How a Text Chat Interview Scores What You Type

What Does a Chat Interview Actually Read?

The content of the answers is what gets read, on the two pages that describe it at all. Vervoe's candidate help page says its AI evaluates what was written, how well it aligns with the task, and how the answer was reached ¹. The same page says the AI has no access to your age, gender, race, education or work history ¹. Sapia says every candidate is assessed on the same set of attributes drawn from their responses ².

Vervoe and Sapia are describing their own products on their own help pages. Neither page publishes the scoring model, the attribute list, or the report the employer sees at the end of it.

A product description is not an audit.

Sapia says the chat interview returns a personality profile and a coaching tip at the end. The page says that document is for the candidate's eyes only and is not shared with the prospective employer ². Read that as a description of one document rather than of the whole exchange. It does not mean the employer receives nothing, and no page says what the employer's own report contains or which attributes reached it.

The research underneath has a gap neither page mentions. The 2022 meta-analysis that put structured interviews at the top of its table of selection methods was built from interviews humans conducted and humans rated ³. Its authors note that as administration and scoring move toward automation, validity has to be evaluated under those changed conditions ³.

That is an open question rather than a verdict, and it cuts in both directions. The paper shows neither that a text-scored interview works as well as a human-rated structured one nor that it works worse.

So the useful posture is neither trust nor dread. Read what the vendor publishes, notice what it declines to publish, and then spend your attention on the part you control, which is what the answers say. What an AI interview scores and what it does not covers the same ground across the video formats.

How Long Should Each Answer Be?

Sapia publishes a range: five to seven text questions, mostly behavioural, with an instruction to use between 50 and 150 words on each one ². The same page says the interview is not timed, that nobody is penalised for how long they take, and that an emailed link lets you save your progress and come back to it ².

Read the range as a shape rather than a quota. One situation, described properly, lands inside it without any counting.

Employers running a written screen are told something close to the same thing. The employer-side piece on what to ask for on an application instead of a cover letter caps the answer at 150 words and asks for one checkable number, date or name. It also tells the reviewer to score every candidate on the same three points, and never to mark an answer down for reading like AI.

That last instruction is worth sitting with, because it puts the weight on what an answer contains rather than on how it was produced.

Whether length itself is scored, the page does not say.

One caution on the format: it belongs to a single vendor. Other text screeners may put a clock on each answer, and a hiring chatbot that texts you knockout questions is a different thing again. Read the invitation before you assume the evening is yours, and if it says nothing, the recruiter who sent it can tell you.

Write the Situation Only You Were In

Specifics are the part a fixed attribute set has something to grab. Name the situation, the decision that was yours, what you did about it, and how it turned out, in that order, and the answer stops being interchangeable with the one after it. Vervoe describes its AI as reading how an answer was reached, not only what it concluded ¹.

• one situation per question, dated if you can date it
• the decision you actually made, and what you gave up to make it
• what happened next, including the number if a real one exists
• the part you would do differently now

There is a measured ceiling on what word categories carry about personality, and it sits lower than an end-of-interview personality profile implies. A 2022 meta-analysis in Psychological Bulletin pooled 31 samples covering 85,724 people and found that the 52 word categories of LIWC, the dictionary tool some academic interview-scoring models use, explained on average 5.1% of the variance in a self-reported personality trait ⁴.

That was general writing scored with a public word-count dictionary rather than any vendor's own model, and it says nothing about job performance. A proprietary system may well read richer features than word counts, and none of them publish what those are. What the number bounds is confidence rather than accuracy: the best-documented linguistic signal is weak, and nothing published on either page tells you how much weight any reading of your words carries in a decision.

So choose the example carefully and stop auditioning adjectives.

Spelling is the other question this format raises. Neither vendor page publishes a rubric line for it, and that silence is not a promise that a typo costs nothing. Read the answer through once, fix what you find, and send it.

Should You Draft the Answers With ChatGPT?

Not where the vendor rules it out, and Sapia's page does: it tells candidates not to copy, and says the chat interview checks for plagiarism from internet sources and can detect whether answers came from a content generator such as ChatGPT ². That is the vendor's stated position. It arrives with no method, no accuracy figure, and no account of what happens to an answer it flags.

Independent work on text detectors does not support reading a flag as a verdict about a person. Seven widely used detectors run over 91 human-written TOEFL essays by non-native English speakers produced an average false-positive rate of 61.3% ⁵.

Those were academic essays and detector versions as they stood in 2023, and that figure is an average across seven tools rather than any single tool's rate ⁵. What the study shows, and no later detector has been shown to fix, is the direction of the error: it lands on people writing in a second language.

What a detector produces is a suspicion.

There is a plainer trap as well. Vervoe compares the first 500 words of every text answer across candidates and flags duplicates ⁶, so the polished paragraph lifted from a forum is the one most likely to arrive twice in the same pile.

The reason to write your own answers does not rest on any of that. A generated answer is fluent and interchangeable, and interchangeable is the quality a written screen exists to sort out. Whether to write a text interview's answers with ChatGPT takes that question on its own terms.

Nobody else can write the afternoon you actually had.

See how it works >