Artificial intelligence and counseling: Four levels of implementation

Artificial intelligence: Description and explanation

Understanding how AI has and will impact the counseling profession begins with establishing reliable definitions. Breaking the term down into its component parts means defining the terms “artificial” and “intelligence.” Artificial implies the synthetic or human-designed rather than the naturally derived. The “artificial” of AI involves mechanics, electronics, or computers. The concept of intelligence—specifically defining and measuring it as a variable, combined with its connotations—has been long debated in the literature (Cherniss, Extein, Goleman, & Weissberg, 2006; Davies, 2002; Fagan, 2000; Schroeder, 2017; Sternberg, 1985). The confusion also exists within the AI community (Legg & Hutter, 2007).

Intelligence is thought to extend beyond a strict cognitive capacity into the emotional realm (Goleman, 2005) and is theorized to have multiple extensions (Gardner, 2006). A useful synthesis of the myriad conceptions of intelligence is offered by artificial intelligence researcher Max Tegmark (2017), who states that intelligence is the “ability to accomplish complex goals” (p. 39). Subsequently, I offer the following definition of AI as the ability of non-biological mechanisms to accomplish goals. The qualifier “complex” is deleted from Tegmark’s definition because intelligence is not a dichotomous concept; rather, both simple and complex goals can be attained. Intelligence in its rudimentary or advanced states occupies different points on a continuum, encapsulated within the same category, differing quantitatively. AI is akin to an operating system, like the human brain. Indeed, neuroscience has informed a substantial portion of prevailing AI research (Hassabis, Kumaran, Summerfield, & Botvinick, 2017; Lecun, Bengio, & Hinton, 2015). The embodiment of AI can take various forms, from a computer screen avatar to a robot.

Machine learning and algorithms

Artificial intelligence brings big-picture, philosophical ramifications, raising ontological and epistemological questions (Copeland, 1998). Yet, AI begins within the purview of the diminutive and precise, requiring mathematics and formal logic as demonstrated by the AI subfield of machine learning. For an AI to progress to the level of a functioning counselor, it must have the capacity to learn. Machines that learn are paradigmatically dissimilar from their traditional predecessors. A major point of divergence is in the agency to possess control. A human who builds a standard machine retains control over the machine. Accidents occur with machinery—an automobile accident, for example—but even then the accident is not caused by the vehicle’s agency. Human error in navigation, human error in construction, or inclement weather may be culprits, but accidents do not occur because the automobile makes a wrong decision.

Conversely, a machine that learns through its own experiences may possess skills and abilities unknown to its human originators. One example is AlphaGo, a computer program designed to play the board game Go (Gibney, 2016). AlphaGo learned by playing thousands of games against human competitors and fellow computers, improving to the point that, in 2016, the program beat world champion Lee Sedol four games to one. During the match with Sedol, the developers of AlphaGo did not know which move it would play next. Their best guess would likely be wrong, lest one of the programmers beat the world champion. The victory of AlphaGo is considered a milestone in the history of machine learning since Go is known as a game requiring not only rote memorization, but strategy and intuition. AlphaGo showed autonomy, acting independently of human input (albeit in a narrow fashion). Nonetheless, this example of machine learning demonstrates that “smart” machines can act in unforeseen ways and outperform humans in tactical proficiency.

Considering that machine learning may only be in its infancy in terms of potential (Arel, Rose, & Karnowski, 2010), it raises numerous questions for the counseling profession. For example, if counselors-in-training can learn, improve upon their mistakes, and eventually cross the threshold to independent practice—and an AI shows the same skillset but learns much more quickly—how might autonomous AIs influence the field? Like Go, counseling too involves intuition and strategy. Would an advanced AI, functioning as a counselor, make moves questionable to even experienced counselors, but that pay dividends in the end?

If AI one day advances to the level of competent counseling practice, it will be through the underlying mechanisms that drive machine learning called algorithms. What culminates in a computer program besting a world champion Go player or, potentially, an AI employing a counseling technique, begins with a set of logic-driven instructions detailing how a task should be performed. The notion of an algorithm does not lend itself well to a rigorous definition (Gurevich, 2012); however, Pedro Domingos (2015) provides a constitutive explanation of an algorithm as “a sequence of instructions telling a computer what to do” (p. 1). AI is a broad area, machine learning is a subfield, and algorithms are specific operations—like written communications that can both therapeutically inform and give conversational voice to the AI.

The road to counseling

The term “artificial intelligence” was devised by mathematics professor John McCarthy, who helped to organize a summer conference at Dartmouth College in 1956 about whether machines could be made to think (Copeland, 1998). McCarthy’s proposal laid out the basic premise of AI research: that if a feature of intelligence, such as learning, could be broken down into its component parts and operationally defined with precision, then a machine could be made to simulate it (McCarthy, Minsky, Rochester, & Shannon, 2006). The conference attendees set out to discover how to make machines use language (see McCarthy et al., 2006, for a complete discussion).

That conference is known as one of many AI milestones of the modern era, including the first meeting of AI and counseling. In many respects, counselors are in the business of communication and depend on various forms: oral, written, non-verbal, art, and music therapy. In 1956, those AI researchers set out to learn how machines can be made to communicate. Ten years after the Dartmouth conference appeared, the first chatterbot capable of communicating in a way reminiscent of a human counselor was developed. Also known as chatbots or virtual agents, chatterbots are computer programs designed to simulate human conversation (Deryugina, 2010). This debut bot, named Eliza, was finalized in 1966 (Weizenbaum, 1966). Designed to replicate a Rogerian therapist, Eliza was known for answering questions with questions (Mauldin, 1994). In their output, machines capable of communication give the appearance of machine-level cognitive ability. At present, chatbots do not literally think, but rather give the illusion of intelligent conversation by imitating it (Abdul-Kader & Woods, 2015; Mauldin, 1994; Warwick & Shah, 2014).

While the metaphysics of AI may be of indirect interest to counselors, a question proposed by AI founding father, Alan Turing, is directly relevant. Turing (1950) proposed a scientific research question: How well can a machine imitate human conversation? The question brought the debate paradoxically more into the empirical and subjective realms. The Turing test places a computer system against human subjective experience. Known as The Imitation Game, the test asks human participants to interact through text with an unknown entity (Saygin, Cicekli, & Akman, 2000). The entity could be a computer program or a human being, typing. If the participant guesses that he or she is conversing with a computer, the computer program fails. If the computer imitates human conversation sufficiently and convinces the participant, the program passes. In a field heavily invested in human conversation, the Turing test may prove pivotal when considering counseling implementation, ethics, working conditions, and accessibility.

Perception is reality to many people. Counselors would be well served to monitor public perception about psychological artificial intelligence. In doing so, counselors could decide that using psychological AI as a supplement to traditional counseling may benefit clients and the profession alike. To a small degree, chatbots like Eliza have mimicked counseling skills for some time. Counselors themselves may disagree. However, if or when the public views psychological AI as relatively synonymous with counseling, counselors would be wise to pay heed.

Four levels of implementation in counseling

The American Counseling Association (ACA) defines counseling as “a professional relationship that empowers diverse individuals, families, and groups to accomplish mental health, wellness, education, and career goals” (D. M. Kaplan, Tarvydas, & Gladding, 2014, p. 366). The definition can be broken down into three pillars of counseling: (a) forming a professional relationship, (b) empowering, and (c) accomplishing goals.

The act of counseling requires the fulfillment of all three pillars. However, we might say that if one or two of the requirements are met by an AI, then that AI is getting closer to functioning as, if not being, a counselor. For example, an AI capable of empowering an individual towards accomplishing a wellness goal is partially functioning as a counselor because two of the three requirements are met. If AI takes on a more prominent role in counseling, we should expect to see the functions of a counselor met—or potentially exceeded—by artificial intelligence.

Based on the premise that AI has been and will continue to be applicable to counseling, I describe four levels of implementation: historical, contemporary, near future, and long-term. The levels propose to help navigate an AI-infused reality by correlating them with time orientation and influence on the field of counseling and comparing them to the ACA-sanctioned definition of counseling. Where the first level, historical, shows that AI’s past involvement with counseling was minimal, the final level has yet to happen but is marked by AI showing sophisticated and highly influential involvement in the field.

Summary

Each implementation level sees AI growing more into the fabric of counseling (see Table 1). The past saw nominal AI implementation to the counseling field, but the present has seen an AI resurgence. There are strong indications of more AI research in the future as the European Commission, U.S., and China devote billions of dollars to funding such endeavors (Cath, Wachter, Mittelstadt, Taddeo, & Floridi, 2018; Kelly, 2018; Larson, 2018). Whether the research surge brings about levels three and four remains to be seen.

OPEN IN VIEWER

Table 1.

Impact of AI level implementation on pillars of counseling process.

		Level 1: Historical	Level 2: Contemporary	Level 3: Artificial General Intelligence	Level 4: Superintelligence
Pillar of Counseling	Professional relationship	No	No	Central ethical question	Yes
	Empowers	Likely no	Unknown	Yes	Yes
	Helps accomplish goals	Likely no	Likely yes	Yes	Yes

Discussion

This article intended to define and explain AI concepts, to discuss how AI pertains to clinical counseling, and to present AI-in-counseling implementation levels from a theoretical viewpoint. Four metalevels of implementation were presented. The levels correspond to time orientation, with level one relating to historical and level four to future implementations affecting humanity in the long-term. I acknowledge that the future is unknowable to some degree, but as climate scientists forecast a hotter world due to global warming based on data patterns, so too are AI prognostications grounded in current research (Hulme, 2016).

Artificial intelligence and counseling already interface. In the future, the extent to which they interweave will depend largely on AI’s rate of growth, which, if current trends continue, will fall somewhere between sequential and exponential. With exponential growth, for example, an AI capable only of posing elementary questions one day could learn advanced assessment, diagnosis, and ways to embody the ethical, cognitive, emotional, and relational characteristics of expert therapists (Jennings, Sovereign, Bottorff, Mussell, & Vye, 2005; Skovholt & Jennings, 2004) essentially overnight. Exponential growth is not certain, but explosive growth is certainly plausible (Pratt, 2015; see Kurzweil, 2006, for a technical explanation of how this might occur).

The presence of AI and high-technology in counseling looks to continue, and even current-level AI implementations in counseling raise a host of practice-oriented and ethical questions regarding how and when AI use is appropriate or effective, to which degree it can be used in place of a human counselor, how it may affect a person seeking human connection via counseling, whether data produced during AI use could be stored in a hacker-proof manner, and whether counselor and client AI are adequately trained and informed on AI practices.

At present, the counseling literature contains a paucity of articles addressing AI from a descriptive, correlative, or experimental basis. More research could inform clinical practice if clinicians employ AI-assisted supplements, such as the psychological AI Tess, to help their clients. Research could also inform thought-leadership if a need arises for the ACA to address AI at a public policy level. Perhaps the most immediate need for research is in counseling ethics.

Using Green’s (2018) outline of ethical concerns surrounding AI as a guide, research must focus on the ways in which AI counseling services can avoid negative side effects, overgeneralizations, and potentially harmful exploration in strategies and techniques. Further, attention must be dedicated to ensuring AI functional transparency, or ensuring that AI actions can be understood by those designing, manufacturing, implementing, and interacting with it. Another ethical concern revolves around data security and privacy practices when implementing AI services. Finally, investigations should seek to determine the extent to which both counselors and clients need to be versed in AI technology and implementation to ensure fairness, beneficence, and non-maleficence in practice and counselor and client safety and wellbeing (Green, 2018).

The counseling community needs further information about the effect AI services could have on people specifically seeking out human interactions because they feel unheard, unseen, and unworthy of the care of others. The shift from human to human-like interactions in counseling, as well as other fields, may bring about a plethora of unchartered existential questions. Coupled with the onslaught of induced unemployment, socio-economic inequality, growing technological dependency, and human de-skilling, these existential questions may warrant closer attention and preparation by researchers and those who specialize in human emotion and crisis, such as counselors (Green, 2018). AI brings power and influence that can be abused. Research helps prepare the profession to address ethical questions when they arise.

More research is needed about psychological artificial intelligence. Considering its burgeoning nature, there is a dearth of research on the topic and noteworthy is the absence of literature about ethical ramifications. This article fills a research gap at the theoretical level, offering a taxonomy with the proposed levels of implementation and providing structure for forthcoming literature. For example, the nature of a clinical ethical dilemma will look different at level one compared to level four. Theoretical pieces carry inherent advantages and limitations. Advantages include providing constitutive definitions to guide future inquiry and high-level context to frame AI implementation and influence on the field. A limitation is the lack of specificity and clinical examples found in an abstract, categorical offering. Further, as AI is developing into a vast interdisciplinary field with weekly or even daily developments, no single article can capture its actual reach and consequence. Examining AI’s impact on a diverse clientele in clinical counseling and identifying ways to prevent bias and discrimination from creeping into AI is a necessary, but yet unexplored focus of research. The intersection of AI and counseling is growing, and a corresponding body of research is needed to match.