set also play crucial roles in managing this cost. For example, researchers used active learning for morpheme annotation in rare language documentation, employing two human oracles to reduce both annotation time and corpus size (Baldridge & Palmer, Reference Baldridge and Palmer2009). Similarly, a study highlights that the interface design can be just as important as the active learning strategy itself in reducing time and cost, with the size of the labeled set directly influencing overall expenses (Druck et al., Reference Druck, Settles and McCallum2009).

Benefits and Limitations

Active learning offers significant benefits, particularly in terms of cost efficiency and managing scenarios with limited data. By focusing on selecting only the most informative samples, it reduces the number of labeled instances required, leading to significant cost savings while maintaining high-quality annotations (Mosqueira-Rey et al., Reference Mosqueira-Rey, Hernández-Pereira, Alonso-Ríos, Bobes-Bascarán and Fernández-Leal2023). This is especially valuable in domains where labeled data is scarce or expensive to obtain, such as rare disease detection, clinical decision-making, or specialized medical image analysis (Budd et al., Reference Budd, Robinson and Kainz2021). For instance, in radiology reports (Hoi et al., Reference Hoi, Jin, Zhu and Lyu2006; Nguyen & Patrick, Reference Nguyen and Patrick2014) or biological research (Luo et al., Reference Luo, Kramer, Goldgof, Hall, Samson, Remsen, Hopkins and Cohn2005), active learning enables models to achieve high accuracy with fewer labeled examples, minimizing annotation costs. Moreover, active learning is highly effective in addressing data limitations, such as class imbalance or constrained annotation budgets. In such cases, it enhances the performance of pre-trained models (PTMs) and large language models (LLMs) by optimizing sample selection in imbalanced datasets, making it a powerful tool for maximizing efficiency in resource-limited settings (Dor et al., Reference Dor, Halfon, Gera, Shnarch, Dankin, Choshen, Danilevsky, Aharonov, Katz and Slonim2020).

However, active learning is not without limitations (Donmez & Carbonell, Reference Donmez and Carbonell2008; Settles, Reference Settles2011). The assumption that human oracles are infallible and always available may not hold in practice (Mosqueira-Rey et al., Reference Mosqueira-Rey, Hernández-Pereira, Alonso-Ríos, Bobes-Bascarán and Fernández-Leal2023). Human annotators can experience fatigue, and labeling quality may vary over time. Additionally, querying samples one at a time can be inefficient, leading to batch querying methods, which, while more efficient, may still not fully address the variability in labeling quality. Furthermore, the cost of obtaining new labels is often assumed to be fixed, yet in real-world scenarios, this cost can vary depending on the complexity of the data. Lastly, in situations where data distributions or model classes change, active learning systems must be adaptable enough to incorporate new knowledge and adjust accordingly.

Definition and Principles

Crowdsourcing refers to an online, distributed problem-solving and production model (Brabham, Reference Brabham2008), where tasks are outsourced to a large group of people – often through platforms like Amazon Mechanical Turk (Paolacci et al., Reference Paolacci, Chandler and Ipeirotis2010). The term was first coined in 2006 by Jeff Howe and Mark Robinson to describe how businesses were leveraging the Internet to “outsource work to the crowd” (Howe, Reference Howe2006). This approach has since become integral to machine learning, advancing key areas such as data generation, model evaluation, hybrid human–AI systems, and behavioral experiments that enhance our understanding of human interaction with AI technologies (Sheng & Zhang, Reference Sheng and Zhang2019; Vaughan, Reference Vaughan2018). By harnessing the complementary strengths of humans and machines, crowdsourcing enables the expansion of AI capabilities through collaborative intelligence.

Crowdsourcing operates on several key principles:

1. (1) Collective Intelligence: The concept of the “wisdom of crowds” suggests that large groups of people, when working together, can often solve problems or make judgments more effectively than individual experts (Kameda et al., Reference Kameda, Toyokawa and Tindale2022). This collective intelligence is fundamental to crowdsourcing, where the input from many contributors can lead to more accurate or creative solutions.

2. (2) Task Submission and Contribution: Crowdsourcing involves two main groups: requesters, who submit tasks to a crowdsourcing platform, and workers, who form the crowd that contributes to solving these tasks. The output is the result of this collective effort, and its quality often depends on the diversity and volume of contributions.

3. (3) Evaluation and Rewards: Requesters may evaluate the quality of the outputs and provide rewards based on performance. In some cases, quality control is delegated to the crowdsourcing platform, where outputs are evaluated and rewarded automatically. Rewards can vary, ranging from monetary compensation to gifts, reputation badges, or other incentives (Daniel et al., Reference Daniel, Kucherbaev, Cappiello, Benatallah and Allahbakhsh2019).

These principles ensure that crowdsourcing initiatives, regardless of their specific application, follow a structured process that leverages collective human intelligence to enhance AI systems (Lease, Reference Lease2011).

Quality Control and Cost Consideration

When implementing crowdsourcing, two primary concerns are cost efficiency and quality control. The challenge lies in ensuring that data labels and prediction models derived from crowdsourced work maintain high quality, especially given the varied skills and motivations of the workers (Daniel et al., Reference Daniel, Kucherbaev, Cappiello, Benatallah and Allahbakhsh2019; Sheng & Zhang, Reference Sheng and Zhang2019).

From a quality model perspective, several key factors affect the final output quality (Daniel et al., Reference Daniel, Kucherbaev, Cappiello, Benatallah and Allahbakhsh2019):

• Data: The quality of both input (e.g., images, text) and output (e.g., labeled data, translated text) is crucial. High-quality data output is vital for the broader acceptance and effectiveness of crowdsourcing systems.

• Task Design: How a task is described, structured, and managed significantly impacts worker engagement, output quality, and overall performance. Clear descriptions, user-friendly interfaces, and well-designed incentives (whether intrinsic, like status, or extrinsic, such as financial rewards) contribute to task efficiency. Additionally, terms and conditions, including privacy, intellectual property rights, and ethical standards, influence worker participation and compliance. Effective resource management is also important, ensuring that the quantity of useful work completed aligns with the resources consumed, thus sustaining productivity in crowdsourcing initiatives.

• People: Both the requesters (who submit tasks and evaluate outputs) and workers (who complete tasks) play critical roles. Characteristics such as fairness, communication, and worker motivation affect the overall success of crowdsourcing tasks. Additionally, group dynamics, such as collaborative teams or the larger crowd, influence task performance and quality.

Table 8.2 outlines various methods used to evaluate outputs in crowdsourcing from a quality assessment perspective.

To effectively manage crowdsourcing quality control, researchers should focus on several key strategies: incorporating more labeled data, integrating hybrid systems, targeting uncertain or diverse data, and conducting ongoing assessments. Additionally, using on-demand evaluations and creating more benchmarks can help minimize the need for reusing data (Lease, Reference Lease2011).

Benefits and Limitations

Crowdsourcing offers several significant benefits that make it a valuable tool in human–AI collaboration:

• Lowered Costs and Improved Speed: Crowdsourcing allows for faster and more cost-effective data collection by tapping into a diverse workforce. This flexibility helps accelerate data gathering and reduces the financial burden on organizations (Vaughan, Reference Vaughan2018).

• Improved Quality: By leveraging the collective intelligence of a diverse group of contributors, crowdsourcing can handle tasks that are difficult for computers alone. Human workers bring common sense, practical knowledge, and real-world experience to the table, leading to more nuanced and accurate outputs (Gomes et al., Reference Gomes, Welinder, Krause and Perona2011).

• Increased Flexibility: Crowds can provide dynamic data for training machine learning algorithms. The ability to generate and aggregate potentially noisy labels has spurred research into methods for improving label accuracy, making crowdsourcing an effective tool for training AI models (Zhang et al., Reference Zhang, Wu and Sheng2016).

• Promoting Diversity: Crowdsourcing encourages input from a broad range of contributors, leading to diverse perspectives and insights, which can enhance the quality and breadth of solutions (Vaughan, Reference Vaughan2018).

• Data Generation and Evaluation: Crowdsourcing platforms are particularly adept at generating large datasets and evaluating machine learning models, for example, evaluating the coherence of topic models and enhancing the interpretability of predictions in supervised learning (Chang et al., Reference Chang, Gerrish, Wang, Boyd-Graber and Blei2009; Ribeiro et al., Reference Ribeiro, Singh and Guestrin2016).

• Hybrid Intelligence Systems: Crowdsourcing supports hybrid intelligence systems that combine human reasoning with machine learning. These systems, drawing on human common sense, subjective beliefs, and life experiences, can outperform traditional AI in specific tasks that require flexible reasoning (Chang et al., Reference Chang, Gerrish, Wang, Boyd-Graber and Blei2009; Ribeiro et al., Reference Ribeiro, Singh and Guestrin2016).

Despite these advantages, crowdsourcing comes with notable limitations:

• Impact on Product Quality: The inclusion of unqualified participants can lead to a significant number of low-quality or unusable contributions (Eapen et al., Reference Eapen, Finkenstadt, Folk and Venkataswamy2023). To mitigate errors, employers often rely on multiple workers to complete the same task, which increases both time and costs (Ipeirotis et al., Reference Ipeirotis, Provost and Wang2010).

• Motivation and Engagement: It can be difficult to maintain consistent engagement and motivation among crowd participants. Varying levels of dedication can result in inconsistent contribution quality, affecting overall outcomes.

• Cultural and Contextual Differences: Differences in cultural background and context among contributors may lead to misunderstandings or inconsistencies in contributions, posing a challenge for maintaining uniform quality.

• Information Overload: The large volume of data generated through crowdsourcing can create challenges in filtering, managing, and processing information effectively.

• Data Security and Privacy Concerns: Crowdsourcing tasks may inadvertently expose sensitive personal information, such as when workers input private data like business card details or share location information through certain tasks, raising concerns over privacy and security (Deng et al., Reference Deng, Shahabi and Demiryurek2013).

Definition and Principles

Interactive Machine Learning (IML) refers to a learning paradigm where humans are actively involved in the machine learning process. Unlike classical machine learning (CML), which operates in a passive, offline manner, IML allows users – whether experts or non-experts – to interactively guide, correct, and refine models as they are being trained. This iterative process ensures that the human agent can adjust the learning process to meet specific needs, optimizing the performance and relevance of the models (Fails & Olsen, Reference Fails and Olsen2003; Porter et al., Reference Porter, Theiler and Hush2013; Ware et al., Reference Ware, Frank, Holmes, Hall and Witten2001).

IML systems enable humans to perform tasks that are complex for machines, such as refining classification boundaries or providing annotations in image segmentation. The human role can vary, from providing input at the beginning (e.g., initial data labeling) to intervening at the end (e.g., validating machine outputs). This interaction allows for a dynamic collaboration between humans and machines, enabling each to focus on the tasks they perform best (Porter et al., Reference Porter, Theiler and Hush2013; Ramos et al., Reference Ramos, Meek, Simard, Suh and Ghorashi2020).

Several key principles underpin IML systems:

• Humans in the ML Loop: In IML, humans play an active role in the learning loop, performing tasks that require their expertise, such as labeling or correcting outputs. The inclusion of human agents makes IML particularly valuable in domains where uncertainty and complexity prevail, such as health informatics (Holzinger, Reference Holzinger2016).

• Incremental and Iterative Learning: IML is characterized by an incremental and iterative learning process. Human input is continuously integrated as the model evolves, ensuring that the system adapts to real-time feedback and improves over time (Dudley & Kristensson, Reference Dudley and Kristensson2018).

• Multiple Human Roles: The individuals involved in IML can assume various roles, from machine learning experts and data scientists to crowdsource workers or domain specialists. The diversity of human roles influences both the interaction and the outcomes of the system (Ramos et al., Reference Ramos, Meek, Simard, Suh and Ghorashi2020).

• User-Friendly Interfaces: The design of the user interface is critical in IML systems, as it determines how users interact with the model and influences the overall learning outcome. Well-designed interfaces can empower non-experts to effectively contribute to the model-building process (Q. Yang et al., Reference Yang, Suh, Chen and Ramos2018).

In IML, quality control and cost considerations go beyond traditional metrics like accuracy. IML involves subjective evaluations that account for factors such as cost, confidence, and task complexity. Unlike classical approaches, IML adopts a human-centered perspective, placing emphasis on the model’s utility and effectiveness for end-users.

Quality Control and Cost Consideration

Quality control in IML integrates both algorithm-centered evaluations, which are typical in Active Learning, and human-centered assessments (Fiebrink et al., Reference Fiebrink, Cook and Trueman2011). This dual approach helps to address the “black-box” nature of machine learning algorithms, offering deeper insights into the model’s performance and making it more interpretable and accessible to users (Boukhelifa et al., Reference Boukhelifa, Bezerianos, Lutton, Zhou and Chen2018). Ultimately, this ensures that IML systems are not only accurate but also practical, efficient, and user-friendly.

The cost considerations in IML include the time and resources required for human involvement. Since humans are actively engaged in the iterative learning process – either by providing annotations or refining models – the cost of human labor becomes a significant factor. However, by focusing on tasks where human input is essential, IML can optimize resource usage while still producing high-quality outcomes that meet the needs of its users.

Benefits and Limitations

IML offers several key benefits, contributing to its robustness, trustability, and lower resource demands (Wondimu et al., Reference Wondimu, Buche and Visser2022):

• Robustness: IML enhances the robustness of models by allowing users to understand how variations in input data affect predictions. This transparency helps in creating more resilient models that can perform reliably across different scenarios and datasets.

• Trustability: By providing clear explanations for the model’s decisions, IML builds trust among users and stakeholders. When people can comprehend and verify the reasoning behind predictions, they are more likely to accept and rely on the model’s outcomes.

• Low Resource Requirements: IML often requires fewer computational resources than traditional machine learning models. This makes it suitable for low-resource environments, enabling the deployment of interpretable models in diverse applications, especially where computational power is limited.

Despite its benefits, IML also presents certain limitations (Mosqueira-Rey et al., Reference Mosqueira-Rey, Hernández-Pereira, Alonso-Ríos, Bobes-Bascarán and Fernández-Leal2023):

• Increased Development Complexity: IML combines machine learning with human–computer interaction, which significantly increases the complexity of developing such systems. Each IML application needs to be carefully designed and studied, complicating the overall development process.

• Dependency on Human Expertise: While IML reduces computational demands and enhances learning efficiency, it relies heavily on human expertise. The system’s effectiveness can be limited by the availability and attention of human experts, as well as the varying levels of expertise, which can introduce inconsistencies and impact performance.

Healthcare

In the healthcare domain, human–AI collaborative intelligence has become a pivotal force in advancing medical practices, improving patient outcomes, and enhancing overall system efficiency. A prominent example is medical imaging, which plays a critical role in clinical decision-making across multiple areas, including detection, diagnosis, and treatment planning (Tajbakhsh et al., Reference Tajbakhsh, Jeyaseelan, Li, Chiang, Wu and Ding2020). While traditional methods for analyzing medical images, such as manual segmentation, are often time-consuming and labor-intensive, human–AI collaboration offers new avenues for improvement. One such approach is interactive segmentation, where clinicians can refine AI-generated initial segmentations. For instance, researchers introduced a dual-CNN-based framework that allows user interaction to correct model inaccuracies, significantly reducing the time needed for precise image segmentation in applications like brain tumor and fetal MRI imaging (G. Wang et al., Reference Wang, Zuluaga, Li, Pratt, Patel, Aertsen, Doel, David, Deprest, Ourselin and Vercauteren2019).

Active learning is another crucial application. In cancer pathology, where extracting key details from reports can be tedious, AI systems can assist by flagging challenging cases for human annotators, thereby reducing the manual workload (De Angeli et al., Reference De Angeli, Gao, Alawad, Yoon, Schaefferkoetter, Wu, Durbin, Doherty, Stroup, Coyle, Penberthy and Tourassi2021). This hybrid method has shown impressive results, achieving similar performance to traditional methods while requiring fewer labeled data. Furthermore, in heart disease prediction, models that solicit expert feedback for the most relevant data points can reduce costs and enhance accuracy (El-Hasnony et al., Reference El-Hasnony, Elzeki, Alshehri and Salem2022). Human experts also play an essential role in preprocessing medical data, ensuring that the AI system receives high-quality inputs. Such collaboration leads to improved outcomes, as seen in the classification of oral cancer tissues (Folmsbee et al., Reference Folmsbee, Liu, Brandwein-Weber and Doyle2018) and radiology reports (Nguyen & Patrick, Reference Nguyen and Patrick2014).

Maintaining effective human–AI collaboration remains essential in healthcare, where the accuracy and reliability of AI systems must meet the stringent requirements of patient safety (Budd et al., Reference Budd, Robinson and Kainz2021).

Finance

Human–AI collaboration is increasingly transforming the financial domain, combining AI’s computational power with human expertise to enhance efficiency, accuracy, and decision-making. In the lending domain, artificial intelligence is being employed to streamline loan processing and ensure responsible lending decisions. However, machines alone are not sufficient for decision-making. For instance, a human–AI collaborative method was introduced to evaluate multiple independent and conflicting pieces of evidence in loan applications (Sachan et al., Reference Sachan, Almaghrabi, Yang and Xu2024). By comparing expert underwriters’ judgments with AI-generated results, the system learns human decision tendencies, optimizing decision quality through continuous human–AI interaction. In international maritime trade, algorithms automatically gather data to generate transaction tables, but human data scientists play a crucial role in reviewing and refining these outputs. Their feedback enhances algorithm performance and adaptability over time (Gronsund & Aanestad, Reference Gronsund and Aanestad2020). Moreover, active learning approaches in stock market sentiment analysis allow machines to query human experts for labels, significantly improving the model’s predictive performance (Smailovic et al., Reference Smailovic, Grcar, Lavrac and Znidarsic2014). This iterative process ensures that the AI continues to improve through expert guidance.

In summary, human–AI collaboration is essential for addressing complex financial challenges. This evolving synergy will continue to drive innovation and ensure fairness in financial services.

Agriculture

Human–AI collaboration has made significant advancements in the agricultural sector by improving efficiency in tasks such as crop monitoring and pest detection. For instance, in crop yield analysis, AI can detect and analyze cereal crops like wheat and sorghum, focusing on panicle density, a critical factor in evaluating yield. Researchers have introduced an automated system that combines weak and strong labeling methods to reduce annotation time and improve accuracy (Chandra et al., Reference Chandra, Desai, Balasubramanian, Ninomiya and Guo2020). This approach allows human annotators to collaborate with AI in a more efficient workflow, producing high-quality data with less manual effort. Similarly, distinguishing crops from weeds is another complex task where human expertise complements AI. Scientists have developed a model to identify samples requiring labeling, which are then manually annotated by experts. This method minimizes human workload while maintaining high classification accuracy (Sheikh et al., Reference Sheikh, Milioto, Lottes, Stachniss, Bennewitz and Schultz2020). In pest classification, research has demonstrated that cleaning redundant public datasets can reduce data size while achieving similar classification results, showcasing the power of human-guided data optimization (J. Yang et al., Reference Yang, Lan, Li, Gong, Zhang and Ercisli2022).

In essence, human–AI collaboration significantly advances agricultural practices by improving efficiency and accuracy, minimizing manual workload and leading to more effective and precise agricultural solutions.

Education

Human–AI collaboration has transformed personalized learning and education. Researchers have developed a system in which AI classifies student profiles, enabling human experts to customize learning experiences. This synergy between AI’s data processing capabilities and human expertise facilitates personalized feedback and content adjustments, leading to improved learning outcomes (De Melo et al., Reference De Melo, Flôres, De Carvalho, De Teixeira, Loja and de Sousa Gomide2014). In the realm of scientific reading, human–AI collaboration is critical for helping students comprehend complex documents. Researchers have proposed a collaborative PDF reader powered by Open Educational Resources (OERs), which uses machine learning to analyze reader behaviors, such as highlighting and questioning, and recommends relevant OERs like videos and slides. Readers provide feedback, allowing the system to refine its recommendations based on user interactions (Jiang et al., Reference Jiang, Liu, Gao and Tang2016; X. Liu et al., Reference Liu, Jiang and Gao2015). For complex mathematical content, the system suggests a Formula Evolution Map that tracks formula development and recommends related OERs (Jiang et al., Reference Jiang, Gao, Yuan, Gao, Tang and Liu2018). Human feedback is crucial, refining the system’s recommendations to ensure greater accuracy and contextual relevance. This collaboration between humans and AI improves understanding of complex concepts, making STEM (Science, Technology, Engineering, and Mathematics) reading more efficient and personalized.

Overall, human–AI collaboration in education leverages AI’s strength in data processing and the human ability to provide customized learning experiences. This partnership enhances personalized feedback, improves comprehension of scientific materials, and facilitates understanding of complex concepts through adaptive content recommendations and user interaction.

Misinformation Control and Cybersecurity

Human–AI collaboration has been applied in tackling misinformation and enhancing cybersecurity. For instance, researchers have used semi-supervised and active learning to detect camouflaged Chinese spam content, integrating human annotations with AI insights. This collaboration improves the accuracy of spam detection, making it more adaptable to evolving text camouflage and data imbalance challenges (Jiang et al., Reference Jiang, Gao, Duan, Kang, Sun, Zhang and Liu2020). Similarly, a security system has been developed by integrating AI with crowdsourced human feedback to combat online misinformation (Demartini et al., Reference Demartini, Mizzaro and Spina2020), while a human-centric vulnerability analysis system was created for cybersecurity (Shoshitaishvili et al., Reference Shoshitaishvili, Weissbacher, Dresel, Salls, Wang, Kruegel and Vigna2017).

These examples show how human–AI collaboration addresses complex challenges in information security, improving both the accuracy of threat detection and decision-making processes.

Law

The legal domain has embraced human–AI collaboration for tasks such as case classification and legal consultation. For instance, researchers have proposed a system where human analysts refine AI decision-making processes by adding new factors and adjusting the algorithm, ensuring flexible and informed legal decisions (Odekerken & Bex, Reference Odekerken, Bex, Villata, Harašta and Křemen2020). Similarly, another legal model employs Positive-Unlabeled Reinforcement Learning (PURL) to dynamically generate diagnostic legal questions. By integrating AI insights with human expertise, this model provides accurate, context-sensitive legal advice (Y. Wu et al., Reference Wu, Wang, Gumusel and Liu2024).

Generally, human–AI collaboration in the legal domain enables more flexible, informed, and context-sensitive decision-making by integrating AI’s analytical capabilities with human expertise, optimizing multiple legal processes.

Expanding Human–AI Collaboration in Other Domains

Human–AI collaboration also extends into other domains. For example, a human–machine interactive image search method was introduced by Kovashka et al. (Reference Kovashka, Parikh and Grauman2015), while human-in-the-loop semi-supervised learning has been applied for random gene regulation (Wrede & Hellander, Reference Wrede and Hellander2019). These efforts demonstrate the adaptability and scalability of human–AI collaboration across diverse industries, ensuring continuous innovation and improvement in performance.

As AI systems evolve, human oversight remains crucial for ensuring transparency, fairness, and efficiency. The integration of human expertise and AI drives innovation, improves decision-making, and enhances accuracy across various applications. These systems foster natural, efficient human–machine interactions, accelerating digital transformation and delivering smarter solutions across diverse industries.

Case 1: Education Domain: Scaffolding with Open Educational Resources (OER)

An exemplary case of effective human–AI collaborative intelligence in the education field is demonstrated through the development of the OER-based Collaborative PDF Reader (OCPR). This system showcases how AI and human intelligence can work together to enhance the comprehension of complex scientific literature for students and junior scholars (Jiang et al., Reference Jiang, Liu, Gao and Tang2016, Reference Jiang, Gao, Yuan, Gao, Tang and Liu2018; X. Liu et al., Reference Liu, Jiang and Gao2015).

A typical interface of the OCPR system is shown in Figure 8.2. The system integrates advanced text mining and heterogeneous graph mining algorithms to recommend relevant Open Educational Resources (OERs), such as videos, slides, source code, and Wikipedia pages, based on students’ interactions with scientific publications. By leveraging human actions, such as highlighting and asking questions, and, most importantly, feedback on the recommended content, the system dynamically adjusts its recommendations, enabling students to better grasp complex concepts.

Figure 8.2 OER-based collaborative PDF reader system

Figure 8.2Long description

The highlighted features include the embedding of student or instructor knowledge within the PDF, user-to-user collaboration for annotating information need, and highlighted text that marks important content. They also encompass system-driven extraction of users’ implicit or explicit information needs, access to further assistance via the “Get Help” drop-down menu, and user-system collaboration through feedback on recommended results.

This human–AI collaboration enables the AI to respond to the emerging information needs of students, while allowing human oversight and intervention to refine the system’s suggestions. The integration of human judgment with AI-driven recommendations creates a learning environment where complex materials are made more accessible, promoting deeper engagement and understanding.

This case highlights the power of human–AI collaboration in educational settings, where the ability of AI to scaffold learning is enhanced by human input, illustrating the future of personalized, intelligent education.

Case 2: Legal Domain: Knowledge-infused Legal Wisdom

One notable example of effective human–AI collaborative intelligence in the legal field is the development of the D3LM (Diagnostic Legal Large Language Model) framework (Y. Wu et al., Reference Wu, Wang, Gumusel and Liu2024). This system illustrates how combining human oversight and AI can transform legal case analysis. As shown in Figure 8.3, the core innovation of D3LM lies in its ability to generate lawyer-like diagnostic questions that guide users in formulating their legal queries more accurately. By incorporating human knowledge through targeted questioning, D3LM extracts crucial case details, which significantly improves the AI’s understanding and predictive capabilities.

Figure 8.3 Comparison of legal service methodologies, focusing on traditional LLMs, lawyer consultations, and the D3LM model

Figure 8.3Long description

The block diagram compares how a legal case is handled using three legal service methodologies: a standard LLM, a human lawyer, and the D3LM model. The process begins with an unclear fact description from the client, which initiates a professional dialog involving follow-up questions—such as inquiries about surveillance footage and intoxication. These interactions help construct a structured fact set, which serves as the basis for generating a legally sound court view. Unlike the other two approaches, the LLM attempts to generate a legal conclusion directly from the vague initial description, without engaging in clarification or fact gathering. As a result, its response is incomplete and overlooks key facts. The human lawyer integrates critical evidence and context into a professional judgment, though at a higher cost. In contrast, the D3LM model autonomously conducts dialog, extracts essential facts, and produces a court-view-level decision that is both cost-effective and legally coherent.

Unlike traditional LLMs that passively respond to user inputs, D3LM actively engages with users to gather missing or overlooked information. It utilizes the Positive-Unlabeled Reinforcement Learning (PURL) algorithm to dynamically adapt its questioning strategy, making it an example of Human-in-the-Loop systems where humans enhance AI’s performance through ongoing interaction. This collaboration results in more accurate legal predictions, enriched by human guidance.

This case highlights how human–AI collaboration can empower individuals without legal expertise to receive professional-quality assistance at a fraction of the cost. Moreover, it demonstrates the potential of integrating human knowledge with advanced AI technologies to bridge gaps in domains where precise and detailed understanding is essential.

Model Robustness and Efficiency

The continuous integration of human expertise into AI models offers a promising pathway to enhancing their robustness and efficiency but it also presents significant challenges. As AI systems undergo iterative training cycles driven by human feedback, issues such as catastrophic forgetting may arise, where models struggle to retain previously learned information when exposed to new data. Addressing this requires strategies like active learning, which ensures models selectively query the most informative data, thereby optimizing learning while maintaining robustness (Bartolo et al., Reference Bartolo, Roberts, Welbl, Riedel and Stenetorp2020; Budd et al., Reference Budd, Robinson and Kainz2021). Incorporating human rationales into these models can further bolster both performance and stability, reducing the chances of losing acquired knowledge over time (Arous et al., Reference Arous, Dolamic, Yang, Bhardwaj, Cuccu and Cudré-Mauroux2021).

Another significant challenge in continuous learning from human experts is the efficiency of data annotation, particularly for complex tasks. Manual annotation is labor-intensive and time-consuming. While innovations like self-annotation frameworks and deep reinforcement active learning have shown potential to streamline the process (Le et al., Reference Le, Sugimoto, Ono and Kawasaki2020; Z. Liu et al., Reference Liu, Wang, Gong, Lu and Tao2019), further advancements are needed to minimize the dependency on large-scale labeled data. For instance, some frameworks improve model robustness by integrating active learning with adversarial examples, which significantly reduces the need for annotations while boosting performance (Guo et al., Reference Guo, Shi, Kang, Kuang, Tang, Jiang, Sun, Wu and Zhuang2021).

Balancing Expertise and Ensuring Feedback Quality

Balancing expertise and feedback quality is another critical factor. While non-experts can contribute valuable insights in some cases, tasks requiring specialized knowledge demand domain experts for accuracy (Fan et al., Reference Fan, Li, Yuan, Dong and Liang2019). Establishing clear guidelines for differentiating tasks suitable for nonexperts from those requiring expert input is essential for improving feedback quality. Additionally, variability in annotator expertise can lead to uneven feedback, which may negatively impact model performance (Jwo et al., Reference Jwo, Lin and Lee2021; Z. Liu et al., Reference Liu, Guo and Mahmud2021). Thus, developing improved methods for evaluating and filtering human feedback is essential, particularly in complex, high-stakes domains where consistent and reliable input is critical (He et al., Reference He, Michael, Lewis and Zettlemoyer2016). For instance, techniques that employ adversarial learning and knowledge graphs have proven effective in tackling these challenges by improving feedback quality and maximizing the utility of human input, even in highly imbalanced datasets (Jiang et al., Reference Jiang, Gao, Duan, Kang, Sun, Zhang and Liu2020).

Real-World Applications and Generalization

Finally, ensuring that AI systems can generalize effectively across real-world scenarios requires robustness to domain shifts and the handling of out-of-distribution samples. Real-world environments often introduce complexities such as inconsistent data quality and ambiguous tasks (Kreutzer et al., Reference Kreutzer, Riezler and Lawrence2021; Ziegler et al., Reference Ziegler, Stiennon, Wu, Brown, Radford, Amodei, Christiano and Irving2020). Research has already begun addressing these challenges, for instance, by using inconsistency-based sample selection to guide human experts in annotating the most uncertain and inconsistent samples, thereby improving model accuracy and reducing labeling costs (Guo et al., Reference Guo, Kang, Duan, Liu, Tang, Zhang, Kuang, Sun and Wu2022). Future research should focus on developing domain-agnostic models that adapt flexibly to diverse scenarios without extensive retraining (Xu et al., Reference Xu, Dainoff, Ge and Gao2023; Ziegler et al., Reference Ziegler, Stiennon, Wu, Brown, Radford, Amodei, Christiano and Irving2020), enhancing the scalability and applicability of human–AI collaborative systems across various domains.