Is the AI Interviewer Adaptive or Reading a Script?

The most useful thing I ever did in a competitor’s AI interview demo took four seconds. I answered a question about my experience with the single word “some,” and then stopped talking.

A well built interviewer treats that as an opening. A scripted one treats it as a completed answer and moves on. This one moved on. The entire evaluation of that vendor was effectively finished at that moment, twenty minutes before the call ended.

If you are running an adaptive vs scripted AI interviewer test on a shortlist of vendors, you do not need a rubric or a spreadsheet. You need four deliberate things to say, and the discipline to actually say them while a salesperson is watching.

Why this is the question that decides everything else

Nielsen Norman Group’s research on AI moderated interviews puts the problem plainly: many AI interviewers cannot meaningfully adapt, in that they do not create new questions, reorder them, skip irrelevant ones, or adjust when a response opens something worth chasing. They march the list.

Candidates have noticed before buyers have. Greenhouse’s 2026 research found that 63% of job seekers have now sat an AI interview and most have not had a good one, which is a striking result for a category whose entire pitch is that the conversation feels natural. A scripted interviewer is a large part of why.

That limitation is invisible in a feature comparison, because “adaptive follow up questions” appears on nearly every vendor page in this category. It is also the single property that determines whether the interview measures anything.

Here is the mechanism. A fixed question list is knowable in advance. In Indian campus hiring, where a batch of 400 students from one college sits the same drive across two days, the list is knowable by lunchtime on day one. It circulates in a WhatsApp group. By the afternoon session you are no longer screening for communication, you are screening for who saw the message.

Add generative tools to that and the problem compounds. A predictable prompt produces a polished, rehearsed, entirely uninformative answer. We wrote about the candidate side of this in detecting rehearsed answers. This post is the mirror image: not how to catch a scripted candidate, but how to catch a scripted interviewer, which is the failure that makes the first problem unsolvable.

The four things to say

Do these in one demo call. Each takes under a minute and each isolates a different capability.

One: give an incomplete answer on purpose. Pick a question about your own experience and answer it in six words with no specifics. Then stop. What you are listening for is whether the next thing you hear names the gap. A real follow up sounds like “you mentioned a hiring process, what was your role in it specifically.” A canned one sounds like “thank you, could you tell me more,” which is a generic prompt wearing the costume of a follow up, or it sounds like the next question on the list.

The tell is specificity. A generic prompt could follow any answer. A genuine probe could only follow yours.

Two: contradict yourself. Say something early in the call, then say something incompatible with it later. Mention you run campus drives, then later describe hiring only senior engineers. A system that is building any model of the conversation has an opportunity here. A system replaying a list will sail past it without noticing, which tells you the evaluation at the end is scoring each answer in isolation rather than reading the conversation as a whole.

Three: answer a question with a question. Ask the interviewer to clarify what it means. Real candidates do this constantly, especially nervous ones, and it is a common failure point. Watch whether the system rephrases usefully, repeats the identical sentence verbatim, or treats your question as your answer and scores it.

Four: go quiet for fifteen seconds mid answer. This one is not about adaptation, it is about whether the system was built by people who have listened to real screening calls. Freshers pause. They restart sentences. They think out loud and then go silent. An interviewer that fills every pause, or ends the call, or repeats itself in a loop is going to do that to a twenty one year old in their first interview, and your completion rate will show it.

What adaptation should not mean

There is a version of adaptive that is worse than scripted, and it is worth naming so a vendor cannot sell it to you.

If an AI interviewer generates its questions freely per candidate, you lose comparability. Two candidates for the same role get different conversations, different difficulty, and scores that cannot be ranked against each other. That is not a better interview, it is an unstructured one, and the research on interview validity has been clear for decades that unstructured beats nothing and loses badly to structured. We covered that literature in structured vs unstructured phone screens.

What you want is a specific shape: a fixed structure with an adaptive layer underneath it.

In our own system that means the core questions come from a structured document built per job description, so every candidate for a role faces the same rubric and the same required areas. Inside each of those areas, the follow up probes are built from whatever the candidate just gave you. Comparability at the top, responsiveness underneath.

The question to ask a vendor is therefore not “is it adaptive.” It is “what stays constant across every candidate for a role, and what varies.” A vendor who cannot draw that line clearly has not thought about comparability, and their shortlists will not be rankable.

There is a way to check the answer rather than take it. Ask for two full transcripts from the same role, ideally one strong candidate and one weak one, and lay the required areas side by side. Every area the rubric claims to assess should appear in both, in the same order, in recognisably the same words. What should differ is everything underneath: the probes, their number, and how far each one pushed. If the two transcripts diverge at the level of the areas themselves, the scores were never comparable and the ranking they produced is decorative.

The disqualifier case, where adaptation matters most

There is one place where the interviewer must deviate from the script, and it is the place most tools do not.

Some answers should end the conversation. A role requires a two wheeler and a licence; the candidate has neither. A role is on site in Pune; the candidate cannot relocate. There is no version of the next twelve minutes that changes the outcome, and running them anyway wastes the candidate’s evening and your interview capacity simultaneously.

Our screens put those disqualifier questions first, in a phase that runs before the main interview. Failing one ends the call early, typically inside the first minute or two, with a reject verdict carrying a tag that names the failed condition, so nobody in the review queue has to reverse engineer it from a low number. Without that tag an early exit is simply unexplained, and unexplained is what you cannot defend. We wrote the candidate’s side of that experience in what happens after a knockout question ends your call.

Ask any vendor what happens when a candidate fails a hard requirement in question two. Should the reply be that it simply runs to the end, they do not have an adaptive interviewer. They have a questionnaire with a voice on top, and you are paying for eleven minutes of compute per disqualified candidate.

What this costs when you get it wrong

The arithmetic is worth doing before a season rather than after.

Take a drive of 1,200 applicants where 20% fail a hard requirement. That is 240 candidates. On a scripted system that runs every interview to completion, those 240 people each spend fifteen minutes in a conversation with a predetermined outcome, and each occupies an interview slot that a real candidate needed. Interview capacity is finite. With a default ceiling of 20 candidates in every 15 minute slot, those 240 pointless conversations eat three hours of a drive.

The candidate experience cost is larger and harder to see. Every one of those 240 people finishes an interview believing they were assessed, then receives a rejection that had nothing to do with anything they said. That is the sort of thing that reaches a campus WhatsApp group with your company’s name attached.

If it is scripted and you have to buy it anyway

Sometimes the decision is not yours. The contract is signed, the tool is embedded in an ATS you cannot leave this year, or the only vendor who will support your volume happens to run a fixed list. That situation is common enough to deserve an answer better than “you are stuck.”

Three mitigations, in descending order of how much they help.

Rotate the question set per session, not per drive. If the tool lets you configure question banks, build three sets for the same role and assign them by session rather than by campaign. The morning batch and the afternoon batch on day one should not face the same list. This does not make the interviewer adaptive, it just makes the list less shareable, and shareability is the specific harm you are managing. It costs you some comparability across sets, so keep the required areas identical and vary only the phrasing and order.

Compress the window. A fixed list leaks over time, so the fix is to give it less time. A drive that runs 400 candidates across five days leaks by day two. The same 400 across one evening largely does not. This is one of the underrated arguments for running screening asynchronously at high concurrency rather than spreading it: the compression is a fairness mechanism, not only a speed one. We worked through the capacity arithmetic of that in running a 3,000 candidate drive in one evening.

Put the unfakeable questions first. Some questions cannot be usefully rehearsed because the answer is a fact about the candidate rather than a performance: do you hold a licence, can you relocate, when can you join, are you willing to work a rotational shift. Those belong at the top regardless of whether your tool supports a formal knockout phase, because knowing the answer in advance does not help a candidate who does not have the licence.

What none of these fix is the follow up layer. A candidate who gives a thin answer will still not be probed, and the resulting score will still measure preparation more than capability. Budget for a heavier human second round accordingly, and be honest in your internal reporting that the first round is filtering rather than assessing.

How to run this test without being obvious about it

One practical note, because vendors are not stupid.

If you announce that you are testing adaptivity, a salesperson will pick the demo path that shows it best. Do the four things quietly, inside a normal sounding conversation, and do them in the vendor’s self serve demo before any human from the company is involved if one exists. Ours sits at hireqwik.in/ai-interview-demo, takes about five minutes, and nobody watches you take it. That is deliberate, and the broader reasoning is in why we let you be the candidate first.

Whatever you learn, write it down immediately after the call while the specific exchanges are fresh, then compare it against what the scorecard says happened. A gap between the two is itself a finding, and reading that scorecard critically is its own exercise.

The take

Adaptive is the most oversold word in AI screening and the easiest to test. It costs you four sentences in a demo you were going to sit through anyway.

Say something vague and stop talking. If the machine does not notice, nothing else on the feature list matters, because every candidate who works out that the list is fixed will beat it, and every candidate who does not will be measured against people who did.