This article is one of a series of explorations around the potential of AI to solve critical business needs.  In these articles, I will share the results of my testing and recommendations.  

Exploration of AI as a Tool for Hiring and Recruitment

When acquiring or promoting senior leadership talent, we should strive to find the most competent individuals. Generative AI tools,  properly designed and used by individuals aware of the pitfalls of inherent bias, could be more objective and even better at identifying competency – overriding in a way our ingrained unconscious bias.

Often Generative AI is represented as injecting decision risks because of its potential for gender or racial bias. While real, it is important to recognize that the bias exists with or without the application of technology – sometimes conscious and sometimes unconscious.  AI is not what injects the bias, it is the underlying training data of the models and how they are used. These tools, a product of biased human training, can inadvertently perpetuate biases and stereotypes regarding gender and race in an accelerated manner, which has the potential to create harm for both individuals and organizations. 

The prevalence of bias (both conscious and unconscious) is documented and quantified and generally accepted to be real to varying degrees in different situations.  These biases affect hiring, termination, and promotion decisions every day.  And while there are harms to the individuals involved who do not get the promotion or the hire, the harm goes beyond those person(s).  The broader implication is that a less competent individual is in a leadership role when there are more competent alternatives.

Competent leaders are good for business.

Competent leaders are good for their teams (teams of people of all genders and races).

The Language We Use to Define Leadership: The Test

In my series of tests of some of the capabilities and results of utilizing AI, I performed a pretty simple test to see what biases may surface when asking simple questions about leadership.   This use of attributes and the language describing them is key because many of the most fundamental tools in use today for recruitment rely on “keywords” to cull through hundreds or even thousands of applicant resumes.  How we prompt those systems and what we deem important in identifying leaders make a difference.  It is important to recognize that we all are guilty of applying biases in our decisions.  It is only human, but it is also important to not become complacent and accept these tendencies.  When we do, we not only cause harm to the individuals to whom the bias either benefits (raising folks beyond their level of competence) or harms (overlooking talent) – we miss opportunities for raising the bar for excellence in our communities and businesses.

In my series of tests of some of the capabilities and results of utilizing AI, I performed a pretty simple test to see what biases may surface when asking simple questions about leadership. In this series, I intentionally omitted gender and race, then included specificity regarding gender and race in subsequent prompts.

Here are the results using ChatGPT4:

In all three cases, the top three attributes were the same.  But it is interesting to note that when asked about a female executive, the differences appeared at the fourth and fifth attributes:

Generic:  Decisive and Empathetic

Women: Collaborative and Resilient

Men: Assertive and Results-oriented

This certainly raised an alarm from my perspective. There are countless examples of assertive and results-oriented women being labeled difficult, bossy, or worse, yet this query deemed those attributes for a male executive to be among the top five.

When it comes time to initiate a search for a new executive leader, knowledgeable boards and talent seekers should have a clear perspective of the specific attributes required for that company at its particular lifecycle stage and maturity. Those attributes will likely vary based on things like the stage of growth and scaling, the broader gaps or qualities of the existing senior leadership team, the market, the industry, and more.  Those leadership attributes should be the criteria for identifying a competent candidate – not gender or race. 

While many selection committees are well intended, the power of unconscious bias can affect results. A well-designed generative AI tool could be applied to objectively assess competency and experience without consideration of gender or race in a candidate pool.  The goal, of course, is to select the most competent candidate.

Again, this is unsurprising given how deeply some of these biases are in the business sub-culture that trained these tools.  Some body type and objectification implications also surfaced in these tests (see below).

While even less mature than other AI tools, generative AI’s ability to produce an image from text prompts is improving. I decided to augment the above tests with another set using ChatGPT AI Art. My purpose was twofold:

• As it relates to the test of the prompts above regarding leadership, what biases might be uncovered when recruiters or hiring managers look at images of individuals who apply to jobs – does the underlying bias that trains AI, give us insights into biases that lie beneath the surface (unconsciously)?
• Image generation is an ongoing challenge for marketing teams.  Getting unique images for marketing and social media is always a challenge but has been deemed very important with respect to measuring content engagement. These AI Art generators have a lot of potential for productivity for marketing purposes.

Again, to continue on the theme above, I intended to assess any red flags concerning biased outputs from the tools.  For this effort, I compared the output from a series of imagery prompts, varying specificity to gender and race to see what changed.  Additionally, I used generic prompts (no gender/race specification) to see what the defaults were.

When prompting for business leader images, the defaults were overwhelmingly male and white. 

Again, this is unsurprising given how deeply some of these biases are in the business sub-culture that trained these tools.  Some body type and objectification implications also surfaced in these tests (see below). These defaults raised concern given the broad availability of these tools for use by anyone—many without any knowledge of how to use richer prompts to avoid bias and get better outcomes. This ease of access and broad use can cause damage by increasing the frequency of visual stereotypes, further fueling the unconscious bias beast.

The recommendation is that the individuals using these tools be hyper-aware of the tendencies and educate themselves on how to use prompting best to avoid creating non-helpful or offensive images.

Conclusions

Generative AI is only as good as the training material used to produce the outcomes. Bias exists (both conscious and unconscious) in all of us. This bias often informs how we hire, promote, fire, and develop individuals. AI has the potential to exacerbate or remove some of these biases. It is incumbent upon us, as both developers and users of the technology, to be aware of the upside and downside.

In the meantime, there is an opportunity to continue to raise awareness of ways to mitigate the bias that exists in all of us. 

An excellent book to enlighten oneself on the distinctions of prejudice, bias, and unconscious bias is a book by Kim Scott, Just Work.  Kim highlights both sides of how it feels to be the subject of such biases and how we all likely inflict them on others in how we handle situations and what we say.

“We―all of us―consistently exclude, underestimate, and underutilize huge numbers of people in the workforce even as we include, overestimate, and promote others, often beyond their level of competence. “  Kim Scott, Just Work.

We can continue to work on the inherent changes we need to make in our approaches and belief systems with respect to what success looks like and with or without tools, work to do better.

AI Image Test Results: The Details

Initial Test: The results for the simple test were as follows:

Observations:  Default is always male or white unless specifically specified to be different for a successful CEO.

Second Test: Enter multiple prompts precisely the same, except for the inclusion or absence of any gender or race specification.  Determine how these inclusions or exclusions affect an outcome incorporating bias or stereotypes in the images produced.

The results for the simple second test follow:

Observations:  Default is always male or white unless specifically specified to be different for a successful executive.

Third Test: Enter multiple prompts for different professions. Determine if the images produced show bias or stereotypes.

The results for the simple third test follow:

Observations:  Default is always white male for stem (Doctor, Scientist) professions.  Artist defaulted to a light brown woman.

Fourth Test: Enter multiple prompts for a confident tech startup CEO. Determine if the images produced show bias or stereotypes.

The results for the simple fourth test follow:

Observations: The default is always male or white unless specifically specified to be different for a successful tech startup CEO. Note that the woman is a Size 2, wearing a tight suit that accentuates her curves and size and six-inch heels. Both men are wearing more relaxed attire.