Welcome to FullStack HR, and an extra welcome to the 48 people who have signed up since last week.
If you haven’t yet subscribed, join the 6900+ smart, curious HR folks by subscribing here:
Happy Friday,
Last week was yet another short week in Sweden; hence, there was no email last week.
But the sun is out this week, and the birds are chirping. I’ve done three outdoor bike rides (yay!) in the last couple of days, and that’s the clearest sign of spring for me.
If you are a Swede, I’m studying the use of AI in HR in Sweden.
Take five minutes to help me out here.
OpenAI has also released GPT4o, and I did a podcast about that.
Listen to it if you haven’t already!
Links out of the way! What’s more to tell you?
Last week, I worked with perhaps my favorite client, giving an online lecture as a follow-up to our in-person lecture a couple of weeks ago.
I got feedback mid-session: “It’s great, but could you slow down a bit…” That’s what happens when I get too excited - I talk too fast. (Sorry for that.)
Overall, I believe the outlook for AI in this particular place is strong. They have dedicated resources internally to utilize the work we started together, and it will be very interesting to follow them on their journey forward.
The other week, I also did a session with board members at one company, which sparked intense discussions. They challenged me like I’ve never been challenged before—I love that.
We had a long and quite fierce debate about what AI will mean for the education system and what assumptions we put on our students.
What will happen when students outsource part of their knowledge to an AI and don’t learn particular parts of a profession? Will it be like when we started to use calculators, everything will be finer or finer, and we will incorporate it into our everyday lives? Or will we get dumber?
I’m in the “we will incorporate”-this camp, but that was under fierce scrutiny!
Which, once again, I really like.
We need to discuss these issues, and I appreciate having the opportunity to talk to different organizations about how to handle them.
Want help in your organization?
Send me an email.
One commonly asked question I get is:
“What’s the difference between the different Generative AI models?”.
Since GPT4o was released and Gemini also got an overhaul this week, I thought it would be a good time to compare the models using some HR prompts.
My answer has usually been that there are slim differences, but that is just according to my subjective analysis, which I’ve gotten by using these models a lot. But what if we threw HR tasks at the models to see how they handle different yet quite common HR use cases for working with these models?
I will use three commonly available Generative AI models, and I’ve chosen them because they are available in most countries. I would say that one is missing from this: Claue. In my opinion, it is one of the best models out there, but as it’s not easily available in the EU (e.g., you can’t access it just by signing up on their site), I’m not going to include it.
The three will instead be ChatGPT-4o by OpenAI, Copilot Pro by Microsoft, and Gemini Advanced by Google.
In theory, ChatGPT-4o & Copilot Pro should be very similar in their responses due to the fact that they share much of the same training data; they both utilize ChatGPT-4 as their foundation when giving answers. However, as we can see, they still use different flavors.
Let’s get going!
Test 1 - The Job Ad
The first one is what I believe to be the most common use case when it comes to GenAI and HR - the job ad.
This is the prompt I’ll be using. It’s a made-up company, which could confuse the models, but I want to see how they handle that. I’ve also deliberately asked the job ad to be between 550 - 800 words as this is a bit longer than most job ads, just to see once again how well it does with these specific instructions.
Act as a recruiter who is also an excellent copy-write. Write a detailed job description of between 550 and 800 words for a Process Technician at MedTechOne; the position is based in Stockholm. Be creative and create and engaging job ad, we want high conversion rates and my career depends on this job ad so make it a good one!
With a focus on creating processes for our pharmaceutical products. Include areas of responsibility such as process research, prototype development, process testing and collaboration with cross-functional teams. Highlight the necessary skills such as creativity, ability to collaborate, analytical ability, communication expertise and extensive knowledge of processes. Application is made via LinkedIn.
ChatGPT chugs down my instructions with ease and creates a decent job ad. Not too hyperbolic, not too creative.
It is not the world’s best job ad, but it is not the worst from my subjective point of view. The job ad is 685 words long, so it falls within the remit of what I asked it for.
Next up is Gemini. Is it trying to make it more “fun”? Is “Process Wizardry” its way of saying, “Hey, I’m trying to be cool?” 440 words, so it’s missing the mark with almost 100 words, despite being given clear instructions on length.
And lastly, Copilot Pro (where you don’t have the availability to screen capture the whole response but if you click this link, you can read the answer.)
It also does a decent job, just as the other models - not great, not too bad.
It outlines the basics but doesn’t in my view articulate the ins and outs of the role as well as ChatGPT does, despite having the same model behind it.
Winner?
For me, ChatGPT is the winner here. I think this is closest to what I had in my (subjective) mind when starting this out.
What do you think?
Test 2 - The Remote Work Policy
You've seen this example before if you’ve heard me talk about AI & HR.
It’s a prompt to showcase how detailed you can be with these models and how you can adjust them to do exactly what you want.
But will they do just that?
I need help designing a remote work policy for my organization. It should include guidelines for the work environment, communication expectations, availability during the workday, and management of company equipment from home. I prefer that the information is presented in an easy-to-overview bullet list and that it covers both the responsibilities of the employee and the employer. Please keep each point short and concise, with no more than two sentences per point. Additionally, include a short introduction explaining the purpose of the policy and its value to the company culture.
ChatGPT follows my prompt closely here and in the end, it also clarifies the employee's and employer's responsibilities. It’s quite formal, and one could perhaps argue around the tone here, but on the other hand, I didn’t specify that either.
This would, in my opinion, be a great starting point.
Geminie does as I intended to as well.
From my point of view, this is slightly better than ChatGPT; it’s easier to overlook and feels more like I would personally write such a policy. It’s detailed yet to the point.
Somehow it makes the language less formal and more like I would write it.
Copilot Pro gives a short answer that I can capture in one screenshot. It’s not wrong, and the things it lists are correct and things I would have included but perhaps a bit too short. Once again, this could be my prompting, but I would guess that most policies around remote work are a bit longer.
From my point of view, Gemini hits closest to home and is the winner in this test.
Test 3 - The Salary Review training
Next, this is how I really like to use Gen AI—to create outlines and syllabuses for training. I think this is one of the best uses of general AI models. They will not create the training for you but give you a head start and concrete and applicable feedback on your training program.
The prompt is quite long and detailed, and if I were doing training, I would see this as a starting point. I’m not expecting any models to get this right on the first try but rather see it as an iterative process. The model and I would go back and forth here until we have it nailed down.
When performing tasks like this in the past, I’ve also simplified them and broken them down into smaller chunks for the AI to handle.
Let’s look at the prompt and how the models handle it.
Act as a highly-skilled L&D professional who’s an expert in creating engaging and useful training. The trainings are liked by the participant and focuses on a blended learning approach with both examples from real-life and work done by the participants, such as discussions and sharing of experiences. Develop a comprehensive 45-minute training session specifically designed for managers in medium-sized companies on how to prepare and conduct salary reviews effectively. The session should aim to equip managers with skills and techniques for: Preparing for salary reviews: Analyzing and understanding the employee's performance and contribution to the company. Gathering relevant market data to support salary proposals. Formulating clear and measurable goals for the employee's future performance and development. Conducting salary reviews: Building open and honest communication. Presenting the salary proposal in a respectful and motivating manner. Handling and navigating the employee's reactions and potential objections. Negotiation techniques: Strategies for handling the negotiation process. Creating a win-win situation for both the employee and the company. Setting realistic limits for salary adjustments. Follow-up steps: Documenting agreements and updating the employee's terms of employment. Following up and evaluating the employee's performance in relation to the set goals. Include practical exercises, examples of dialogues and case studies to make the session interactive and engaging. Also provide tips on digital tools and resources that can support managers in their preparation and conduct of salary reviews.
ChatGPT outlines a 45-minute training in a couple of seconds, and, I could see myself hosting something very close to this to prepare managers for an upcoming salary review.
It includes what I want and what I outline in my prompt.
Once again, I would see this as an iterative process, going back and forth. It’s not done training yet, but that isn’t the point here either. The fact is to have a starting point, and idea generation.
Gemini also delivers a solid overview of a process. Same as with ChatGPT, I would have downplayed or switched the “Negotiation” part of the training, but other than that, it’s something I definitely could have used as a start.
Copilot also handles the task well. (Link to the full answer.)
As with the above, I would have downplayed negotiation, but all in all, it was a solid suggestion and a great starting point for further detailing of the training.
This was a hard one to judge, but I do think that ChatGPT and Gemini hit closer to home here as well. However, it’s hard to make a call on which one is better. I would probably have used ChatGPT due to the fact that I’m more comfortable with the model and know more of its limitations, but Gemini shows that it’s not too far off.
The point of all of this?
The main takeaway is that all three models—ChatGPT-4o, Gemini, and Copilot Pro—offer solid starting points for various HR tasks. While there are nuances in their outputs, the differences are often minor. The best model for you might simply be the one you're most familiar with. Using it effectively can significantly enhance your HR processes regardless of which model you have.