UPDATE 2 (October 8, 2025): Just in! Google has wrapped up its pilot of onsite interviews for software engineers at its largest sites (Bay Area, Seattle, New York, Poland, Bangalore). The format appears to be: 2 virtual interviews, then 3–4 onsite. The plan is to “actively” roll this out for all engineering roles over the next 2–3 months.
UPDATE: This piece was also covered in The Pragmatic Engineer newsletter.
A year and a half ago we published “How Easy Is It to Cheat on Technical Interviews with ChatGPT?” In it we shared our experimental findings: if companies ask LeetCode problems verbatim (or even with small tweaks), gaming the system is fairly easy. Fully custom questions proved nearly cheat-proof.
Based on that, we predicted that technical interviewing would have to change, and that companies would be forced to move away from cookie-cutter LeetCode problems toward more personalized questions. Even with that prediction, we were confident that—even if the content and format shifted—algorithmic interviews were not going away.
Now we are starting to see some movement, and technical interviews have drawn a lot of press lately. As is often the case with the media, the headlines are mostly clickbait. They hint that companies (Meta and Anthropic in particular) now allow AI in interviews. Naturally, many engineers hopefully read those headlines as a sign that algorithmic interviews are on the way out… but is that actually true?
And what else is changing? Will we return to onsite interviews? Will questions get harder? How widespread is cheating, and what are companies doing about it? If candidates are allowed to use AI, what will the new interview formats look like? And how does all of this differ between FAANG and FAANG-adjacent companies versus startups?
And finally, the most important question: can we say, “We told you so!”?
To unpack all of this, we surveyed our interviewers (all from FAANG and adjacent companies) and some of our users (who interview at their own companies). Respondents mostly represented FAANG, as well as companies like Stripe and Uber.
In addition to the survey, we spoke with Amanda Richardson (CEO of CoderPad) and Ilya Tilis (VP of Product at Codility) to understand how their customers are integrating AI into interviews and assessments.
What is changing at FAANG and nearby?
At FAANG—especially Meta—some things really are changing. But at FAANG-adjacent companies, almost nothing! Here is how a few of our interviewers from those companies described the situation (though most simply answered “Not applicable”) when we asked how AI had affected their processes:
“My company moves slowly, despite numerous cheating incidents and other problems.”
“Honestly, little has changed from the interviewer's side. I watch for whether the candidate is cheating, but that's about it.”
That is surprising, because FAANG-adjacent companies are smaller and, at least in theory, more agile than their “giant” siblings. But like it or not, FAANG sets the tone and structure of technical interviews in our industry, and everyone else follows. In the 1990s it was Microsoft. Today it is probably Meta (we will explain later why Meta will lead on bringing AI into technical interviews). As one of our interviewers put it:
“I believe the precedent Meta creates by allowing AI in interviews will be adopted by other large tech companies as well.”
So what is changing at FAANG?
- Algorithmic questions are not going away (at least not yet).
- Cheating (or at least the perception of it) is a real problem.
- Although algorithmic questions remain, many companies have changed the questions themselves, and many interviewers are changing HOW they ask them.
- The most visible process change at FAANG is the rollout of cheating-detection systems (but not every FAANG company uses them).
- Meta will launch an AI-assisted interview as part of the onsite loop next quarter, but it will be part of the onsite stage and will NOT replace the algorithmic phone screen.
- It is unclear whether onsite interviews will return.
Startups are a different story entirely. They are making all kinds of AI-driven changes—to questions and to processes. We cover startups at the end of this post.
Although there are not many changes at FAANG+ yet, the ones that exist are encouraging. Years ago, a Twitch interviewer said something that became our favorite quote about interviewing: “The interviewer's job is to figure out whether they and the candidate can be ‘smart together.’”
That is it. That is the essence of a good interview.
These changes will hopefully put our industry back on a healthier path—away from the arms race in which candidates memorize ever more LeetCode problems while companies raise the bar expecting instant recall of those answers.
Now let's dig into the details!
Methodology
We sent a survey to our interviewers and a subset of users. We received 67 responses, mostly from FAANG and FAANG-adjacent companies, plus a few startups.
In the survey we asked how processes and questions at their companies had changed with the rise of AI. We also asked about the prevalence of cheating and expected changes going forward.
We instructed respondents to answer solely based on their own experience—not on rumors or what they had seen in the press.
Algorithmic questions remain—for now
First, the main finding. Of 52 FAANG respondents, NOT A SINGLE ONE said their company had dropped algorithmic questions.
Still, more than half of respondents believe that within 2–5 years algorithmic interviews will no longer be as prominent as they are today. About 20%, however, are confident they will never disappear entirely.
Cheating is real, and some companies are investing in detection
About a third of FAANG interviewers said they had actually caught someone cheating during an interview, and 81% said they suspect candidates of using AI to cheat.
Are companies responding by rolling out detection systems? Yes. When asked whether AI had changed the interview process at their companies, the only change interviewers named was the introduction of cheating-detection tools. That said, it is still not widespread—only 11% of FAANG interviewers said their companies use such software. And nearly all of those interviewers (except one from Microsoft) were from Meta.
Meta really does seem to be betting on fighting cheating. As one of our Meta interviewers put it:
“Fighting cheating is front and center at Meta right now. We now have to flag whether we suspect a candidate of cheating on virtually every interview type and level (previously this was coding only) and provide a rationale. We also require candidates to share their entire screen and turn off all background filters (including blur) in most interviews.”
We also spoke with Amanda Richardson, CEO of CoderPad, for more context on cheating-detection tools. CoderPad is clearly a leader in technical-interview tooling, and its customers include Meta, individual Microsoft teams, and many FAANG-adjacent companies. Amanda confirmed that CoderPad has enabled cheating detection by default for all customers (monitoring suspicious actions such as copy/paste or switching browser tabs).
Despite that, NOT A SINGLE ONE of the surveyed FAANG-adjacent engineers said they use anti-cheating tools. One even noted:
“My company moves slowly, despite numerous cheating incidents.”
So where is the disconnect? Maybe interviewers simply do not know detection tools exist, or they turn them off—but more likely it is still too early for a consistent policy across every team, especially in high-chaos companies.
Although we expect cheating detection to become more standard in the coming months, the main way FAANG+ fights cheating is by adapting questions and changing how they are asked.
Questions are changing (and may be getting harder)
Even though algorithmic questions have not gone away, 58% of FAANG interviewers said they have adjusted the types of algorithmic questions they ask. Here is what changed, in the interviewers' own words:
Meta
“I treat AI use like memorizing LeetCode problems—the skill I look for in an interview is not recalling a popular question, so a small tweak to a popular question is a very easy way to catch that. More open-ended questions that test thinking rather than applying a known pattern. The question description has gotten much longer; the setup is harder.”
Microsoft
“We choose questions that require fluent explanation of concepts, and harder coding tasks where LLM solutions often contain ‘tells.’ Variants of LeetCode problems. Or checking whether a candidate can extend existing logic using a class structure instead of rewriting everything. More focus on systems thinking and more complex/practical implementation tasks deliberately designed to elicit overly specific/rare and incorrect answers from LLMs.”
Amazon
“I create problem variants to foil potential AI use (and also to bypass people who memorize problems). Less copying from LeetCode, more thinking from the same concept. Whatever I pick will not be a straight lift from LeetCode.”
“No LeetCode-style problems; instead, multi-subproblem tasks that require a lot of problem-solving rather than a lot of coding. Different terminology and levels of abstraction that AI tools were not trained on. Likely requiring a combination of 2+ data structures/algorithms. Focused on more abstract ideas that are not tied to a single answer. Problems with multiple solutions.”
Roblox
“Because you can instantly get an answer to most LeetCode questions, we make sure our question cannot be solved by a standard chatbot.”
In our view, anything that moves the industry away from having to memorize questions is very good. Of course, companies are improving processes not because it is the right thing, but out of necessity—and cheating is becoming the forcing function.
Unfortunately, many companies still take a LeetCode problem and tweak it lightly. In our cheating study a year and a half ago, we saw that small changes are not enough—there was no meaningful difference between candidates who cheated on verbatim versus lightly modified LeetCode questions.
If you truly want your questions to be cheat-resistant, you need to put in the work and create your own. We cover how to do that in the appendix at the end of this article.
At this point you might wonder: are the questions getting harder? Judging by the answers, “not really.” Only 21% of FAANG interviewers admitted they deliberately ask harder questions. But even if the questions themselves are not getting harder, our data shows that the bar for passing interviews has indeed risen over the past few years.
Interviewers are also changing how they ask questions
One of the subtler changes we noticed in the survey was not which questions people ask, but how they ask them. About a third of surveyed FAANG+ interviewers noted this.
Here is what they say:
Meta
“I pay attention to patterns in the candidate's approach. AI has a certain pattern, and if a solution looks AI-generated, I start asking follow-ups. When evaluating a candidate, I look not only at correctness but at whether they understand what they are doing. Much more attention to WHY, not HOW. I also try to ask follow-ups with different parameters/constraints to quickly test adaptability.”
Microsoft
“I ask more foundational clarifying questions about the basic data structures used in the problem, to check whether the candidate knows the data structure or just knows how to apply it. More detailed questions, e.g.: ‘What does this line of code do?’, ‘What happens to this function if we remove this line?’, and ‘non-standard’ questions like ‘Can you do this differently?’
Amazon
“Deep understanding of topics/subjects, not just knowing facts. More focus on probing to assess understanding and expanding questions.”
Netflix
“I ask more clarifying questions. Sometimes I go back to earlier questions and ask them again, rephrased, after 5 minutes to see whether the answer changes.”
Atlassian
“I think it is not about which question is asked, but about how you ask follow-ups or push the candidate to explain their thinking/code.”
DoorDash
“Faster pace. I do not paste the full question for the candidate.”
Like the changes to the questions themselves, these shifts in interviewer behavior are a very encouraging sign. They point to deeper interviewer engagement, probing questions, and more collaboration with the candidate.
What is Meta's new AI interview? And will it replace the algorithmic interview?
At the start of this post we said Meta would likely lead in defining what AI-assisted interviews look like. As we have already seen, Meta appears to be the only FAANG company that has fully invested in cheating detection. And as far as we know, Meta is the only FAANG company experimenting with AI interviews.
Several interviewers mentioned that Meta is testing AI interviews, and that was also the subject of a WIRED article we referenced earlier. Will this new AI interview replace the algorithmic interview—or complement it?
Here is what you need to know:
- This new AI interview is currently being tested at Meta.
- It is planned to launch for some candidates this/next quarter.
- It does NOT REPLACE the algorithmic phone screen. Rather, it will become part of the onsite loop, possibly replacing one of the current interviews in that cycle.
As one of our Meta interviewers put it:
“Meta is actively testing this and plans to launch in Q3, so it is not far off. That does not mean AI interviews will run for every role/candidate by the end of this quarter, but they will start being used with real candidates by then… Like most things at Meta, it will likely take the form of ‘start small and iterate fast.’ I think that most likely means… an onsite-loop interview for a subset of candidates, and then we see what comes next based on results and feedback.”
Ilya from Codility:
“Based on our conversations with technology leaders, most treat AI skills as an additional stage in the interview process, not as a replacement at every stage.”
And as several of our interviewers said:
Microsoft: “AI is hovering over interviews, but not materially yet.” Google: “AI killing LeetCode is overhyped.”
Depending on Meta's experiment results, we expect other FAANG+ companies to follow, though some may take longer—and we see no evidence that AI interviews will replace algorithmic ones.
Returning to onsite interviews? Maybe…
50% of FAANG interviewers predict companies will soon return to onsite interviews, versus only 29% among FAANG-adjacent. Interviewers from the same company often disagreed, though. For example, although 55% of surveyed Meta interviewers think Meta will soon return to onsites, one of them said:
“We tried [onsite] interviews at Meta again, and it was a complete failure. Unlikely we will go back.”
Similarly, although half of Amazon interviewers think onsites will return soon, some explicitly noted that Amazon has not tried it yet:
“I expect interviews to return offline, but Amazon has shown no signs of that.”
“Amazon has not considered this internally yet. We need to do it as soon as possible.”
Here is a table showing what share of interviewers from each FAANG company expect a return to onsite interviews:
| Company | Share expecting a return to onsite interviews |
|---|---|
| Amazon | 50% |
| Apple | 100% |
| 67% | |
| Meta | 55% |
| Microsoft | 78% |
| Netflix | 50% |
Judging by these quotes and numbers, we expect that—except possibly Apple, and despite whatever Sundar Pichai may have said on a podcast—there has been no top-down mandate on this at FAANG.
And at FAANG-adjacent companies, we believe a return to onsite interviews is not planned for now.
Startups are a different story—and they are adopting AI in interviews fast
Although our survey was mainly aimed at FAANG interviewers, we did get a few responses from startup engineers, and they told a very different story.
67% of startup respondents said AI had substantially changed the interview process at their company, versus 0% at FAANG and FAANG-adjacent. Those changes included:
- Dropping algorithmic questions
- Adding an AI interview to the onsite loop
- Dropping take-homes (which makes sense, since take-homes are far more common at startups than at FAANG)
- Replacing the phone screen with an AI-assisted one
Amanda from CoderPad:
“We are watching the rapid death of algorithmic take-homes. It is over. But I think we are now seeing people replace them with something like a project-based take-home.”
So what is next? Both CoderPad and Codility say true AI interviews are still early, but they are already emerging and in demand at a growing number of companies.
According to Amanda from CoderPad, about 20–30% of customers are already actively using AI in their interview processes: “where candidates use AI to understand a codebase… debug with it and write code the way many people imagine programming in the future.” CoderPad customers have run more than 35,000 AI-assisted interviews.
And according to Ilya from Codility: “Usage of our built-in AI assistant has accelerated significantly over the past couple of quarters… It is embedded directly in our VS Code environment and includes capabilities developers are used to seeing in their day-to-day tools.”
Have AI interviews gotten easier?
One of the questions we asked in the survey: “Do you think AI assistance in coding interviews lets weaker candidates pass interviews they would otherwise fail?” 75% answered “yes.”
And judging by the reaction to news that Meta might bring AI into its interviews, many candidates also believe using AI will give them an edge.
But before building too many assumptions, remember your first open-book exam. You were probably thrilled and relieved—at first. Then, if your experience was like ours, that relief turned into real horror when you saw the exam itself and realized that nothing—and certainly not an open book—would save you.
When we discussed with Business Insider a few months ago whether companies should let candidates use AI in interviews, we used the open-book versus closed-book exam analogy, and that analogy made the article's headline. It fits because open-book exams are fundamentally different: they no longer require reproducing memorized knowledge. They require synthesizing information you were already supposed to understand and using it to solve hard, novel problems.
Open-book exams do not lower the bar. They assess something else… and possibly something harder.
Here is what Amanda from CoderPad says:
“AI tasks are a 1,000–2,000-line codebase, and you need to add a feature in a short time. Open book maps better to the job, but it makes interviews harder and impossible without AI.”
And Ilya from Codility:
“Imagine an interview where the task is not simply to solve a small isolated problem. Instead, the candidate is dropped into a large, unfamiliar codebase and asked—using integrated AI tools—to complete a realistic task. That might be implementing a new feature, fixing a list of security vulnerabilities, or refactoring a service for maintainability.”
So to everyone prematurely celebrating this new interview type, we want to recall Amanda's excellent line:
“AI interviews are an ‘add-on,’ not a replacement. I cannot imagine a team saying, ‘We don't need to know whether you can code.’ That still matters.”
Conclusion
So where does that leave us? At the start of this post we wondered whether we could say, “We told you so!”
This time—unlike our forecasts about when hiring would rebound—we think we got it right. Algorithmic interviews are not going away, but AI-fueled cheating has become a forcing function for better questions and better interview mechanics.
Looking ahead, we expect AI interviews to become part of every interview loop without replacing it wholesale—just as open-book and closed-book exams serve different purposes.
And what does all of this mean for interviewer quality? After all, the main reason technical interviews have a bad reputation is not the questions—it is the interviewers. We would love to say that the new focus on deep engagement with candidates, checking understanding, and follow-up questions will push companies to invest more carefully in both interviewer training and incentives—because if you want engineers to become better interviewers, you have to reward that behavior rather than treating interviews as an interruption between code commits.
But after many years in this industry, we are not sure companies will make those investments—at least not until some other future forcing function makes it unavoidable. Still, hopefully someday the image of the disengaged interviewer (sitting in another tab waiting for the candidate to reproduce a perfect answer) will be a relic of the past.
Appendix: How to invent good (if not cheat-proof) custom questions
- Do not just re-skin old problems – Simply changing the wording or context of a LeetCode problem is not enough. You need unique inputs and outputs.
- Keep a shared team problem journal – Create a shared document where the team logs interesting problems they have solved. Those notes become the foundation for interview questions.
- Adapt real problems to interview format – Strip excess complexity, focus on the core of the solution, and remove company-specific context.
- Plan for iteration – Test questions internally and improve them based on feedback.
- Do not be afraid to keep a DS&A component – Data structures and algorithms remain a powerful evaluation tool in the hands of good interviewers.