AI is only as good as the experts behind it

There has been some negative media attention lately surrounding the use of Artificial Intelligence (AI) in the recruitment space with warnings ranging from the fact that AI produces a shallow candidate pool to more serious things like amplification of bias.

There are many instances of AI being used in a way that has harmful outcomes, but it is important to clarify that this is about how AI is being implemented and not an issue with the use of AI itself.

?When AI is used appropriately, responsibly, and following regulatory guidelines it is an incredibly powerful tool that can create fair outcomes for candidates who are selected without bias - in a way that no other tool at our disposal can.

?This is why we think it’s worthwhile that more people better understand AI and some of the differences in the way it is used and implemented.

THERE IS NO UNIFIED AI

?Most media articles refer to AI as if it represents a singular master algorithm and fail to identify?how varied the implementations of it are. Almost all AI we have today falls into the category of “narrow AI”, in other words algorithms, mostly machine learning, built to solve a specific problem. E.g. classify sentiment, detect spam, label images, parse resumes. These purpose built AI are highly dependent on the nature of the underlying training data and the expertise of the developers in making the right assumptions and tests of validity of their models.?When built in the right way and used responsibly, AI has the ability to empower humans. This is why at Sapia.ai we have made various conscious design choices and adhered to a framework called FAIR ? that tests for bias, validity, explainability and inclusivity of our AI based tools.

?TEXT OVER VIDEO

The biggest cause for alarm is when AI is applied to analysing video, which can lead to irrelevant inputs like clothing, background, and lighting being used as predictors of personality and job-fit. Video and speech patterns also make it nearly impossible to remove demographic information like race and gender as inputs. Additionally, analyzing facial expressions is problematic, especially when evaluating certain candidates like those with Autism Spectrum Disorder or other forms of neurodiversity.

This is why Sapia.ai does not, and will not, use AI scoring for video interviews or even voice transcriptions from videos or audio given the word error rate introduced in transcribing speech. Instead, we opt for text - which we implement in a friendly no pressure environment that feels like you are texting a friend.

?It’s worth noting that no data other than the answers given by the candidate are used in the ‘fit score’ calculation – that is, we never use demographic data, social media, CV or resume data (which also contain demographic signals, even when de-identified), or behavioral metrics such as time to complete.

Even a candidate’s raw text itself contains gender and ethnicity signals that can introduce bias, if not mitigated. This is why we only use feature scores (e.g., personality, behavioral competencies, and communication skills) derived according to a clearly defined rubric in our scoring algorithms, which our extensive research shows contain significantly less gender and ethnicity information than raw text.

AIM TO UNCOVER HIDDEN TALENT WHILE MEASURING POTENTIAL

Another common concern is that AI will result in more uniformity rather than diversity in the workforce as algorithms narrow the pool in order to search out an employer's ideal candidate. There are several things worth noting here.

?First, identifying what the ideal candidate is –?that is, what knowledge, skills, abilities, and other characteristics are important for success in the role?–?is what a job analysis is for and should, legally, be what your selection tool is designed to measure.

This is also not specific to AI, as all selection systems are designed to identify which candidates have a profile of traits and characteristics that indicate they will likely be successful in the role. This doesn’t automatically mean that every hire is going to be exactly the same, though. When you focus on the traits and characteristics that will set someone up to be successful, considering potential more than background or pedigree, you’re more likely to uncover hidden talent and hire more successful people from a broader, more diverse range.

Relying solely on past data to build your model also runs the risk of introducing historical data biases. This is actually why it is so important to consider the ideal candidate profile and use that to inform your scoring model. We strongly believe in keeping the human in the loop, which is why our scoring models are centred around the human-determined (via job analysis) ideal candidate profile and then optimized to ensure all bias constraints (e.g., 4/5ths rule and effect sizes) are met.?

?Using this approach, Sapia has helped clients achieve their DEI goals and increase their diversity hires, including impressive statistics like hiring 3x more ethnic minorities, 1.5x more women, and 2x more LGBTQ+ candidates in just 3 months.

KEEP AI PROCESSES TRANSPARENT

Lastly it’s worth acknowledging that there is often a?“black box” mystery of how AI recruitment tools work. People don’t trust what they don't understand. While we don’t expect everyone to be an expert in AI or Natural Language Processing, we do strongly believe in building trust through transparency and work hard to make sure that our models are easily understood and open to scrutiny. From third-party audits to detailed model cards to in-depth dashboarding and reporting, we aim to maximize transparency, explainability, and fairness.

?We believe a fairer future can only be achieved when AI is used responsibly. AI is not the enemy, rather it’s the experience and motivation behind those promoting it that can make the difference between what is good AI and what is harmful AI.