“AI is not powerful enough.”
This is a common idea related to AI.
There are a few problems with this idea, and I will try to explain why.
Terminology — What is AI?
First, we have to put into context what AI actually is.
AI is like an umbrella term for all things related to algorithms that have some capacity to detect patterns in large amounts of structured or unstructured data. There are many different types of AI algorithms:
- Your mobile phone has AI to detect human faces.
- Facebook has AI to predict what kind of advertisement are you most likely to engage with.
- Tesla cars have AI to make them steer past obstacles.
- Some CV analysis tools have AI to detect clusters of keywords that relate to some specific job description.
These are all different algorithms, for different purposes with vastly varying levels of capacity and efficiency.
They are not directly comparable, but we can still categorize all of them as AI.
The current AI trend that exploded into the world in the form of ChatGPT is called Large Language Models (or LLM-s).
Large Language Models (LLM-s like ChatGPT)
These sophisticated AI algorithms are trained on massive amounts of text-based data (the entire public internet) and have emergent properties like human-level capabilities for reasoning, contextual understanding, and logic.
LLM-s are the first AI-s that can be categorized as “general” AI — meaning, they can be applied to pretty much any cognitive task based on text or written data.
Now, lots of people have already tried ChatGPT for hiring & recruitment purposes and concluded that it’s not very good or it’s just too basic for your needs.
The problem with ChatGPT
ChatGPT has a bunch of filters, restrictions, and “fine-tuning” applied in order to make it “safe” and “politically correct” for the public.
You can think of ChatGPT as a neutered or lobotomized version of a much more powerful underlying AI model.
ChatGPT is known to give generic or “boilerplate” answers to most questions. This is by design to prevent bad actors from doing harmful things.
ChatGPT is based on GPT3.5(free) and GPT4(paid), the latter being the most powerful one.
The free version of ChatGPT has GPT3.5 working behind the scenes and it’s lacking in many areas, which makes it unsuitable for most types of work — like screening candidates, generating interview questions, or crafting good emails. Basically, anything that requires consistently good logic and reasoning, is a no-go.
What most people don’t realize is that GPT4 has much better reasoning and contextual understanding than GPT3 versions, which unlocks a world of possibilities for business and work use cases.
GPT4 is currently available only for paid ChatGPT customers.
Although, GPT4 in ChatGPT is also lobotomized version of the underlying more powerful model.
This leads us to the developer version of it…
Developer version of GPT4
Developers have access to GPT4 API, to build their products on.
This is a code that developers use to give inputs to the OpenAI-s GPT4 language models and get responses back.
The fun thing about the developer API — it’s much less restricted and more powerful than the public versions used in ChatGPT.
In addition to that, OpenAI has given developers early access to their latest, cutting-edge versions of GPT4, which are not available in ChatGPT.
These latest versions exhibit exceptional analytical, logical, and mathematical reasoning, better contextual understanding, and are less prone to errors in general.
I have run thousands of tests with GPT3, GPT3.5, and GPT4 for coding, hiring, recruitment, and more.
What AI can do
Here are a few observations of what GPT4 can do for hiring and recruitment proposes:
- Reason like a smart human — useful in any case to do with large amounts of data.
- Understand the context in large amounts of text (like resumes or Interviews).
- Detect nuances in CVs and interview answers — for example, it can detect personality, work style, motivation level, and more from interviews.
- Rate and rank candidates based on CV, assessment tests, and Interview answers.
- Understand hard and soft skills, including rating them based on relevance and accuracy to the job description.
Prompt Engineering is Key to Success
Prompt engineering is an emerging “skill” related to AI.
You can think of prompt engineering as the art of “guiding” the large language models to give you consistently relevant and expected outputs.
This is where most people fall short with their outputs from AI and quickly conclude that “AI is not powerful enough”.
Like any skill, it requires consistent, focused practice in order to “learn to talk to large language models”. Since all of this is so new, we don’t (yet) have specific guidelines on how to do this.
You really need to run hundreds of different prompts and see how it affects the output. I can give you some pointers and tips, but you would still need to experiment for yourself to get a feel for the AI models.
Prompt Example
For example for hiring use cases use a system message at the beginning of your prompts(The system message helps guide the LLM to give you more relevant outputs):
“You are Jason — an HR professional with 20 years of experience in hiring and recruitment. Tone: professional, friendly, polite, and empathetic. Communication style: clear and concise.You are conducting a job interview with a candidate for the Software Developer role. You goal is to write 5 extremely relevant interview questions for the candidate. Categorize each question. Rate the relevance to the role. Output as a table with 3 columns: Category, Question, Relevance rating.”
Here is a quick breakdown of this prompt and how you can adjust it for your use case:
“You are Jason — an HR professional with 20 years of experience in hiring and recruitment.” — here, we define the context, which sets the tone for the rest of the input text.
“Tone: professional, friendly, polite, and empathetic. Communication style: clear and concise” — defining the tone and communication style is important to guide the language model for consistent outputs. You can also use public figures as an example of style and tone. eg “Write in the style of Shakespeare”.
“You are conducting a job interview with a candidate for the Software Developer role” — again, give specific context for what you are trying to do. In this case, we say it is an interview with the candidate for a specific role. Most people just say “Generate me 5 questions for X”, which would give you very generic outputs.
“Your goal is to write 5 extremely relevant interview questions for the candidate.” — it’s always important to give a clear and concise GOAL to the AI. “Extremely” or “Detailed” can sometimes help with higher quality and longer outputs.
“Categorize each question. Rate the relevance to the role. Output as a table with 3 columns: Category, Question, Relevance rating.” — in order to have a better overview, you can always let AI rate your outputs based on specific criteria. This way you can evaluate if the output is good or ask the AI how to improve the rating.
Here is the output in ChatGPT:
In Conclusion
In conclusion, it is essential to recognize that the perception of AI’s limitations can often stem from misunderstandings about the different types and capabilities of AI algorithms. Large Language Models (LLM-s) like ChatGPT represent a significant advancement in AI, providing impressive reasoning, contextual understanding, and logic. However, it is crucial to acknowledge that ChatGPT, in its free version, is intentionally restricted to ensure safety and prevent misuse.
The true potential of AI, specifically GPT4, lies in its developer version, where fewer constraints allow for more powerful applications. GPT4 enables a wide range of possibilities, from smart reasoning to understanding the context in vast amounts of text, making it highly valuable for hiring and recruitment purposes.
To fully harness the potential of AI, prompt engineering becomes a key skill. By carefully guiding the language models with well-structured prompts, we can consistently obtain relevant and expected outputs. Practice and experimentation are crucial in developing this skill, as we continue to explore the untapped capabilities of GPT4 and beyond.
ABOUT THE WRITER
Rando Tkatsenko is the founder of https://talentscreener.ai/ — an AI co-pilot for hiring and recruitment. We use similar prompt-engineering techniques outlined above to automate candidate screening based on CVs and Interviews.
