Some behavioral questions are now scored by software before a person sees your answer. The usual format is an AI behavioral interview on recorded video, early in the process. You answer on camera with no one on the call. Then a tool scores the recording. You can often find out in advance. In New York City, employers must give notice before an automated tool screens you. Elsewhere, you can ask the recruiter directly. The public record also says what these tools weigh most. HireVue, a large vendor, dropped facial analysis and says it mainly scores the words you say. So prepare an answer that works as a transcript. Name the question’s skill, say what you did in the first person and give a concrete result. Tone and pacing appear to matter far less than the content. The rest of this post takes one recorded answer and rewrites it for a transcript.
An AI behavioral interview usually sits at the start of the process
The best public description of these tools comes from the vendor side. SHRM, the HR professional association, covered HireVue in a February 2021 article. At that point HireVue had hosted more than 19 million video interviews for over 700 customers. SHRM reported it was most often an automated screen at the start of hiring. High-volume employers were the main users. Candidates answer the same questions on recorded video. The software then assesses their fit for the role.
For a candidate, that means a one-way video step before the loop, not a panel conversation. The 2021 date matters, since products change. Treat the details as HireVue’s description at that time.
The software mostly scores your words, not your face
The same SHRM piece covers what changed. HireVue said it stopped facial analysis in March 2020. Its internal research found that language analysis had become more predictive. Visual analysis, in its words, “no longer significantly added value.” HireVue’s CEO, Kevin Parker, said the tool transcribes each answer to text and assesses its content. He said it doesn’t analyze accent or diction. Its chief data scientist, Lindsey Zuloaga, said speech pauses and tonality were small factors and under bias review.
These are a vendor’s claims about its own product. The same article quotes critics too. AI ethics lecturer Merve Hickok argued that natural language processing can’t yet judge an answer’s nuance or context. So don’t assume the tool reads you the way a person would. It reads a transcript. A transcript only holds what you say out loud.
In New York City, the employer has to tell you first
New York City’s Local Law 144 covers automated employment decision tools. The city’s Department of Consumer and Worker Protection enforces it. Employers can’t use such a tool unless it had a bias audit within the past year. A summary of that audit must be public. Candidates must also get notice at least 10 business days before the tool is used. If an employer skips any of those, you can file a complaint with the department.
So for a New York City role, check the careers page for an audit summary or notice. Elsewhere the rules vary. You may get no notice at all. Ask the recruiter in one line.
Is the recorded interview step reviewed by a person, scored by software, or both?
It’s a normal logistics question. The answer tells you whether to prepare for a transcript or a listener.
The example: one recorded answer, before and after
Say the screen asks, “Tell me about a time you made a decision without complete information.” Here is a first draft many engineers would give on camera.
So, yeah, this was a while back. We had an issue with the app where things were kind of slow for some users. We weren’t totally sure what was going on. We looked at a bunch of stuff. Eventually we went ahead with a change to how we load images. It seemed to help. People were happier after that.
A person might fill in the gaps from your tone and follow up. A transcript can’t. It has no named decision, no “I”, no missing information stated and no result. Here’s the same story written for a transcript.
I made a decision without complete information when our Android app’s feed got slow for some users. Our traces pointed to image loading. We couldn’t reproduce it on test devices. I decided not to wait another week for more data. I shipped a smaller image size behind a flag to 10 percent of users. My reasoning was that the change was easy to roll back. Within three days, slow feed loads in that group fell by about half. We rolled it out to everyone.
The second version opens by echoing the question’s skill. It says “I decided” and names what was unknown. It gives the reasoning and a number. None of it depends on how you sound. The numbers here are illustrative, so use your own. The first person matters with a human listener too. Our post on saying “I” instead of “we” covers why.
How to practice for a recorded screen
Record yourself answering three likely questions, then read the auto-generated captions instead of watching the video. The captions are closer to what a scoring tool works from. Check each transcript against four things.
- The first sentence names the skill the question asked about.
- The decision or action is stated with “I” and a verb.
- The result has a number, a date or a named outcome.
- Filler like “kind of” and “a bunch of stuff” is gone, because it adds words without facts.
Keep each answer within the time the platform allows. Keep delivery steady and clear, because a human may still review the recording later. But spend most of your practice on content, since that’s what the vendor says it scores.