AI Cheating in Coding Interviews: Rounds' Anti-Cheat Solution
#Background
In order to succeed in technical interviews, candidates had to be skilled in data structures and algorithms and demonstrate that knowledge in multiple rounds. This type of skill could be honed by doing many LeetCode problems. Since the past decade, LeetCode has been at the center of preparing for coding interviews. However, there have been countless debates on whether doing LeetCode problems equates to being a great software engineer. Although doing LeetCode problems helps with enhancing problem solving skills, it does not test how well someone can use AI, how well someone can communicate their thought process and whether they can build a feature end-to-end. This is where Rounds comes in.
#How Rounds Started
The initial iteration of Rounds involved a web app that would compete with GlassDoor when it came to sharing interview experiences. However, we noticed a big gap in the software engineering interview process when tools such as Interview Coder were released. There were countless stories of people cheating on big tech interviews, only to get caught or receive offers, with people even bragging about it! These tools would allow candidates to cheat during virtual interviews where they would remain hidden to the interviewer but visible to the interviewee. Some companies went as far as to ban AI in coding interviews, but this was no different than banning internet use. Many veteran engineers used forums such as Stack Overflow or other blogs to learn about how they can enhance a feature or fix a bug. Therefore, we decided to build an assessment that would encourage the use of AI but also not allow the AI to be able to solve that assessment with one shot.
The assessment itself took inspiration from the Arc Challenge by Francois Chollet, a French software engineer and AI researcher who wanted to create a challenge that would allow humans to be able to find patterns in abstract problems. This is something that AI hasn’t been able to achieve, and if there’s an AI that could solve these types of problems in one shot, then the inventor of that AI could win one million dollars. But here was the challenge: how could we build an assessment that could support multiple languages, run different test cases and run in a container outside of the web app itself. Initially, a hackathon was done within 3 days along with an all-nighter which led to the assessment being created with support for Python as well as an internal compiler that could run Python code. Little did we know that there was something out there that could do all this work for us. Once this MVP was created, Rounds became viral which led to thousands of people taking the assessment out of curiosity.
#What Skills Does the Assessment Measure?
Judge0 was recommended by someone on LinkedIn and it was the best decision to buy both Judge0 CE and Judge0 Extra CE. Judge0 Extra CE would allow support for NumPy in Python as well as providing support for different modules across different programming languages. Judge0 became a part of the success behind the Rounds assessment, where people could use AI to help them solve the problem, but they could not one-shot the problem into an AI such as Gemini or ChatGPT 5 and get the right answer. Overall, Rounds has been able to provide a great signal to companies on how well candidates can perform on the job. This is what Rounds measures in a candidate:
- Prompting skills
- Communication skills
- Coding skills
Each problem in the assessment has a visual representation of a real-world problem, and then it allows the candidate to be able to think of ways in which they can solve that problem using all the resources available at their disposal, whether it’s using AI or going through forums online such as Stack Overflow. This allows the candidate to simulate how they would solve a problem rather than having to reverse a linked list or traverse a binary search tree. Instead of rote memorization, Rounds encourages the use of AI, allows candidates to use their imagination and provides better signals to companies on how candidates could perform on the job. This is all shown with a report that provides information such as the code submission for each problem, the video explanation for certain problems and the prompts used for each problem. There have been successful case studies where a company that makes 10 million ARR mentioned that the candidate they hired who scored a 100% on the Rounds assessment ended up performing really well on the job. Another company mentioned that they were able to find a stellar candidate who was also the highest scoring individual on the ASVAB in the state of Texas, which is a test that measures aptitudes on skills such as verbal reasoning, mathematics, science and puzzle-solving.
As more users ended up using the platform to test their skills, there needed to be better scalability. At first, the assessment couldn’t handle 50 users at once. Now, it can handle thousands of users taking it. Another fun fact is that GPT 5 scored an 18% on the assessment, while human takers scored much higher, with an average score of 42%.
#Why This Matters Now
In the age of AI, there’s always excitement on how certain tools will make things easier for the human race, such as robots doing mundane tasks, self-driving cars and security systems that can send alerts for suspicious activity. But this also comes at a cost, where some people can use AI for the wrong reasons, and for Rounds’ case, it involves cheating and then having companies take AI completely away from the interview process. We’ve recently seen how Meta is now allowing candidates to use AI during their interviews and Rounds is playing a major role in changing the interview process in the tech industry while also matching pre-vetted students to startups that have internship openings.
Thanks to Judge0, Rounds has been able to act resilient when there are thousands of users taking the assessment, while also allowing support for multiple languages such as Python, Java, C#, C++ and JavaScript. It allowed Rounds to create “AI-resistant” assessments in a way that is fun, interactive and valuable to both the candidate and the company.