What can I do with you, machine?

To begin the article, the authors have researched that the concept of management evolved through three stages:

• Classical theories: depicted managers as authority figures responsible for planning and organizing their subordinates’ work. Contributed by Henri Fayol and Frederick Winslow Taylor.

• Neo-classical theories: subsequently shifted focus toward social interaction, thereby emphasizing managers’ role in leading and controlling workers. Contributed by Chester Barnard and Mary Parker Follett.

• Modern era systems perspective: recognized the managerial role of knowledge workers who, although lacking formal authority, make decisions that impact an organization’s performance significantly. Contributed by Peter Drucker and Peter Senge.

Integrating these perspectives, the authors brought up figures such as organizational theorist Richard L. Daft, who defined management as "the attainment of organizational goals in an effective and efficient manner through planning, organizing, leading, controlling, and [deciding on] organizational resources." He later writes that management occurs in an organization, described as "a social entity that is goal-directed and deliberately structured," and applies "to all organizations, including both profit and non-profit ones."

On the other hand, consistent with the systems perspective, Austrian American consultant and educator Peter Drucker believes that management is not confined to those with "command over people" but extends to everyone bearing "responsibility for contribution".  That while managers impact others hierarchically through their positional authority, knowledge workers do so laterally through their knowledge authority. Consequently, this stance on management extends beyond business managers to include professionals (e.g., doctors and lawyers).

In 1967, Drucker anticipated that technology would promote the decentralization of management. To quote his words: "With the computers taking over computation, people all the way down in the organization will have to learn to be executives and to make effective decisions."

Subsequent research provided evidence of the way technology enables managerial authority to be delegated to individual experts throughout an organization. Professor Nicolai J. Foss, for instance, documented how Danish hearing aid manufacturer Oticon’s adoption of advanced information technology to coordinate plans and actions facilitated managerial work’s decentralization.

While there are more examples referenced in the article, the example above should provide a collective consensus on the evolution of management.

Defining AI in a multi-lens, scholarly manner, a prominent figure the authors referenced was John McCarthy, one of the founders of AI as a discipline. He described AI in terms of "making a machine behave in ways that would be called intelligent if a human were so behaving." In this tradition, AI is not a specific technology, but "a moving frontier of next-generation advancements in computing."

In the late 1980s, for example, rule-based automation was introduced into expert systems providing managers with decision support. These early systems are deterministic since they follow pre-programmed rules to perform repetitive tasks, so University of Notre Dame professor in IT (and more) Nicholas Berente et al. concluded that "expert systems were widely considered a type of AI then, but most would not consider it a type of AI today."

Contemporary AI differs from prior technology vintages by its ability to learn and adjust its behavior on the basis of the data to which it is exposed. Machine learning enables algorithms to learn from experience without the need for explicit human programming.

For example, organizations use deep learning based on artificial neural networks, a type of machine learning that distinguishes itself from earlier technologies by its exceptional capacity for autonomous learning and action. This capacity is crucial for management, which, unlike rule-based automation, cannot just carry out orders, but requires autonomy over the direction, the content, and the quality of the work or the methods of its performance.

Unlike earlier technologies, contemporary AI is distinguished by its ability to make the rules and to learn, adapt, and act autonomously. To the authors, we refer to contemporary AI applications that have the capacity to learn and can, therefore, improve and adapt based on experience.

In practice, contemporary AI evolved in two ways:

• Predictive AI: Learns patterns from existing data in order to anticipate specific task domains’ outcomes. Take capital investment’s future returns as an example.

• Generative AI: Create new data based on learned patterns, which organizations apply across a wide range of task domains. For example, generative AI helps managers develop strategies, solve problems, and provide employee feedback.

Collectively, predictive and generative AI have the transformational potential to create intelligence as a general systems property, rather than a specific human attribute.

However, even contemporary AI systems fail to provide 'strong' AI or artificial general intelligence, both of which refer to futuristic visions of algorithms able to perform all tasks just as well as humans if not even better. Contemporary AI lacks consciousness or intrinsic motivation, therefore requiring humans to set its goals. Moreover, managerial tasks’ computational complexity means that AI can only produce outputs that approximate reality.

With recognition to the human desire of familiarity, human intuition and judgment are essential to align AI outputs with practical realities. Furthermore, data constraints can degrade AI performance, yielding biased outcomes or hindering AI application entirely. Owing to these challenges, humans remain involved in managerial processes and interact with machines on a wide range of tasks… for now.

In recent years, scholars have begun to expand traditional management theories from humans onto AI algorithms. Although this emerging body of research lacks a unified definition of managing with AI, it is possible to discern common elements across studies. Collectively, these studies focus on algorithms performing managerial tasks, thereby diverging from traditional management theories’ exclusive attention to humans undertaking these tasks.

The core analytical focus is on humans’ interactions with algorithms, whereas traditional management theories examine interactions between humans. Like traditional management theories, research on managing with AI occurs within organizations. In synthesizing this work, the authors defined managing with AI as humans’ interaction with algorithms performing managerial tasks in organizations. This definition includes all core managerial tasks — planning, organizing, leading, controlling, and deciding.

A task is managerial if it substantially affects an organization’s ability to attain its goals effectively and efficiently. Furthermore, this definition encompasses all humans who interact with AI, including managers and knowledge workers who use algorithms to manage, as well as employees and workers whom algorithms manage. These algorithms have the capacity to learn, adapt, and act autonomously, which allows them to perform managerial tasks.

Lastly, the author's definition applies to human-algorithm interactions in all types of organizations, whether they are business firms with profit goals or public entities with nonprofit goals.

The human-AI collaboration (HAIC) and algorithmic management (AM) perspectives provided the authors diverging conceptualizations of managing with AI along four key dimensions:

• Context: the organizational tasks and relationships in which managing with AI occurs.

• Agency: a temporally embedded capacity to act with intent, which is assigned to both humans and AI algorithms.

• Interaction: a relational exchange that takes place between humans and algorithms.

• Outcome: the impact of AI implementation.

HAIC research portrays executive decision-making in an enabling context, emphasizing retained human agency, augmented interactions, and improved task performance. HAIC scholars draw on conceptual roots in the behavioral theory of the firm and on foundational human-AI collaboration research, both of which depict decision-making as the key managerial task.

Consistent with these roots, HAIC scholars explore executive decision-making contexts. They define decision-making as "the process of selecting the alternative that is expected to result in the most preferred outcome." Such structured decisions are the crux of most empirical HAIC studies, although there is emerging interest in AI’s use for unstructured decisions requiring problem solving.

Drawing on behavioral theory foundations, HAIC scholars assume that executives' bounded rationality motivates human-AI collaboration in decision-making. AI algorithms have information processing skills that differ from those of humans, helping organizations overcome some of their traditional cognitive limitations. AI algorithms support executive decision-making with their superior prediction skills, which reduce cognitive biases, ensure greater consistency, and provide greater speed and efficiency.

Another AI benefit is its ability to analyze extensive datasets and identify hidden patterns. In turn, HAIC scholars assume that humans outperform AI in terms of judgment skills based on their intuition, contextual understanding, and creativity. 

HAIC research explores executive decision-making embedded in traditional organizational settings, such as multinational enterprises and public institutions, providing an enabling context for human-AI collaboration. Specifically, scholars analyze the decision-making structures in which humans and AI collaborate. These structures define how decision tasks are divided between humans and AI, as well as how these agents’ activities are integrated.

Conceptual research has also further distinguished between sequential and interactive human-AI decision-making structures. While sequential structures assign different decision tasks to humans and AI, interactive structures allow humans and AI to work jointly on the same decision task. Empirical studies provide evidence of sequential and interactive structures’ use in management practice. A few HAIC studies also describe organizational-level structures that enable human-AI collaboration, such as knowledge sharing processes, high performance work systems, and corporate technology centers.

HAIC researchers focus on first-party actors who collaborate with AI, comprising managers but also knowledge workers such as consultants, controllers, accountants, physicians, and traders. Moreover, HAIC research on public management investigates civil servants, while marketing scholars analyze sales experts. A few scholars also examine data scientists and system developers interacting with the AI algorithms they develop.

Meanwhile, HAIC scholars focus on first-party actors with the expertise to complement AI. Most scholars highlight these actors’ domain expertise or the "skills and knowledge accumulated through prior learning within a domain." Such domain expertise helps mitigate AI biases by allowing humans to use their tacit knowledge, which algorithms do not have, to correct or overrule AI's predictions. More on domain expertise, it allows humans to complement AI's outputs. A few scholars also explore AI expertise, or the "skills and knowledge accumulated through prior familiarity of tasks with the technology."

HAIC research also suggests that first-party actors' ability to retain their agency when collaborating with AI is contingent on their levels of domain expertise. Novices with little expertise lack the absorptive capacity to collaborate effectively with AI. For example, Anthony observes that a group of young, inexperienced knowledge workers forfeited their agency to AI algorithms by "taking them for granted without understanding them."

Higher levels of expertise allow first-party actors to probe AI algorithms more effectively. More experienced first-party actors can retain their agency and sideline AI by maintaining their work routines or by ignoring AI when they feel its advice is not helpful or is incorrect.

HAIC scholars conceptualize humans' interactions with algorithms as augmentation, defined as humans and AI collaborating to undertake a managerial task. Augmentation is based on a complementary relationship arising from humans' and AI's different skills. Complementarities emerge from humans' and AI’s heterogeneous (diverse in character) cognitive skills and AI's cognitive skills and humans' social skills, which, for example, allow humans' emotions to be considered when making decisions or when communicating decision outcomes to humans.

Conceptual HAIC studies assume that executive decision-making problems require augmentation. While AI algorithms handle levels of complexity that quickly overwhelm boundedly rational humans, executive decision-making problems also "encapsulate an element of uncertainty […] rendering human input indispensable." Consistent with these assumptions, empirical HAIC studies suggest that uncertainty may reduce AI's predictive accuracy, which manifests in biases associated with AI's data inputs and outputs. Furthermore, complexity can mean that humans find AI algorithms' inputs, processing, and outputs opaque.

The HAIC literature describes various engagement practices that first-party actors use to address these problems. Engagement practices refer to interactions during which these actors integrate "AI knowledge claims with their own." Prior research distinguishes three types of engagement practices:

• Envelopment of AI inputs: refers to first-party actors' involvement with choosing and curating training data, as well as with selecting, tuning, and governing AI algorithms.

• Auditing of AI processing: Describes practices that examine, validate, and correct an algorithm’s processing. It and envelopment can contribute to AI quality by reducing data and processing biases.

• Translation of AI outputs: Practices that make AI outputs comprehensible and actionable for users, which reduce AI’s opacity.

Collectively, prior HAIC studies provide rich evidence of engagement practices, but some scholars also observe first-party actors’ disengagement practices, such as regularly ignoring AI’s input or accepting it without much reflection.

HAIC research also provides insight into the drivers of first-party actors’ engagement or disengagement. The author's references find that disengagement in the form of overreliance on AI is due to managers’ strong perceptions of AI’s quality, which make them trust AI blindly. Relatedly, managing expectations about AI’s quality leads to first-party actors’ greater engagement. Moreover, first-party actors often respond to AI opacity with disengagement practices, but those who engage subsequently create increasing transparency.

With regard to expertise, HAIC research indicates that experts with the intuition and tacit knowledge to complement AI, ironically, tend to be the first-party actors least likely to engage with AI, due to their algorithm aversion. On the other end of the spectrum, novices with little experience tend to follow AI blindly due to their cognitive overload. Lastly, AI expertise is positively related to first-party actors’ willingness to engage with AI.

In terms of outcome, HAIC researchers are primarily interested in first-party actors’ task performance. They explore their effectiveness, such as service quality, decision accuracy, and customer satisfaction, as well as their efficiency, such as task duration. Empirical studies comparing the task performance of humans collaborating with AI to that of humans without AI arrive at inconclusive findings, reporting increased or reduced task performance, as well as mixed findings.

The HAIC literature also identifies supportive conditions enabling first-party actors to perform well when interacting with AI. These conditions are linked to the context, agency, and interaction dimensions. In terms of the context, granting first-party actors' discretionary power to override algorithmic decisions reduces their overall task performance but increases it in terms of highly complex decision-making tasks. Furthermore, the authors have found evidence suggesting that decision-making structures enabling direct human-AI interaction are more beneficial than those only allowing indirect interaction. A sequential process during which AI delegates tasks to humans is more effective than one where humans refer tasks to AI, which is due to humans’ lack of sufficient meta-knowledge for making appropriate delegation decisions.

Regarding agency, the HAIC literature reports mixed findings. More experienced first-party actors tend to be better at discarding inaccurate AI advice but also often ignore correct AI advice, which can reduce their task performance. Conversely, another finding by the authors shows that both novices and experts increase their task performance by collaborating with AI. Other reported findings indicate that a combination of domain expertise and AI skills is required for knowledge workers to perform well when interacting with AI on innovation tasks.

The authors believe that there is sufficient evidence that engagement leads to a higher task performance, whereas disengagement reduces it when compared to humans working without AI. In addition, scholars show that first-party actors perceived higher AI quality in human-AI interactions increases their task performance. Additionally, research reports that first-party actors perceived higher AI trustworthiness is associated with a better task performance. Nevertheless, increased explainability does not necessarily lead to a better performance in human-AI collaboration.

AM studies depict managerial control in a coercive context, in which human agency is restricted, interactions are automated, and individuals are personally affected. For comparison, while HAIC research describes executive decision-making, AM literature explores human-AI interaction in the context of algorithmic control, which it defines as "efforts to align worker behavior with organizational goals".

AM studies’ main conceptual roots lie in labor process theory, which depicts work relations as a 'structured antagonism', perceiving everything in terms of conflict and managerial efforts to control workers more completely. Furthermore, AM studies have roots in research on technology use to control workers. Consistent with these roots, AM research portrays algorithmic systems as 'contested instruments of control.'

The AM literature focuses on two main managerial control mechanisms:

• Algorithmic monitoring: Combines collecting real-time data on workers’ behavior with comparing this data with the expected behavior, which serve to evaluate workers’ job performance. These outputs further enable algorithmic feedback (see below).

• Algorithmic feedback: text. Uses disciplinary means (i.e., fines and suspension) to sanction undesired behaviors, and enabling means (i.e., coaching and rewarding) to incentivize desired behaviors.

An example of algorithmic monitoring and feedback found by the authors, written by Indian professors Shalini Parth and Dharma Raju Bathini show how a ride-hailing platform uses subtle, real-time behavioral nudges in text messages and other notifications to influence drivers’ behavior and align it with the organizational goals.

Whereas HAIC research explores executive decision-making in the enabling contexts that traditional organizations provide, the AM literature primarily investigates managerial control in digital platforms’ coercive context. In these organizations, the algorithmic control systems’ information processing is invisible to the workers, which is due to platforms’ deliberate efforts to conceal their AI systems’ inner workings and the absence of dialogue and explanation regarding the systems.

Once again referencing HAIC research, while it focuses on first-party actors who use AI algorithms to manage, AM research describes second-party actors who are managed by these algorithms. These are often legally independent workers, such as ride-hailing drivers, food delivery couriers, and crowd workers. Adding to this list, marketing scholars are interested in sales agents and customers, public management scholars focus on citizens, and HR scholars attend to job candidates and employees subject to algorithmic control.

Actors’ domain expertise plays a more marginal role in AM research than HAIC studies. The former's focus is on their vulnerability which can be due to second-party actors’ socio-economic status, employment conditions, and gender. In terms of status, platform workers’ often low education levels and temporary residence status make these actors vulnerable to algorithmic control.

For more real-life examples, precarious employment conditions, such as dependence on a platform for work and income, reliance on the state for social welfare, independent contractor status, early-stage employment, and job insecurity are some of the many sources to second-party actors’ vulnerability. Furthermore, research on gender differences showed the authors that fearing discrimination by human evaluators, who are perceived as favoring men, leads female second-party actors to prefer algorithmic over human evaluation.

While HAIC scholars contend that humans retain their agency, AM scholars assume that AI algorithms 'constrain' human agency. This assumption reflects the AM literature’s roots in labor process theory, which describes workers’ lack of agency in employment relations. Accordingly, AM scholars portray workers as 'prisoners'. While companies justify the use of algorithmic control as enabling workers to become entrepreneurs and enjoy greater freedom than they would have in traditional work settings, these practices actually constrain workers’ already limited agency even further.

Their constrained agency reduces second-party actors’ abilities to navigate their roles, develop personal competencies, and engage in workplace activities. One particular study finds that citizens interacting with a state-run algorithmic system have little opportunity to report problems with the system or to object to its decisions. Accordingly, some conceptual AM studies envision a dystopian future of behavioral control that "gains steam until humans lose agency", effectively resulting in the 'end of choice' when humans’ agency is muted entirely.

Some empirical AM studies also suggest that second-party actors are, under certain conditions, able to reduce these algorithmic constraints and reclaim agency. However, scholars emphasize that second-party actors’ ability to reclaim agency decreases over the labor process’s subsequent stages and when companies take retaliatory measures, while continuous changes made to algorithmic systems constrain this ability even further. Furthermore, AM scholars observe that platform workers’ vulnerability and dependence on a platform reduce their ability to reclaim agency.

While the HAIC literature conceptualizes human-AI interaction as augmentation, the AM literature describes it as automation (i.e., replacing humans with AI algorithms in a managerial task). One example even stated that managing activities are transferred from humans to sophisticated algorithmic technological systems and management is no longer a human practice, but a process embedded into technology.

More sources defined automated interactions as the "absence of hierarchical reporting relationships", where discretionary feedback, reviews and ratings are calculated, interpreted and rendered actionable by largely inscrutable and opaque processes of automation.

Similar to HAIC scholars, AM scholars identify AI opacity as a defining characteristic for humans’ interaction with algorithms. However, while said scholars understand opacity as a starting condition, which can partially be overcome by first-party actors’ interactions with AI, they describe opacity as an AI characteristic that organizations implement strategically or accept willingly in order to reinforce information asymmetries and to control their workers.

AI opacity creates uncertainty for second-party actors regarding their performance’s measurement and the consequences of not following instructions, thereby eliciting their compliance. Furthermore, AI opacity constrains these actors’ learning.

Continuing forward, while HAIC research describes first-party actors’ engagement, AM research focuses on second-party actors’ acceptance. Acceptance practices allow workers to adapt their behavior to comply with algorithmic control. A group of researchers observed an online talent marketplace’s distinctive acceptance practice, which include staying under the radar (i.e., refraining from voicing concerns), purposefully curtailing outreach (i.e., avoiding difficult clients who could give bad ratings), keeping emotions in check (i.e., suppressing negative ones), and undervaluing work (i.e., lowering hourly rates to improve clients’ ratings).

AM studies describe acceptance as the dominant response to algorithmic control but also provide evidence of resistance, which refers to practices allowing workers to block or bypass algorithmic functions by, for example, logging off strategically, refusing or canceling requests, ignoring instructions, using wrong identities, and communicating outside the algorithmic system.

AM research also offers insight into the conditions under which second-party actors accept or resist algorithmic control. This work focuses on said actors’ perceptions of AI’s legitimacy, quality, and trustworthiness. AI algorithms are perceived as legitimate when their actions are desirable, proper, or appropriate within some socially constructed system of norms, values, beliefs, and definitions. Second-party actors’ positive AI legitimacy perceptions are associated with acceptance, whereas their negative ones are linked to resistance. Besides legitimacy, low perceptions of AI quality trigger resistance through gaming and workaround tactics.

AM scholars also explore second-party actors’ perception of AI’s trustworthiness, defined as their expectation that algorithms are sufficiently competent and benevolent to rely on them. AI trustworthiness is associated with acceptance of algorithmic monitoring and feedback

While HAIC studies focus on task performance outcomes, AM studies are primarily interested in human-AI interactions’ personal impact on second-party actors. These studies focus on job-related, career-related, and psychological impacts. Job-related impacts include second-party actors’ job satisfaction, commitment, and engagement; career-related impacts refer to these actors’ capability development and career success, feelings of social isolation and diminished well-being.

The AM literature focuses on algorithmic control’s negative personal impact. Scholars of this literature have, for example, found that algorithmic control practices’ adoption reduces second-party actors’ employee commitment due to their perception of AI algorithms as reductionist, unfair, and impersonal. However, a few AM studies have also reported a neutral or positive personal impact. Owing, for example, to AI’s greater ability to provide personalized attention and feedback, the adoption of algorithmic control practices can lead to similar or increased levels of employee commitment and perceived job satisfaction.

Whereas HAIC studies identify supportive conditions for first-party actors’ task performance, AM researchers focus on containing conditions that reduce algorithmic control’s negative impact along the context, agency, and interaction dimensions.

Regarding the context, a study saw that enabling AI control practices (i.e., advice) increases employee commitment, whereas constraining AI control practices (i.e., sanctions) decreases it. Another observation detailed that first-party actors in highly skilled jobs generally experience enabling AI contexts, which allow them to build new capabilities, whereas second-party actors in low-skilled jobs face constraining AI contexts that hinder their capability’s development.

In terms of agency, the authors brought up that established platform workers experience less anxiety than newcomers, who are more vulnerable due to their lack of positive reviews. In conjunction, they also found that their lower socio-economic status and education levels make platform workers more prone to social isolation.

On the positive side, the authors dug up a report about how algorithmic feedback enables second-party actors’ capability development; this increases even more when human feedback is added to the algorithmic feedback to reduce these actors’ algorithm aversion and cognitive overload, which would otherwise obstruct their learning.

Furthermore, a research paper said that greater AI transparency can reduce workers’ negative emotions, such as anxiety, and help them develop their competencies, both of which enable them to pursue new career opportunities.

With respect to interaction, studies associate second-party actors’ perceived AI legitimacy with their job satisfaction. For example, a study described how workers’ AI legitimacy perception mediates algorithmic practices’ negative effect on their commitment. Relatedly, another study found that AI legitimacy has a direct positive effect on employee commitment. Lastly, it has been demonstrated that if second-party actors perceive AI’s legitimacy and trustworthiness as strong, they experience fewer negative emotions.

The authors' initial review suggested that HAIC and AM analyzed different parts of managing with AI. This practice led to fragmented and partly contradictory micro-level perspectives, they self-admittedly. Such micro-level views are reductionist since organizational phenomena cannot be fully understood by analyzing their isolated parts, fail to explain how these parts interact at a higher level, and cannot capture the collective behaviors and properties emerging from these interactions.

The authors therefore drew on systems theory to reanalyze their sample with a focus on making latent linkages between HAIC and AM perspectives visible. Specifically, they used key systems concepts — hierarchy, interconnectivity, emergence, and scale — to develop building blocks that bridge the two perspectives. This integrative framework redirected scholarly attention from tasks to the organizational context, from individual to collective agency, from local to systemic interaction, and from micro- to multi-level outcomes.

A paradigm is the pattern of thinking, and a paradigm shift is a fundamental change in approach or underlying assumptions. In simpler terms, it means a major change in how people think and get things done that upends and replaces a prior paradigm.

In 2023, the emergence of pre-trained models such as GPT — on which the artificial intelligence (AI) system ChatGPT is based — has created a paradigm shift in AI, transforming it from an application to a general-purpose technology that can be configured to specific uses. Whereas AI models were trained to do one specific task well, it is now usable for a variety of tasks (e.g., chatting, decision making, item generation) potentially unrelated to one another without explicit training. In short, AI applications are slowly becoming applicable to anything involving intelligence.

One of the pioneers of AI who coined the term itself, Herbert Simon, described it as a 'science of the artificial'. In contrast to natural sciences, which describes the existing world, the science of the artificial is aims to create machine intelligences instead. According to Simon, this makes AI a science of design and engineering.

More on the topic of science, the article proposes that the evolution of AI can be understood through Thomas Kuhn's theory of scientific progress in terms of paradigm shifts, which describes science as a process involving occasional 'revolutions' stemming from crises faced by dominant theories, followed by periods of 'normal science' where details of the new paradigm are fleshed out.

Over time, as the dominant paradigm fails to address an increasing number of important anomalies or challenges, the center of gravity of the field shifts toward a new set of theories and methods — a new way of thinking that might better address them —even as the old paradigm may co-exist. An example of this from the author is that Newtonian physics is still used and taught, despite its breakdown when it comes to subatomic particles or objects travelling at the speed of light. 

To further understand the state-of-the-art of AI and its current developmental direction, the author proposed tracking its scientific history, notably the bottlenecks that stalled progress in each paradigm and the degree to which they were addressed every paradigm shift. Here are the AI paradigms that the author provided on, from expert learning to modern deep learning:

• AI paradigm in late 1950s: dominated by game-playing search algorithms that led to novel ways for searching various kinds of graph structures, which provided limited insight into the concept of intelligence (likely general, not just machine).

• AI paradigm across mid-1960s to late 1980s: expert systems progressed AI by representing domain expertise and intuition in the form of explicit rules and relationships that could be invoked in the form of an inference mechanism, albeit manually by a knowledge engineer.

  • Example: INTERNIST-1 was an early successful demo of AI in medicine that performed diagnosis in the field of internal medicine and represented expert knowledge using casual graphs and hierarchies relating diseases to symptoms.

  • Hurdle: the 'knowledge engineering bottleneck' is a consequence from the challenging task of extracting reliable knowledge from human experts — for human reasoning and language is too complex and heterogenous to be captured by top-down specification of relationships. Researchers also found that expert systems would often make errors in common-sense reasoning, which seemed to intertwine with specialized knowledge.

• Last AI paradigm in late 1980s and 1990s: the breakthrough that breached the key barrier that prevents progress is machine learning through self-supervision, which is what contributed to transforming AI into a more general-purpose technology. Instead of being a passive repository of knowledge, the machine became an active 'what if' explorer, capable of asking and evaluating its own questions. This enabled data-driven scientific discovery, where knowledge was created from data.

  • Argument: the author argues that the epistemic criterion in ML for something to qualify as knowledge was accurate prediction. This criterion conforms to Austro-Hungarian philosopher Karl Popper’s view that the predictive power of theories is a good measure of their strength — that theories seeking only to explain a phenomenon were weaker than those making bold ex ante (forecast) predictions that were objectively falsifiable.

  • Exemplar: the exemplar — the community's consensus on the best examples of scientific research in a field — in this paradigm was largely shaped by logic and the existing models of cognition from psychology, which viewed humans as having a long-term and short-term memory, and a mechanism for evoking them in a specific context.

  • Hurdle: what was needed was the machine ability to directly deal with raw data emanating from the real world, instead of relying on humans to perform the often-difficult translation of feature engineering. Dubbed the feature engineering bottleneck, machines needed to directly ingest raw data such as numbers, images, notes, or sounds, ideally without curation by humans.

• Most recent AI paradigm: deep learning solved the last AI paradigm's feature engineering bottleneck by engineering machines to directly consume raw input in the form of images, language, and sound like humans do. Advances in hardware for parallel processing (and thus parallel learning) were critical in making deep neural networks (DNNs) feasible at scale.

  • Exemplar: the exemplar in deep learning is a multi-level neural network architecture. Adjusting the weights between neurons makes it a universal function approximator, where the machine can approximate any function, regardless of its complexity, to an acceptable degree of precision. In a sense, the architecture enables the successive refinement of input data by implicitly preprocessing it in stages as it passes through different layers of the neural network to its output.

  • Example: the above is applied to language learning models (LLMs) from which systems like ChatGPT are built. LLMs learn the implicit relationships among things in the world from large amounts of text, then they use a special configuration of the transformer neural architecture that represents language as a contextualized sequence, where context is represented by estimating dependencies between each pair of the input sequence.

  • Hurdle: the learning process of pre-trained DNNs is unclear — a pseudo-black box. Whereas expert systems store specified relationships in self-contained chunks, DNNs' are smeared across their countless weights and are harder to interpret. Another social key barrier believed to be in this paradigm by the author is lack of trust for the widespread adoption of AI.

What the author considers most important is the shift from building specialized applications of AI to one where knowledge and intelligence do not have specific boundaries but transfer seamlessly across applications and to novel situations. Regardless of whether modern AI systems satisfy the Turing Test for intelligence, the debate itself is considered indicative of the significant progress in AI and the palpable increase in its capability.

The upcoming AI paradigm after deep learning is general intelligence (GI). Brining up an example of a pneumonia-predicting AI in a hospital, the performances of previous AI applications would not necessarily transfer to use on patients in other locales. GI, on the other hand, is about the ability to integrate knowledge about pneumonia with other diseases, conditions, geographies, and so on from all available information, and to apply the knowledge to unforeseen situations.

More broadly, GI refers to an integrated set of essential mental skills — which include verbal ability; reasoning; and spatial, numerical, mechanical, common-sense, and reasoning abilities — which underpin performance across all mental tasks. A machine with GI maintains such knowledge in a way that is easily transferrable across tasks and can be applied to novel situations.

While discussing the development towards GI, each paradigm shift moved AI closer to GI and greater expanded the scope of applications from the former. To list each paradigm starting from expert systems:

• Expert systems structured human knowledge to solve complex problems. 

• Machine learning brought structured databases to life.

• Deep learning went further, enabling the machine to deal with structured and unstructured data about an application directly from the real world, as humans are able to do. 

• Recent advances in DNNs have made it possible to create entirely new forms of output, previously assumed to be the purview of humans alone, which has greatly expanded the scope of AI into the creative arena in the arts, science, and business.

At the moment, there are no obvious limits to these dimensions in the development of GI. On the data front, for example, in addition to language data that humans will continue to generate on the Internet, other modalities of data, such as video, are now becoming more widely available. Indeed, a fertile area of research right now is how machines will integrate data from across multiple sensory modalities — such as vision, sound, touch, and smell — just as humans can do.

We are currently in a transitional phase where deep learning remains the dominant paradigm, but GI is rapidly emerging as a transformative frontier thanks to advancements in the former. This transition would see the continued improvement of pre-trained models and GI with increases in volume and variety of data and computing power. However, this should not distract us from the fact that several fundamental aspects of intelligence are still mysterious and unlikely to be answered solely by making existing models larger and more complex.

Nevertheless, it is worth noting that the DNN exemplar of GI has been adopted by a number of disciplines which seek to explain the concept of intelligence itself. This has arguably made AI more interdisciplinary by unifying its engineering and design principles with these perspectives.

In relation to intelligence, within our current paradigm, explaining and understanding the behavior of DNNs in terms of a set of core principles of its underlying disciplines is an active area of research.

The progression toward general intelligence has followed a path of increasing scope in machine intelligence from earliest to latest: learn from humans, learn from curated data, learn from any kind of data, and learn from any kind of data about the world in a way that transfers to novel situations.

In correlation to other technologies, great strides made in both sensor and servo technologies have led to concomitant leaps forward in the field of robotics, including autonomous vehicles. Should this trend continue, AI will be central to the development of the mobile artificial life that organic life creates, which could well become a more resilient life form than its biological creator.

We are currently at the early stages of fleshing out the details in the next paradigm (after GI). Despite our current limitations, pre-trained AI models have unleashed applications in language and vision, spreading out to cover support services and embedding in a broad range of industries and functions.

Economists use the term general purpose technology, of which electricity and the Internet are examples, as a new method for producing and inventing that is important enough to have a protracted aggregate economic impact — a long-lasting and widespread effect on the overall economy — across the economy. Manuel Trajtenberg and Timothy Francis Bresnahan Bresnahan and Trajtenburg describe general purpose technologies in terms of three defining properties:

• Pervasiveness: The technology is used as inputs by many downstream sectors.

• Inherent potential for technical improvements: The technology exhibits a foundational capacity for continuous and significant technical advancements, enabling sustained evolution in performance and functionality over time.

• Innovational complementarities: The productivity of R&D in downstream sectors multiplies as a consequence of innovation in the general-purpose technology, creating productivity gains throughout the economy.

The author questioned how well the GI paradigm of AI meets these criteria before stepping back to point out that AI is already pervasive, embedded increasingly in applications without human realization.

For example, with the new high-bandwidth human-machine interfaces enabled by conversational AI, the author presumes that the quality and volume of training data that machines such as ChatGPT can now acquire as they operate is unprecedented. Sensory data from video and other sources will likely continue to lead to improvements in pre-trained models and their downstream applications.

The last of the three properties — innovation complementarities — may take time to play out at the level of the economy. With previous technologies, such as electricity and IT, growth rates were below those attained in the decades immediately preceding their arrival.

Borrowing the words of American academic Erik Brynjolfsson et al., substantial complementary investments were required to realize the benefits of general-purpose technologies, where productivity emerges after a significant lag. With electricity, for example, it took decades for society to realize its benefits since motors needed to be replaced, factories needed redesigns, and workforces needed to be reskilled. IT was similar, as was the Internet, and so likely will AI.

In fact, AI is similarly in its early stages, where businesses and governments are scrambling to reorganize processes and rethinking the future of work. Just as electricity required the creation of an electric grid and the redesign of factories, AI will similarly create an 'intelligence grid' requiring perhaps multiple redesigns of processes to realize productivity gains from this new general-purpose technology. Such improvements can take time to play out.

The author conclusively argues that we should not assume that we have converged on the 'right paradigm' for AI as the current one will undoubtedly accommodate others that address its shortcomings. History has proven that paradigm shifts do not always improve on previous paradigms in every way, especially in their early stages, and the current paradigm is no exception. New theories often face resistance and challenges initially, while their details are being filled in.

For example, the deep-learning paradigm is far superior at perception relative to earlier paradigms, although it is poorer in terms of explanation and transparency.

Indeed, despite the current optimism about AI, which has been fueled by its success at prediction, the author emphasizes that the current paradigm faces serious challenges as AI permeates different aspects of our lives — one of the biggest of challenges being one of trust.

This challenge stems in large part from its representation of knowledge that is opaque to humans (aka the fear of the unknown). For example, a recent exciting application of AI is in the accurate prediction of protein-folding and function that is essential for drug development and disease intervention. However, the behavior of such machines is still sensitive to how the input, such as a molecule, is represented, and they struggle to explain the underlying mechanisms of action.

While a fertile area of research in the current paradigm, the current opacity of these systems hinders human understanding of the natural world, which places a high premium on explanation and transparency.

Likewise, even though systems such as ChatGPT can be trained on orders of magnitude more cases than a human expert could possibly encounter in a lifetime, their ability to introspect from multiple viewpoints is severely limited relative to humans. A related problem proposed by the author is that we can never be sure they are correct, and not 'hallucinating,' that is, filling in their knowledge gaps with answers that look credible but are incorrect. It is like talking to someone intelligent whom you cannot always trust.

The author had listed the following problems that AI creators will need to be addressed if humanity is to trust AI. Since the data for pre-trained models is not curated, machines pick up on the falsehoods, biases, and noise in their training data. Systems using LLMs can also be unpredictable, and systems based on them can exhibit undesirable social behavior that their designers did not intend. While designers might take great care to prohibit undesirable behavior via training using something like Reinforcement Learning via Human Feedback (RLHF), such guardrails do not always work as intended. Like some parts of its creators, the machine is relatively inscrutable.

Making AI explainable and truthful is another big challenge; it is not obvious whether this problem is addressable solely by the existing paradigm. The unpredictability of AI systems built on pre-trained models introduces new problems for trust. For example, the output of LLM-based AI systems on the same input can vary, a behavior we associate with humans but not machines. On the contrary, we expect machines to be deterministic, not 'noisy' or inconsistent like humans. Until now, we have expected consistency from machines.

While we might consider the machine’s variance in decision making as an indication of creativity, a human-like behavior, it poses severe risks — especially given its inscrutability and an uncanny ability to mimic humans. Machines are already able to create 'deep fakes' or fictional movies indistinguishable from human creations. We are still seeing the emergence of things (e.g., AI-generated pornography) made using real humans and their creations (i.e., art, documents, and personas) whose basis is hard to identify.

Likewise, it is exceedingly difficult to detect or even define plagiarism or intellectual property theft, given the large corpus of public information on which LLMs have been trained. Existing laws are not designed to address such problems, and they will need to be expanded upon to recognize and deal with them.

Finally, inscrutability also creates a larger and existential risk to humanity, which could become a crisis for the current paradigm. For example, in trying to achieve goals that we give the AI, such as 'save the planet,' we have no idea about the sub-goals the machine will create to achieve its larger goals. Unlike previous AI machines, which were designed for a specific purpose, general intelligence has emerged without such a purpose, which makes it impossible to know whether the actions of systems built on it will match our intentions in a particular situation.

This is known as the 'alignment problem,' which arises from our inability to determine whether the machine’s hidden goals are aligned with ours. In saving the planet, for example, the AI might determine that humans pose the greatest risk to Earth’s survival due to pollution or nuclear war, and hence they should be contained or eliminated. This is especially concerning if we do not know the data the machine is using to base its decisions, and the credibility of the data used.