Rethinking Data Science Interviews in the Age of AI

AI is rewriting the day-to-day of data scientists. , information scientists should discover ways to enhance productiveness and unlock new prospects with AI. In the meantime, this transformation additionally poses a problem to hiring managers: the best way to discover the most effective expertise that can thrive within the AI period? One crucial step in constructing a robust AI-empowered information staff is to revamp the hiring course of to raised consider candidates’ means to work alongside AI. 

On this article, I’ll share my perspective on how information scientist interviews ought to (would) evolve within the age of AI. Whereas my focus right here is on Information Scientist Analytics (DSA) roles, the concepts right here additionally apply to different information positions, reminiscent of Machine Studying Engineers (MLE). 

I. The Conventional Information Scientist Interview Loop

Earlier than speaking about how issues will change, let’s undergo the present construction of information scientist interviews. Except for the preliminary recruiter name and hiring supervisor screening, a typical information scientist interview course of consists of:

1. Coding interviews: SQL or Python coding questions to check syntax and fundamental logic.
2. Statistics interviews: Statistics and chance questions, in addition to the most typical statistical functions in information science workflows, reminiscent of A/B testing and causal inference.
3. Machine studying interviews: Deep dive into machine studying algorithms, experiences, and circumstances.
4. Enterprise case interviews: Focus on a hypothetical downside to check analytical considering and enterprise understanding — metrics, funnels, progress, retention methods, and analytical approaches.
5. Behavioral interviews: Customary “stroll me by way of a mission / a time whenever you XXX” to grasp how candidates deal with particular conditions and if they’re a cultural match. 
6. Cross-functional interviews: Information Scientist is a technical function, however additionally it is extremely cross-functional, aiming to drive actual enterprise influence utilizing information. Subsequently, many information scientist interview loops right now embrace a cross-functional interview spherical to speak with a enterprise companion to evaluate the area information, communication abilities, and stakeholder collaboration. 

From the record above, you may see that information scientist interviews normally have a superb mixture of technical and non-technical evaluations. However with AI getting into the sport, a few of these interviews will change considerably, whereas some will grow to be much more necessary. Let’s break it down.

II. How Interviews Will Shift within the Age of AI

For my part, how the interview loops are going to alter is determined by two issues: 1. Can AI deal with the duty shortly? 2. Does it inform how the candidate makes use of AI thoughtfully? 

Coding Interviews: Most More likely to Change First

What can AI do shortly? Easy coding duties. Subsequently, the coding interview might be the primary one to be impacted. 

In the present day’s coding interviews ask candidates to write down SQL and Python code accurately. The SQL questions normally require easy joins, CTEs, aggregations, and window features. And the Python questions may very well be easy information manipulation with pandas and numpy, or simple LeetCode-style questions. However let’s be trustworthy, these interview questions may be solved by AI simply right now. In my article one 12 months in the past, I evaluated how ChatGPT, Claude, and Gemini carry out in easy SQL duties, and was impressed already by all three — Claude 3.5 Sonnet even bought full factors in my take a look at. 

Let’s take one step again. For information scientists, the true coding problem right now comes from 1. Understanding the info and finding the proper tables and fields; 2. Translating your information questions into the proper question/code. In different phrases, right now’s coding interviews largely take a look at fundamental syntax, which is likely to be truthful for entry-level candidates, however have been failing to guage precise problem-solving for a very long time, even with out the evolution of AI. The truth that AI can reply them shortly solely makes this spherical much more outdated. 

So, how can we make the coding interviews extra significant? I feel, firstly, we should always permit candidates to make use of AI instruments like GitHub Copilot or Cursor throughout the coding interview to imitate the brand new work setting with AI. I’ve seen this taking place regularly within the business. For instance, Canva introduced AI-assisted coding interviews just lately, and Greenhouse also says, “We welcome clear use of generative AI within the interview course of for sure roles with the expectation that candidates can completely clarify the prompts they create and/or talk about in-depth the technical selections they make.” I feel permitting candidates to make use of AI is healthier than attempting each means to stop them from dishonest with AI, as they may use (and are anticipated to make use of) AI at work anyway :). 

In the meantime, as an alternative of asking easy SQL/Python questions, I’ve a few concepts:

1. Ideally, we might arrange an setting with a number of documented tables and ask the candidates to do a reside problem-solving session with the assistance of AI. As a substitute of asking questions like “write a question to calculate MAU since 2024”, ask extra open-ended questions like “how would you examine buyer churn since 2024?”. The analysis won’t solely be primarily based on code accuracy, but in addition on how the candidates body their evaluation and interpret the outcomes. And when the candidate interacts with the AI instrument, how do they immediate, iterate, and consider the output. Although this does make interviewers’ lives more durable — they should be very acquainted with the datasets and have the ability to comply with the candidates’ logic, ask follow-up questions, and assess the responses. 
2. Alternatively, we will ask candidates to guage the AI outputs — that is most likely simpler to arrange and fewer nerve-racking and time-consuming than the above format. Whereas AI might help with coding, it’s nonetheless people’ accountability to guage the output. Not each AI-generated code is right, even when it runs with out errors. The interviewer can describe what they’re attempting to do and present AI-generated code, then ask the candidates to determine if the logic is right, if it ignores any edge circumstances, if there’s any higher alternate options, or if the code may be optimized additional — this requires the candidate to totally perceive the best way to interprets between the enterprise logic and the code. It’s also simpler to design a typical rubric with this downside setup. 

Statistics and Machine Studying Interviews: Much less Principle, Extra Context

Subsequent, let’s speak about statistics and machine studying interviews. AI is a good trainer — it explains fundamental stats and machine studying ideas clearly and might help brainstorm totally different methodologies — strive asking ChatGPT, “clarify p-value to me like I’m 5”. Nevertheless, understanding the theories doesn’t all the time imply making use of the suitable strategies primarily based on enterprise eventualities. Yow will discover a superb instance in my Google Data Science Agent evaluation article — it does an incredible job establishing a modeling framework with useful starter code, nevertheless it requires a transparent downside assertion and a clear dataset. Human experience can also be mandatory for characteristic engineering, selecting the most effective domain-specific information science practices, and tuning the fashions. Retaining that in thoughts, I feel statistics and machine studying interviews ought to ask fewer theoretical questions or coding fashions from scratch, however combine extra with enterprise case interviews to check if the candidates can apply theories to a enterprise context. So as an alternative of asking remoted questions like “What’s the distinction between Ridge and Lasso Regression?” or “Find out how to calculate the pattern measurement for an A/B take a look at?”, current a real-world downside and observe how the candidates strategy the questions analytically, if the proposed strategies make sense, and if they impart their concepts logically. It’s not like we now not want the candidates to have stable stats and ML information, however we are going to take a look at the information extra seamlessly within the case dialogue. For instance, when going by way of a hypothetical fraud detection case, we will ask why the candidate proposes XGBoost over Random Forest, and whether it is higher to impute lacking values in family earnings because the median or zero.  

The excellent news is we’ve already seen many of those technical + enterprise case interviews within the business. My prediction is that AI will make it much more predominant.  

Behavioral & Cross-functional Interviews: Largely Unchanged, However With New Twists

For the remaining two interview sorts, behavioral interviews and cross-functional interviews, they may probably keep right here. They consider the candidates’ delicate abilities, reminiscent of cross-functional collaboration, communication, battle decision, and possession, in addition to their area information. These are the issues AI can not substitute. Nevertheless, there may very well be some shifts in what questions individuals ask. Interviewers can add questions concerning the candidates’ previous expertise with AI instruments to get extra sign on how they use AI to spice up productiveness and resolve issues. For instance, a product supervisor may ask, “How can we use AI to enhance buyer onboarding?” These conversations can floor the candidates’ means to determine AI use circumstances that drive actual enterprise worth.

Take-home Assignments: Nonetheless Controversial, However Helpful

In addition to these frequent interview codecs, there’s additionally a controversial one which comes up in information science interview loops once in a while — Take-home assignments. It’s normally within the format of offering a dataset and asking the candidates to do an evaluation or construct a mannequin. Generally there are guiding questions, typically not. Deliverables vary from a Jupyter pocket book to a sophisticated slide deck. 

I do know there are candidates who actually hate it. It takes loads of effort — although recruiters all the time say common candidates take about 4 hours, the precise time you spend is normally considerably longer, as you need to be complete and showcase your finest work. And what makes it worse is, the candidates could find yourself being rejected with out the chance to even speak to the staff — how irritating! Unsurprisingly, I heard from my staff’s recruiter some time again that take-home task results in a excessive drop-off fee within the hiring course of (so we eliminated it). 

However take-home assignments do have worth. It exams end-to-end abilities from downside framing, coding, writing, to presentation. And the character of working together with your native setting together with your most popular instruments now means you may search AI’s assist to finish the task quicker and higher! Subsequently, take-home assignments can simply evolve and grow to be extra frequent on this new period, with increased expectations for depth, interpretation, and originality. The problem, although, is for hiring managers to provide you with an task that AI can not simply resolve or will solely generate the minimal acceptable answer. For instance, a easy information manipulation process won’t be applicable, however an open-ended query that requires making assumptions primarily based on area information, tradeoff dialogue, and prioritization will work higher. And a follow-up reside interview is all the time useful to validate the understanding. 

Now let’s summarise the standard interview codecs vs. the brand new codecs beneath the AI period:

III. What This Means for Candidates

Above is my tackle how information scientist interview loops will remodel beneath the age of AI. Nevertheless, these shifts should take some time to occur, particularly at giant corporations with a standardized and well-established recruiting course of.

So, what ought to the candidates do to organize themselves higher forward of time? 

1. Know when and the best way to use AI thoughtfully. As corporations begin to permit the usage of AI and even consider how you employ AI throughout interviews, understanding the best way to use it thoughtfully turns into crucial. Don’t simply immediate and paste. You must perceive what AI does properly and the place it falls quick, and the best way to consider the outputs. To not point out that AI can also be an excellent useful instrument in interview preparation. It may well assist you to perceive the place higher, arrange a preparation plan, and do mock interviews — I can write an entire article on this (perhaps subsequent time). 
2. Perceive the enterprise deeply. Now that technical abilities are getting simpler with AI help, enterprise understanding and area information grow to be the important thing for a candidate to face out. Subsequently, everybody ought to collaborate extra with stakeholders at work to develop their enterprise information. And whenever you put together for interviews, spend time doing firm analysis to grasp its product — what can be the important thing metrics, the best way to develop the product additional with information, and what must be the retention technique. 

Thanks for studying! In case you’re a hiring supervisor, I’d love to listen to how your staff is adapting. And in case you’re a candidate, I hope this helps you put together smarter for the way forward for interviews.