What a Sapia.ai Chat Interview Records and Scores

If an employer has sent you a Sapia.ai chat interview, you are about to take an interview that records nothing but your typed text. There is no webcam, no screen recording, no lockdown browser, and no timer. That removes most of the surveillance a candidate expects from an online assessment and replaces it with a single integrity surface: the words you type, checked as you type them for plagiarism and for signs of generative-AI authorship. This guide explains what the platform captures, what it infers from your writing, how its AI-content flag works, and who ends up holding the transcript.

Sapia.ai (formerly PredictiveHire) is an Australian hiring platform whose main product is a structured text interview used mostly for volume hiring in retail, hospitality, contact centres, and frontline roles. It is a different category of tool from the coding assessments and video interviews covered elsewhere on this blog, and it fails or flags candidates for different reasons.

What the chat interview actually is

Sapia's candidate explainer describes the format as "a text-based interview and assessment which allows prospective employers to interview every applicant." The mechanics it documents, as of September 2026:

• You are asked "between 5-7 text-based questions," which are behavioural rather than technical.
• The interview "is not timed and you are not penalized for how long it takes."
• You should "aim for your responses to be between 50 and 100 words per question," with a hard ceiling of "500 words in each of your responses."
• On cameras and microphones, the page is unambiguous: "No and no! There is no video recording. There is no camera used."

Sapia states on its chat interview platform page that the chat interview is available in English, Spanish, French, Danish, Dutch, Finnish, German, Italian, and Swedish. It is typically taken on a phone — in Sapia's own Woolworths case study, the company reports that half of candidates completed their interviews on mobile.

What it records

The recorded artefact is your text and the metadata around submitting it. Sapia's candidate page states that "all you're assessed on is your written answers to the questions, and nothing more."

Sapia's privacy policy, last updated 16 July 2024, describes the broader collection as "interview responses, opinions, answers to questions, and any specific information requested about you by a Client (or us)." The same policy notes that the platform can carry "sensitive information" such as "information about your age, health status, background, racial or ethnic origin," and states that "sensitive questions asked via our Platform are always optional." Those are usually separate diversity-monitoring questions, not the interview itself.

What is not recorded matters as much. Because there is no camera, no screen share, and no secure-browser install, none of the device-level signals that drive integrity flags on proctored coding platforms exist here: no monitor count, no tab-switch counter, no keystroke-timing profile. The absence of those signals is precisely why the textual check carries so much weight.

What it scores

Sapia's platform page states the chat interview assesses "soft skills, behavior traits, and overall cognitive ability" from text responses. Its research arm describes the underlying method as inference from language: "lexical analysis of language helps us to discover and understand personality traits," aimed at identifying "personality traits, behavioral competencies, and language skills."

The company does not publish, on those pages, the specific trait framework, validity coefficients, or scoring thresholds behind that inference. Treat the mechanism as documented at a high level and unverified in detail. What the vendor does commit to publicly is the output shape: employers receive a ranked shortlist, and every candidate receives an automated personality summary. Sapia's candidate page states you will get "an email with your My Insights personality profile, and a coaching tip to help you with your job search," regardless of outcome.

One design consequence is worth stating plainly. Because scoring runs on language, your writing register is part of the assessment, not a neutral wrapper around it. Short, clipped answers of a dozen words give the model very little to work with. That is the practical reason for the 50–100 word guidance, and it is a more useful instruction than it first appears.

How it flags AI-written answers

This is the part candidates most often get wrong, because the interview feels unmonitored. It is not.

Sapia's candidate explainer states directly: "Don't copy answers! Chat Interview checks for plagiarism from various internet sources, and can detect if answers come from content generators like ChatGPT."

The company announced the underlying model on 21 June 2023, describing it as detecting responses "sourced by generative AI tools like ChatGPT in real-time." Sapia reports the model achieved "a ROC-AUC of over 95%" against answers generated by GPT-2, GPT-3, and ChatGPT across multiple role types. That figure is vendor-reported against models that are now several generations old, and we have seen no independent replication of it; treat it as a statement of intent rather than a measured success rate against current systems.

The workflow around the flag is documented more usefully than the accuracy claim. Sapia states that "candidates will be alerted in real-time as they respond when their answers are likely to be AI generated content (AGC), giving them an opportunity to change it ahead of the final submission," and that the system "will then flag to the decision-maker the likely inclusion of AGC in the candidate's response for further review." So the flag is not a silent disqualification: you are warned before you submit, and a human sees the flag afterwards.

What this means if your writing gets flagged

Two things follow from a text-only classifier, and both cut against candidates who did nothing wrong.

First, this class of detector keys on surface features of writing — regularity, hedging, uniform sentence structure, generic phrasing. People who write formally by habit, who draft and then heavily polish, or who compose in a second language and smooth the result can produce text that resembles the pattern. No public evidence establishes a false-positive rate for Sapia's model on those populations, and the company has not published one.

Second, because the alert fires before submission, a flag you ignore is a flag you chose to submit. If you see one, rewrite the answer in looser, more specific language: name the actual shift, the actual customer, the actual thing that went wrong. Concrete detail is both what the scoring wants and what a generic-text classifier is least likely to flag.

Who holds the transcript afterwards

Under EU and UK data protection law, Sapia's privacy policy states the employer is in charge, not Sapia: "Our Clients will remain the data controllers of your personal data — they are responsible for how it is collected and used," with Sapia acting as processor. Retention follows from that: Sapia says it will "retain your personal data for the period set by our Client, the data controller." There is no single retention window you can look up; it is whatever the employer configured.

The policy also states that Sapia "use[s] de-identified information to improve the Platform, create new software products, and conduct academic research (which may be published)," and that it may "use de-identified information indefinitely without notifying you." If you want your interview deleted, the policy directs EU/UK candidates to "contact the organisation with which you applied for a job" first. That routing — rights exercised through the employer, not the vendor — is the same pattern described in what AI hiring rules entitle you to know in 2026, and it is worth knowing before you ask the wrong party.

Scale, and what the vendor's numbers do and do not show

Sapia reports having "completed more than 10 million interviews across over 47 countries," an average candidate happiness score of "9.1 out of 10" for the chat interview, and a 75.4% completion rate for its combined chat-and-video flow, in a post dated 29 April 2026. Its Woolworths case study reports an 82.6% chat interview completion rate and 27,000 hires in under ten weeks.

All of those are vendor-reported and self-selected. They are credible evidence that the format is deployed at genuine scale and that candidates who finish it do not hate it; they are not evidence about scoring accuracy, adverse impact, or how a given employer weights the result. Sapia links to an independent bias audit from its site footer, which is more than most vendors in this category publish, but the audit's scope and findings are not summarised on the candidate-facing pages.

Where the chat interview sits in the wider process

A chat interview is almost always a screening stage, not a final one. Sapia's own description of the flow is that employers "decide whom you want to approach for video or final interviews" after reviewing the shortlist — and Sapia sells a separate video interview product that many employers run immediately afterwards. Clearing the chat stage usually means a recorded video or live interview is next, which is a completely different recording surface; what HireVue records and scores covers what changes when the camera comes on.

If being screened by software at all is the objection rather than the mechanics, that is a separate conversation with the employer, and there is more room to have it than most candidates assume — see what you can actually request when you are sent an AI interview.

What to do with this

Treat a Sapia chat interview as a writing sample that is read by a model and re-read by a person, not as a form to clear. Answer in 50–100 words. Use specific, first-person detail rather than general competence statements. Take the time the untimed format gives you. And if the real-time alert tells you an answer looks AI-generated, rewrite it then and there rather than submitting through the warning.

Live AI assistance is the wrong tool for this stage, and Control is not built for it: an asynchronous typed interview with a text-authorship classifier running on the input is close to the worst possible fit for a real-time desktop assistant, and using one invites exactly the flag this article describes. The place where a desktop assistant is actually relevant is the live video or screen-share round that often follows. For the stage in front of you, the more useful preparation is structured practice: our guide to preparing for an AI interview covers how to build answers that hold up whether a model or a hiring manager reads them first.