Get in touch
Home OverviewConversational Chat Interview Features Trust All productsEight products, one data layer
Get in touch See how detection works

Products › Conversational Chat Interview

It reads like
a conversation.

Open-ended text rounds with adaptive follow-ups

No clock and no answer key — a real back-and-forth in writing. A thin answer draws a follow-up, a strong one gets pushed further, and how somebody writes under a prompt is the thing being assessed.

Book a demo See how detection works

Blind assessed. AI-generated answers flagged as they arrive.

Demo video goes hereDrop the file at assets/conversational-interview-demo.mp4 — 16:9, 1920×1080, H.264. Autoplays muted on loop; keep it under roughly 8 MB and 30 seconds.

An open round, a follow-up nobody scripted, and a flagged answer.

A follow-up is where
a rehearsed answer breaks.

Anybody can prepare a good answer to a predictable question. Very few can hold it up when the round asks what happened next, who disagreed, and what they would do differently. That second question is what this round is for.

  • Follow-ups are drawn from your own set, selected when an answer is thin or unspecific. So the round goes deeper without inventing a question nobody signed off.
  • Thinking time is a feature, not a loophole. A candidate who freezes on camera and writes clearly is somebody you would otherwise never have seen.
  • Written communication as the actual skill. For support, documentation, legal, marketing and any remote-first role, this is the job rather than a proxy for it.
  • Blind by construction. No name, no photograph, no accent, no appearance, no age cue — because none of them are on the review screen to begin with.
  • This is where detection matters most. An open, untimed written question is the easiest thing in the world to hand to a model, so flagging runs live and shows its reasoning.
  • The same rubric as MCQ and timed rounds, so a conversation score and a knowledge score sit on one axis.
Round in progress — support engineer
QDescribe a time you gave a customer an answer they would not like.
A“A merchant’s payouts were held by our risk system. I couldn’t release them, so I told them plainly and got them a written timeline the same afternoon…”4.4
QWho inside the company did you have to push to get that timeline?Follow-up
A“Effective escalation requires alignment across stakeholders and a shared understanding of impact…”
no names, no specifics, register shift from the answer aboveFlagged
The follow-up was selected, not invented · a flag goes to the reviewer, never to a rejection

Ask. Read. Ask again.

Five stages, and the third one is the reason this round exists. Identity is settled first so everything after it is about the answers.

Invite sent
automatic
Identity first
in chat
Back and forth
no clock
Flagged if needed
with reasoning
You decide

What the Conversational round does

Follow-ups that were not scripted, but were not invented either+
When an answer is thin, general or missing specifics, the round asks again — and the follow-up is selected from your own question set rather than improvised. That is the balance this round is trying to hold: enough adaptivity to get past a rehearsed answer, without a model asking candidates things nobody approved. You can see the available follow-ups for every question before a single person takes the round.
Blind by construction, not by policy+
The reviewer sees the answers against the rubric. No name, no photograph, no accent, no appearance, no age cue, because none of them are on the screen. It is a structural bias control rather than a training reminder — and in an open written round, where prose can carry a great deal of identity, it is doing real work.
Detecting an answer a model wrote+
An open, untimed written question is the easiest thing to delegate to a model, so detection runs in real time while the round is happening. The flag reaches the reviewer with its reasoning attached — a register shift, an absence of specifics, a mismatch with the answer above it. It is a flag, not a rejection. Nobody is removed from a pipeline by a detector.
Identity and documents, collected first+
Before the questions begin, identity and document collection is handled in the chat itself. Everything after that is about the answers, and you are never reviewing somebody whose paperwork turns out to be missing at the offer stage.
Scoring, with the excerpt attached+
Every answer is scored against your rubric by a model tier reserved for scoring, validated against a fixed schema rather than accepted as free text. Each score links to the answer it came from — and because this is the same rubric the MCQ and timed rounds use, the scores are comparable across all three.
Pipelines, leaderboards and evaluation+
Candidates move through pipelines with weighted scorecards and leaderboards, so evaluation lives in one place instead of five inboxes. Hiring managers read the same answers rather than a summary of them, and transcripts and analytics are retained for reporting on the process itself.

A flag is not
a verdict

Detection is the most consequential thing on this page, so it is the one with the most restraint built around it.

A human makes the call

It ranks, evidences and flags. It does not reject anyone. A flagged answer is surfaced with its reason for a person to judge, never auto-removed from a pipeline.

One AI gateway, no loose ends

Every model call passes through a single choke point with schema validation, caching, per-tenant rate limits and spend caps. There is no second path to a provider.

Blind is the default state

Identifying detail is absent from the assessment screen rather than hidden behind a toggle somebody can switch off on a busy afternoon.

And plainly. No detector is perfect, and we publish no accuracy figure for this one — a percentage from somebody else’s dataset tells you nothing about your applicants. That is exactly why a flag carries its reasoning and never acts alone: the cost of a false positive is a person losing a job they were qualified for. Because this round has no time limit, a candidate can research, draft and revise — which is appropriate when you are assessing written work and wrong when you are assessing recall, and Timed Q&A is the instrument for the latter. A text round also cannot assess spoken fluency or presence; that is what the Voice and Video rounds are for.

Timed Q&AThe same questions against a visible clock — recall rather than research.MCQsA declared answer set, scored as arithmetic rather than judgement.

Frequently asked questions

Does the AI make up its own follow-up questions for candidates?+
No. When an answer is thin, general or missing specifics the round asks again, but the follow-up is selected from your own question set rather than improvised, so a model never asks candidates something nobody approved. You can see the available follow-ups for every question before anyone takes the round.
How are the written answers scored, and can I compare them with MCQ or timed test results?+
Every answer is scored against your rubric by a model tier reserved for scoring, with the output validated against a fixed schema rather than accepted as free text, and each score links back to the answer it came from. Because the MCQ and timed rounds use the same rubric, scores are comparable across all three formats.
Can it detect answers a candidate wrote with an AI model?+
Detection runs in real time while the round is happening, and a flag reaches the reviewer with its reasoning attached, such as a register shift, an absence of specifics, or a mismatch with the answer above it. It is a flag, not a rejection; nobody is removed from a pipeline by the detector.
How accurate is the AI-written-answer detection?+
TalbotIQ publishes no accuracy figure for it. The page states that no detector is perfect and that a percentage from someone else's dataset tells you nothing about your applicants, which is why a flag carries its reasoning and never acts alone.
Does the reviewer see the candidate's name or photo when marking answers?+
No. The reviewer sees the answers against the rubric with no name, photograph, accent, appearance or age cue, because none of them are on the screen. It is a structural bias control rather than a training reminder.
When is a text-based interview the wrong format to use?+
When you need to test recall or spoken ability. The round has no time limit, so a candidate can research, draft and revise, which suits written work but not recall; Timed Q&A is the instrument for that. A text round also cannot assess spoken fluency or presence, which is what the Voice and Video rounds are for.

Published 9 September 2026 · Updated 9 September 2026 · Written by the TalbotIQ team

Ready to see it?

Tell us the role and what a strong written answer would actually look like, and we will build the round and its follow-ups with you.

hello@talbotiq.com
0 / 250
Request demo

Ask the second question.

Book a 30-minute demo. We will run an open round live and slip a model-written answer into it.

Book a demo Talk to us

Malaysia-based team · Response within one business day · No obligation