Can an AI coach be as good as a human coach?

5 Nov

Spoiler alert: For the kind of coaching that I do, and my colleagues at Spring Leadership do, the research suggests that the answer to the question, “Can an AI coach replace a human coach”, is probably no. Certainly not for a long time and not without some as yet unimagined solutions to very hard problems. To the best of my (non-academic) ability, I here lay out the evidence-based rationale and research that I found.

I find myself genuniely curious about AI coaches and cautious about my own bias towards humans. I value the human “lived experience” and the developmental opportunities created from our experiences. Humans are wonderfully complex, gifted, flawed, and unique. We grow, we learn. The person we’re becoming next may feel quite different from the person we’ve been before. (As an aside - check out vertical development if you’re interested in what this might mean for you). Recently, one of my colleagues reflected back a conversation she’d had with her partner (who is in computing/systems/IT). She was telling him about some of the challenges, suprises and complexities of successfully coaching a human being and he said “So why would you ever think AI could do this?”. Huh. Right! Why would we?

But yet, I sometimes hear commentary such as:

Artificial Intelligence (AI) will take over coaching.”
“AI coaches are just as good, maybe better, than human coaches because they don’t suffer human failings such as getting emotional, having “bad days”, blinded by triggers and blind spots, lacking experience."
“It’s only a matter of time before AI coaching programmes learn to be at least as good as their human equivalent.”

The use of Artificial Intelligence (AI) in coaching is certainly growing quickly. It comes in various forms including:

AI coach matching algorithms - where the algorithm offers you coaches based on your preferences and situation;
AI supported learning curation for coaching clients - where the algorithm suggests learning, tools, videos etc on what it thinks you might find most useful;
AI supported assessments - for example, the VMI, a vertical development assessment scored by AI or recruitment processes using AI to weed out and select candidates;
AI based coaching, based on an algorthm - for example: Cultivate, Leadx and Coach M. This is a hot market with much more to come.

All categegories pose concerns. All face major ethical, practical, and privacy issues. My commentary here relates to all, but it’s also the last category that I’m particularly interested in at the moment and I wondered what research I might be able to find on this topic:

Can an AI coach do what a human coach can?
In what kind of situations might an AI coach be helpful, or harmful? A basic tenet of professional coaching is to “do no harm”. Do AI coaches and organisations hold a similar set of enthical standards? Who holds them to account? (No-one? Just because an organisation can do something, does it mean it should?)

Some time ago, I trialed one of the AI coach applications for myself. Although it asked me about goals and offered quite good nudges, questions and prompts, it was linear, superficial and narrow. If my goals and learning needs were straight forward, transactional, easily repeatable no matter who the person, or very specific (sports, fitness coaching?), then perhaps an AI coach might be quite effective? Maybe? Or maybe not - what if the human has a whole lot of other things going on in their life? It wasn't much good for me. I just got annoyed. I couldn’t change my goals, couldn’t widen my view, couldn’t add in new context, couldn’t seek clarity or insight in sharing thoughts into the space between us, and, most importantly, I couldn’t “feel” seen, heard, understood, nor feel like my experience was shared or cared about.

But, I wondered. I'm clearly biased and maybe I'm also blinded by ego. What if an AI coach could learn to be like a human coach? Is it inevitable that AI will improve so much that an AI coach will actually end up better?

No, I think not and here's how I've been thinking it through (so far, at least!)

My own coaching experience, since 2007, working as a professionally credentialed coach;
Issues relating to models, theories, tools, content, diversity and equity;
Issues relating to differences in vertical development stage;
Wider research relevant to AI and coaching;
AI ethics;
My experience as a coach for a large coaching platform organisation heavily using AI and investing millions of dollars in further development.

My own coaching experience

I often find myself reflecting back on a coaching session with some fascination. For example:

I find myself surprised as the direction of the conversation, or the speed of an insight, or the sudden depth in which we find ourselves. A few examples:
- Recently, deep into a conversation, I noticed my client ever so subtley shift. It was a micro-moment, almost a nothing, but not a nothing at all. It was so subtle and my client wasn’t aware of it herself. I said, “oh, what’s happening, have you slightly backed off?” She looked at me, paused for a few moments and then said, “Yes, I think I have. But, I’m not sure why”.
- With another client, I made an observation that her body language that was subtley contradicting her words. While she was saying “I’ve made sense of this”, I suddenly got the sense she was rattling - and I said that. She paused, took a breath and said, yes, rattling, and she continued to reflect in a follow-up email after the session. It was something important for her.
- For a long time (almost an entire lifetime), another client has been trying to figure out her unique contribution, and in a moment, I just said back what she’d said about her values of bravery and fun and asked, I wonder if this also is also what you are driven to offer the world? She had never thought of it like that. I’d disrupted her thinking loop. Her eyes widened slightly, and she took a breath. I noticed and she noticed. “Yes! What if!” she exclaimed.
I find myself reflecting on the challenge of some conversation. It feels hard. At the end of the session, I may almost wipe my brow and I literally feel like I have reached out into every corner of my experience to find ways to support my client. As an example, recently I had a client in what felt like “a game of two halves”. The first half was hard. My client did not want to be there, did not want to engage, but also did. Although this was somewhat invisible to herself - and me at times! The second half was completely different. The walls came down, and she started to reveal herself to me - and to herself too. I was there WITH her. And that made all the difference. A co-regulation process had emerged. She felt safe, seen, heard and was starting to see the doorways, for her own growth and wellbeing, opening.
I notice what organically emerges in the conversation and what the client decides are the things that most resonate with them. Quite often these things are not the things I might have predicted. (Though, an interesting research study that might be!) To make even more complex, sometimes, the client speaks to what’s resonated and then as we chew over their insights, and yet another insight or direction emerges, that’s even more core to their current situation.
Co-regulation. There’s something special about pure human presence and the conversation that emerges between two humans. How can this happen in the same way with an AI coach? (I don’t think it can).
Dr Dan Siegel defines the mind as an “emergent, self-organizing, embodied, and relational process that regulates the flow of energy and information.” When I completed Dr Siegel’s course on Interpersonal Neurobiology, it became clear that our “mind” is scientifically proven to not just be the thing within our skull. The mind is an energy and information flow within and between us (thus, Dr Siegel’s “MWE” model - we are not just Me, but part of the larger MWE). Part of that is the energy between coach and client. We see our clients, sense the energy and co-regulate to match, or calm or ease our clients state.

Issues relatiing to coaching models, tools, approaches

I find many questions troubling, such as:

Do AI coaches rely reavily on models that are assumed to be “good”? How well do these models or approaches meet the needs of each client? For example, how do AI coaches create respectful and aware ways of dealing with cultural differences, and diversity? In “The Liberated Coach”, David Clutterbuck describes the dangers of a limited model of coaching.
To what extent to the AI Coach algorithms rely on blanket “goodness” of certain approaches? For example, mindfulness may be helpful sometimes, but not as a blanket approach. Or the commonly used “reframing” approach in coaching. Yes, reframing is a powerful tool - sometimes. Other times, if a client is trying to come to terms with a situation, then an effort to force through “reframing” verges on toxic positivity and may do more harm than good. It places the emphasis on the individual. The implicit message is that if you’re not achieving, you’re not trying hard enough, not good enough, too weak (etc).

Issues relating to differences in vertical development stage

Taking a Vertical Development lens to coaching, the adult development stage of the coach and the client is important. According to Otto Laske’s research, “if the coach functions at a lower stage of maturity than the client, ethical issues can arise since in that case the coach may arrest or delay, and thus harm, the client’s mental growth”. This raises many concerns about AI coaches. What stage are AI coaches programmed to centre, and is this a conscious or unwitting choice by their creating organisations?

Research related to AI, the human brain, and coaching

“Artificial intelligence has made great strides since the deep learning revolution, but AI systems still struggle to extrapolate outside of their training data and adapt to new situations.”
“We believe it is helpful to study the best solution to the intelligence problem that evolution has discovered - the human brain”.
- Daniel C. Elton - Cognitive Systems Research - “Applying Deutsch’s concept of good explainations to artificial intelligence and neuroscience” - December 2020

The brain uses two types of processes. How the human brain works matters when we consider how AI may be able to, or not, learn to equate a human in certain circumstances.

In “Thinking, Fast and Slow”, Daniel Kahneman, describes the two systems used by the brain:

• System 1 operates automatically and quickly, with little or no effort and no sense of voluntary control. [Thinking fast]
• System 2 allocates attention to the effortful mental activities that demand it, including complex computations. The operations of System 2 are often associated with the subjective experience of agency, choice, and concentration. [Thnking slow]

Kahneman describes the difference with an example: by glancing at someone’s face, you may instantly think you recognise an emotion because they have, for example, raised eyebrows, or a furrowed brow. This is an example of “thinking, fast”.

On the other hand, Kahneman offers the contrast of System 2 thinking:

17 x 24

He says, you probably knew this was a maths problem. Perhaps you thought you would be able to solve it with pen and paper. You might even have a rough idea of the solution - e.g. not near 80 or 20,000. But the exact solution is not quick to come to mind. This is an example of thinking slow. Your thinking proceeded through a sequence of steps. The process included a mental effort.

If both System 1 and 2 processes are important to “intelligence”, then the problem of developing a “super-intelligence” (generalised intelligence) gets much more complex and further away. Assuming that super-intelligent, general purpose AI might operate in a relatively similar fashion to the human brain. If the brain relied mostly on System 1 processes, AI development may be relatively more straigtforward and rely on “deep learning” and increasingly larger datasets as input.

“If hard-to-vary* explanations are an important part of human intelligence, the problem of how to program an AI to generate/discover such explanations must be solved first before truly human-like AI can be produced.”
- Daniel C. Elton - Cognitive Systems Research - “Applying Deutsch’s concept of good explainations to artificial intelligence and neuroscience” - December 2020.

Elton’s exploratory article considers Deutsch’s rebuttal to Occam’s Razor (simplifying won’t work) and induction (you can’t just infer conclusions and use “brute force” to work them through).

Elton’s article casts many questions on the ability of deep learning AI to become generalised superintelligent. He quotes: “As Francois Chollet has beautifully argued recently, the true test of intelligence is how well a system can adapt to solve new environments and tackle entirely new problem situations.” This last quote gave me pause. Coaching conversations are vastly, deeply, fascinatingly unique. Just as situations and people are vastly, deeply, fascinatingly unique.

As Marcus Buckingham in his latest book “Love + Work” says: We have one hundred billion neurons in our brains that link up to form a n et work of one hundred trillion unique connections. How big a number is that? Since there are approximately four hundred billion stars in the Milky Way galaxy, and there are a thousand billions in one trillion, your brain has more connections within it than five thousand Milky Ways.

Elton argues that a key issue for AI development is around how to build systems that can generate “good explanations” of the world that would enable AI to deal with new situations.

At the moment at least, and perhaps for the long-term, this seems beyond the reach of AI coach working with humans about topics relating to complex human needs, challenges, aspirations and contexts.

Ethics of AI in coaching

Another slant on AI in coaching is one of ethics. This is a huge topic and I certainly don’t feel qualified to do justice to it here, other than to mention a couple of examples of questions I wonder about.

Do no harm. How does an AI coach determine the subtley and invisiblity of a real person (a client) verging towards trouble? The client themselves may not be fully aware so how would an AI coach become aware and what would they do? As far as I understand it, AI deep learning development is narrow in scope, but people are not narrow in scope.
When and why should you explain how your AI works? Reid Blackman and Beena Ammanath, (HBR, August 31 2022) say that while we may understand the inputs (the variables) going into the AI, we may not understand the how that then creates the outputs. There’s effectively a “black-box” between inputs and outputs. So therefore, how do we know whether to trust the AI? Can we trust a tool we don’t understand when the stakes are high? If the stakes aren’t high, perhaps it doesn’t matter much. We give it a go. But when we are dealing with real people and complex and unpredictable conversations and situations, the stakes are high. The authors argue that “explainability” is important when the stakes are high, and they are high when the outputs of AI impact how people are treated. But attempting to explain how the AI works takes time and resources. Perhaps it also risks revealing intellectual property. Therefore I assume, organisations are less likely to do so.
How do AI coaching organisations manage privacy and confidentiality and protect ourselves and our clients from data harvesting? How do those same organisations ensure fairness, transparency and equity? And, who keeps them accountable for a wider societal responsibility? (No-one?)
How do AI coaching organisations embrace diversity? (And again, who calls them to account?)

My own experience as a Coach for a large coaching platform heavily investing in AI to support coaching at scale

I worked as a “platform coach” for a large tech coaching company for several years (ending 2022). Over the period of time I was with them, I experienced a massive increase in AI supported applications. I did not personally experience an AI coach with this platform. However, some of the ethical, research and human issues I pose in this article, came to the fore in many situations. For example:

AI algorithm for offering a selection of coach “matches” to each client. We, the coaches, were in the dark about how the algorithm worked. The test of “explainability” posited by Reid Blackman and Beena Ammanath was not met and the lack of transperancy had a considerable impact on coach’s incomes.
In addition, AI algorithm collated survey feedback after every coaching session about coach performance and client satisfaction. Coach renumeration was directly linked to feedback ratings and yet was not explained (the “black box” problem) and as the power imbalance was large, the division in values considerable (profit motive), coaches were not able to influence the system in a meaningful way, and many left (like me). Probably to be replaced with more AI and new internally trained (cheaper) coaches using the argument that scaling cheaper coaching systems is necessary to get coaching out to the masses who would otherwise not have access. Hmmm. I wonder about many issues and questions with this. Is the heart of coaching about human conversations? How do you feel about living in a world that’s not prioritising human interactions? How do the humans at your work or in your life develop better conversation skills? How does the system support this? The ever-increasing drive to scale up and profit comes at what cost and to whom? How and why is data collected, harvested and used? Who could have influence in holding the organisation to account? Clients? Conscious leaders? Teams? Professional bodies such as the International Coach Federation? Conscious shareholders or venture capatalists looking at impacts wider and longer than dollar returns?
AI algorithm supported an ever-increasing amount of automated resources and tools offered to clients. Were these useful? Who knows? Results and outputs were not reported back. And as far as I know, neither client not coach got “good explanations” - the “black box” issue again.

My conclusion

Firstly - well done reading all the way to here! Impressive! There are a lot of questions, concerns, perspectives I’ve not included here - and that haven’t even occurred to me, (thank goodness I have wonderful colleagues and friends on this journey with me), but this is at least, a start.

My answer to the question I posed - Can an AI coach be as good as a human coach? The answer is... no. Not for the kind of complex coaching that I (and my colleagues) offer. AI deep learning development is at best not sufficient and at worst potentially ethical questionable.

Or at least, not for a long time, and only with some as yet unimagined solutions.

Notes:

David Deutsch (who I think of like a modern day Einstein) writes in “The Beginning of Infinity” about the importance of “good explanations”. A good explanation of a theory needs to be “hard to vary” (HTV). How much must the model be varied in order to deal with new situations?
Reid Blackman and Beena Ammanath use “good explanations” and “explainability” in a completely different context. They instead refer to a good explanation being one that is intelligible to its intended audience. (And knowing who those audiences are, is important).

References:

https://www.scientificamerican.com/article/kahneman-excerpt-thinking-fast-and-slow/

https://brucenielson.medium.com/hard-to-vary-vs-induction-a-response-to-kieren-26f8f779c4bb

https://hbr.org/2022/08/when-and-why-you-should-explain-how-your-ai-works? (Refers to https://www.amazon.com/Ethical-Machines-Unbiased-Transparent-Respectful/dp/1647822815)

https://interdevelopmentals.org/wp-content/uploads/2020/02/2006d-Why-Maturity-Matters.pdf

https://www.researchgate.net/publication/347917488_Applying_Deutsch%27s_concept_of_good_explanations_to_artificial_intelligence_and_neuroscience_-_An_initial_exploration

https://www.researchgate.net/publication/233357425_Coaching_reflection_The_liberated_coach

https://joshbersin.com/2021/07/ai-enabled-coaching-is-hot-and-theres-lots-more-to-come/

https://drdansiegel.com

https://www.youtube.com/watch?v=uo8Yo4UE6g0

https://www.porchlightbooks.com/blog/staff-picks/love-and-work

https://hbr.org/2022/04/marcus-buckingham-why-love-is-the-key-to-career-success