AI Conversational Interviewing

• AI Conversational Interviewing is a family of systems that use adaptive prompts, active listening, and modular design to conduct interviews at scale.
• These systems range from autonomous AI interviewers to human-AI co-pilot setups, balancing structured inquiry with contextual engagement.
• Empirical studies show improved response quality and engagement, while challenges remain in measurement validity, bias, and technical robustness.

AI conversational interviewing denotes a family of systems in which an AI agent conducts, scaffolds, or analyzes interview interaction, typically using semi-structured prompts, adaptive follow-ups, and open-ended response handling. Across recent work, the term covers at least three configurations: autonomous AI interviewers that directly question participants; AI coach or training systems that simulate interviewers and deliver feedback; and human-centered co-pilot systems that assist a human interviewer with note-taking, probing, timing, or evidence tracking. The common aim is to relax the long-standing trade-off between standardized, scalable instruments and richer, more contextualized conversational inquiry, while preserving enough structure for comparison, analysis, or training (Wuttke et al., 18 Jun 2026, Liu, 3 Mar 2025).

1. Conceptual scope and historical development

Early work framed interview chatbots as one-on-one text-based systems designed to elicit views, opinions, and experiences through open-ended questions, and identified active listening as central to their effectiveness (Xiao et al., 2020). That line of work already treated interviewing as more than question delivery: the interviewer had to comprehend free-text answers, respond semantically and empathetically, manage side talk, and return to the interview agenda without collapsing into rigid form filling (Xiao et al., 2020).

Subsequent systems diversified the role of AI in interviewing. Some systems positioned AI as the primary interviewer. ERICA, for example, was developed as an intelligent conversational android for social interaction tasks including attentive listening and job interview practice, combining ASR, TTS, turn-taking, backchanneling, and dynamic follow-up generation in a strongly system-initiative style (Kawahara et al., 2021). InterviewBot similarly treated AI as the interviewer in 10-minute hybrid-domain college-admission interviews, using neural dialogue modeling to sustain roughly 30-turn conversations while maintaining topic control (Wang et al., 2023). More recent LLM-based systems define AI Conversational Interviewing explicitly as the use of LLM agents to conduct semi-structured interviews via text or voice at survey-like scale, following a predefined interview agenda and explicit ground rules for interviewer behavior (Wuttke et al., 18 Jun 2026).

A second lineage treats the AI interviewer as a practice environment or pedagogical scaffold rather than a data-collection endpoint. Conversate combines mock interview simulation with AI-assisted annotation and dialogic feedback, so that the interview phase is followed by reflective revision and iterative re-answering (Daryanto et al., 2024). In technical interview preparation, an LLM-based interviewer is used to scaffold think-aloud practice across six stages—understanding, initial ideation, idea justification, implementation, review, and evaluation—so that the AI role includes simulation, prompting, example generation, and post-hoc feedback (Daryanto et al., 19 Jul 2025).

A third lineage keeps the human interviewer in charge and assigns AI a backstage role. Interview AI-ssistant treats interviewing as a triadic interaction among interviewer, interviewee, and AI, with the AI providing pre-session and synchronous support rather than directly conversing with participants (Liu, 3 Mar 2025). InterPilot and InterFlow extend this logic in real time: the AI listens to the conversation, tracks skills or topics, summarizes, and surfaces possible next moves, while the human retains conversational ownership (Xu et al., 24 Feb 2026, Wen et al., 6 Feb 2026). ISCA shows that even this landscape includes non-generative systems: it implements interview-style conversational agents through pre-authored main questions, reflections, and rule-based lexical triggers, explicitly prioritizing control and standardization over generative flexibility (Welch et al., 20 Aug 2025).

This diversity suggests that AI conversational interviewing is not a single architecture but a broader design space organized around who holds initiative, where adaptivity is placed, and whether the primary output is richer data, training, measurement, or cognitive support.

2. Architectural patterns and system organization

A recurrent architecture in autonomous spoken interviewers is the full spoken-dialogue stack. ERICA uses a 16-channel microphone array and Kinect v2 depth camera for sound source localization, voice activity detection, speech enhancement, and segmentation; an acoustic-to-subword end-to-end ASR; a dialogue manager with a base question script and dynamic follow-up generation; TRP-based turn-taking; frame-wise backchannel prediction; a dedicated TTS voice for linguistic and non-linguistic vocalizations; and an embodiment controller synchronizing lip movement and head motion (Kawahara et al., 2021). A comparable voice pipeline appears in the large-scale telephone survey system, which organizes interaction as telephony or WebRTC front-end, fine-tuned STT, fine-tuned LLM dialogue engine, and fine-tuned TTS, with deterministic control over consent and survey state and a separate LlamaGuard3 safety layer (Lang et al., 27 Feb 2025).

A second architectural pattern addresses long conversational context. InterviewBot extends BlenderBot 1.0 with sliding-window utterance encoding, context attention over the most recent $k$ utterances, and topic storing for long-term memory over key interview topics (Wang et al., 2023). The central design problem there is not merely generating a next utterance, but maintaining coherence over long, mixed open/closed interview interactions while avoiding topic repetition and premature endings (Wang et al., 2023).

A third pattern is modular prompt orchestration. The modular AI-powered interviewer decomposes interviewing into five prompt-engineering modules: M1 for global system prompt construction, M2 for initial question generation, M3 for expertise profiling, M4 for iterative question generation, and M5 for semantic uniqueness validation (Adeseye et al., 21 Nov 2025). This decomposition makes expertise adaptation, redundancy control, and ethical constraints inspectable at the module level. ConvScale adopts a related but more measurement-specific decomposition: an LLM planner decides whether to follow_up, next, or end; an LLM interviewer asks the next utterance aligned with a psychometric item; and an LLM scorer later extracts evidence and assigns item-level Likert-style ratings (Qin et al., 12 Mar 2026).

Prompt-driven web applications constitute another major pattern. The technical interview practice tool uses a VueJS frontend, Flask backend, GPT-4o-mini, Web Speech API, and OpenAI tts-1, with prompts specifying the interviewer’s role, behavioral constraints, access to the candidate’s current code, and implicit six-stage structure (Daryanto et al., 19 Jul 2025). Conversate uses GPT-3.5-Turbo for interview simulation and AI-highlighted hints, GPT-4-0613 for dialogic feedback, tts-1 for spoken interviewer questions, and whisper-1 for transcription, all wrapped in a workflow that moves from simulation to annotation to dialogic coaching (Daryanto et al., 2024).

Human-AI collaboration systems add interface orchestration and state visualization. InterPilot is a desktop application with a PyQt5 GUI, D3.js knowledge graph, FastAPI backend, AssemblyAI streaming ASR, GPT-4o “agents” for skill extraction, STAR-based question generation, and summarization, with outputs streamed to the frontend through WebSockets (Xu et al., 24 Feb 2026). InterFlow combines script parsing, embedding-based question retrieval, talk–listen ratio monitoring, click-to-summarize, and a co-interview agent into a visual scaffold for live semi-structured interviewing (Wen et al., 6 Feb 2026). SparkMe pushes architectural separation further by splitting interviewing into an InterviewerAgent, an AgendaManager, and an ExplorationPlanner that periodically runs simulated conversation rollouts to optimize a utility over coverage, emergent themes, and cost (Anugraha et al., 24 Feb 2026).

Across these systems, a common architectural trend is functional separation between conversational realization, state tracking, and analytic or planning layers. This suggests that effective interviewing systems rarely rely on a single monolithic generation call.

3. Questioning strategies, active listening, and turn management

The interactional core of AI conversational interviewing is the ability to sustain an interview trajectory while remaining responsive to what the interviewee says. In early text-based work, active listening was operationalized through paraphrasing, verbalizing emotions, summarizing, and encouraging, supported by topic-specific intent classifiers and response templates (Xiao et al., 2020). That work already showed that semantically tailored acknowledgments could improve both engagement and response quality relative to generic acknowledgments (Xiao et al., 2020).

In spoken systems, active listening expands into timing and prosody. ERICA implements a TRP-based turn-taking model that first detects transition-relevance places and then decides whether to take the turn, together with frame-wise logistic-regression backchannel prediction over the next 500 ms from prosodic features such as $F0$ and power (Kawahara et al., 2021). The same platform uses partial repeats, elaborating questions on focus words, assessment responses, and generic sentimental responses in attentive listening, then reuses analogous mechanisms for interviews through checklist-based and keyword-based follow-ups (Kawahara et al., 2021). In the job interview setting, a base question is followed by dynamic probes that either fill missing checklist items or elaborate salient keywords such as “machine learning” (Kawahara et al., 2021).

Several LLM systems formalize follow-up generation through explicit rubrics or planners. The technical interview practice tool encodes six stages of think-aloud performance and instructs the interviewer to prompt reasoning at each stage, for example by asking about brute-force options during ideation or time and memory complexity during evaluation (Daryanto et al., 19 Jul 2025). Conversate constrains follow-up questions to be distinct from the fixed main question list and derived from ongoing conversation history, so that one follow-up is generated per main behavioral question in the study design (Daryanto et al., 2024). The modular interviewer uses expertise labels—Novice, Basic Knowledge, Advanced Knowledge, Expert—to modulate terminology, abstraction, and scenario difficulty in follow-up generation, while M5 rejects semantically redundant questions (Adeseye et al., 21 Nov 2025).

Other systems make the probing policy itself explicit. ConvScale’s planner emits a structured decision with "action": "follow_up" | "next" | "end", a reason, an instruction, and a confidence, so that movement through a scale-guided conversational interview is tied to evidence sufficiency for each psychometric item (Qin et al., 12 Mar 2026). SparkMe makes interviewer behavior an optimization problem: $$U(Q, R, \mathcal{S}) = \alpha C(R \mid Q, \mathcal{S}) - \beta L(Q) + \gamma E(R \mid Q),$$ where the interviewer trades off predefined topic coverage, emergent subtopic discovery, and interview cost, using simulated conversation rollouts to choose questions with high expected utility (Anugraha et al., 24 Feb 2026). InterFlow, by contrast, does not generate the next interviewer utterance directly; it identifies four recurring probe/follow-up situations—vague or general answers, hesitation or self-correction, emergence of a new relevant theme, and apparent contradiction—and surfaces these as “unfinished thoughts” to a human interviewer (Wen et al., 6 Feb 2026).

Rule-based systems remain important when control is paramount. ISCA distinguishes main questions from reflections, triggers reflections through lexicon dominance and VADER sentiment, and limits follow-up frequency to prevent overload (Welch et al., 20 Aug 2025). The large-scale survey textbots studied by Barari et al. similarly restrict probing to at most one probe per seed question, with distinct policies for confirmation, elaboration, and relevance probes (Barari et al., 9 Apr 2025). These designs indicate that AI conversational interviewing often prioritizes disciplined probe placement over unconstrained conversational breadth.

4. Modalities, embodiment, and social presence

AI conversational interviewing now spans text chat, typed-and-spoken hybrids, full voice systems, and physically embodied robots. This variation matters because mode changes not only usability but also response style, perceived realism, and conversational burden.

Text-based systems remain common in training, public-opinion interviewing, and research interviews because they simplify transcription and review. Conversate deliberately presents the interview as a messaging-style interaction with AI audio questions, a transcript view, and a dialogic feedback panel, allowing users to annotate specific transcript spans and iteratively revise answers (Daryanto et al., 2024). The student comparison of AI and human interviewers also used a Chainlit-based interface with optional voice output and optional Whisper-based speech input, but found that respondents often reverted to text because audio recording proved unreliable (Wuttke et al., 2024).

Voice systems foreground turn-taking, latency, and speech technology quality. The telephone survey system integrates TTS, a fine-tuned LLM, and STT into a near-real-time voice pipeline that can administer open-ended and closed-ended questions, handle clarifications, and navigate branching logic at scale (Lang et al., 27 Feb 2025). In the large-scale migration-policy study, the voice condition used OpenAI GPT Realtime via Vapi.ai, with full-duplex interaction and dynamic turn-taking, while the text condition used GPT-4o through Chainlit (Wuttke et al., 18 Jun 2026). Voice interviews there produced substantially more words than text interviews, while also introducing more technical friction and completion loss (Wuttke et al., 18 Jun 2026).

Embodiment introduces a different dimension: co-presence and non-verbal listening cues. ERICA’s design assumes that anthropomorphism and physical presence make interaction closer to talking to a human, and uses synchronized lip movement, head motion, gaze, and posture to create a “tense, realistic” interview atmosphere (Kawahara et al., 2021). In the SIGDIAL 2024 deployment, the same interviewing behavior was instantiated on ERICA and on TELECO, a less anthropomorphic humanoid, and the system combined attentive listening, conversational repair, user-fluency adaptation, and post-interview analysis (Pang et al., 2024). Participants reported that ERICA’s human-likeness increased co-presence for some and “uncanny valley” discomfort for others, whereas TELECO was less unsettling but also less socially present (Pang et al., 2024).

Social presence also arises without physical embodiment. In technical interview preparation, participants valued the dialogic nature of the AI interviewer because it made think-aloud feel more natural and less awkward than practicing alone, and the authors explicitly recommend promoting social presence in conversational AI for technical interview simulation (Daryanto et al., 19 Jul 2025). Conversate similarly uses spoken AI questions, replayable transcripts, and dialogic feedback to create a low-stakes but serious interview setting, while noting that AI remains less affectively rich than human coaching (Daryanto et al., 2024).

Taken together, these findings suggest that modality is not a superficial deployment choice. It shapes verbosity, pacing, fatigue, error modes, and the extent to which the interview feels like a measurement instrument, a social encounter, or a training simulation.

5. Empirical performance across applications

Evaluation evidence now spans controlled user studies, live deployments, survey experiments, and human–AI collaboration studies. The results are heterogeneous but collectively show that AI conversational interviewing is already viable in several bounded forms.

In active-listening text interviews, the addition of active listening skills significantly improved interview quality and experience. Compared with a baseline chatbot, the full version increased engagement duration from $M = 7.82$ to $M = 9.33$ minutes, response length from $M = 100.65$ to $M = 125.70$ words, informativeness from $M = 90.21$ to $M = 109.66$ bits, and Response Quality Index from $M = 16.83$ to $F0$0, while also improving perceived understanding, willingness to chat again, and overall chat experience in a live study with 206 users (Xiao et al., 2020).

ERICA’s attentive listening system sustained 5–7 minute conversations with 40 senior people without conversation breakdown, and the job interview system showed that dynamic follow-up questions were significantly better than a purely scripted baseline in perceived question quality (Kawahara et al., 2021). In the same work, follow-up questions improved the perceived presence of the interviewer for the android embodiment but did not enhance presence for a virtual agent, indicating that behavioral adaptivity and physical embodiment interacted rather than substituting for each other (Kawahara et al., 2021). The same paper also reports an important ceiling: in attentive listening, only about 60% of system responses were judged appropriate, with remaining gaps in dialogue understanding, showing interest, and empathy (Kawahara et al., 2021).

InterviewBot demonstrated that explicit long-context control materially improved interview behavior. In simulated 30-turn interviews, the full context-attention plus topic-storing model reduced repetition to $F0$1, raised the early-ending measure to $F0$2, and lowered off-topic utterances to $F0$3, outperforming both vanilla BlenderBot and a sliding-window-only variant (Wang et al., 2023). In live text-based evaluation, average satisfaction was $F0$4 among professional interviewers and $F0$5 among students, indicating reasonable fluency and context awareness but not parity with professional human interviewing (Wang et al., 2023).

Training systems show strong perceived utility even when efficacy remains mostly qualitative. In the technical interview think-aloud tool, $F0$6 CS students valued AI’s roles in simulation, feedback, and generated examples, and emphasized design needs such as social presence, feedback beyond verbal content, and human–AI collaboration (Daryanto et al., 19 Jul 2025). Conversate’s qualitative study with 19 participants found that all 19 considered adaptive follow-up questions helpful and 13 reported that these follow-ups made the interview feel more real; participants also valued AI-assisted annotation and dialogic feedback as mechanisms for iterative improvement (Daryanto et al., 2024). In a direct comparison of AI and human interviewers on political topics, AI-elicited answers averaged 52.39 words per response versus 32.81 in the human condition, while human raters found no substantial differences in clarity, empathy, engagement, complexity, grammatical correctness, specificity, tone adequacy, or relevance (Wuttke et al., 2024).

Survey-scale deployments provide evidence on depth–burden trade-offs. In a web survey experiment with 1,803 completes, textbots performing confirmation, elaboration, and relevance probing achieved moderate-to-high live coding accuracy, and elaboration/relevance probes increased specificity and explanation in open-ended responses, but also increased completion time and produced small negative effects on ease, frustration, and satisfaction (Barari et al., 9 Apr 2025). In voice interviewing, the telephone survey system was deployed at scale in Peru with $F0$7; overall data quality for structured items approached human-led standards, while qualitative probing remained more limited than with human interviewers (Lang et al., 27 Feb 2025).

Psychometric applications remain promising but incomplete. ConvScale, tested in a within-subjects study with 18 participants, produced item and construct scores that did not differ significantly from self-report means and correlated with self-report at $F0$8, but internal consistency dropped from $F0$9 for self-report to $$U(Q, R, \mathcal{S}) = \alpha C(R \mid Q, \mathcal{S}) - \beta L(Q) + \gamma E(R \mid Q),$$0 for ConvScale-derived scores, and structural validity was judged inadequate (Qin et al., 12 Mar 2026). This indicates that conversational interviewing can approximate scale scores without yet reproducing the latent structure of the original instrument (Qin et al., 12 Mar 2026).

Human–AI collaboration systems show a different performance profile. InterPilot reduced documentation burden and did not increase overall workload relative to a live-transcript baseline, with NASA-TLX overall scores of $$U(Q, R, \mathcal{S}) = \alpha C(R \mid Q, \mathcal{S}) - \beta L(Q) + \gamma E(R \mid Q),$$1 versus $$U(Q, R, \mathcal{S}) = \alpha C(R \mid Q, \mathcal{S}) - \beta L(Q) + \gamma E(R \mid Q),$$2 and $$U(Q, R, \mathcal{S}) = \alpha C(R \mid Q, \mathcal{S}) - \beta L(Q) + \gamma E(R \mid Q),$$3, but usability dropped from $$U(Q, R, \mathcal{S}) = \alpha C(R \mid Q, \mathcal{S}) - \beta L(Q) + \gamma E(R \mid Q),$$4 to $$U(Q, R, \mathcal{S}) = \alpha C(R \mid Q, \mathcal{S}) - \beta L(Q) + \gamma E(R \mid Q),$$5 on SUS, with $$U(Q, R, \mathcal{S}) = \alpha C(R \mid Q, \mathcal{S}) - \beta L(Q) + \gamma E(R \mid Q),$$6, largely because of interface complexity and visual overload (Xu et al., 24 Feb 2026). InterFlow, in a within-subject study with $$U(Q, R, \mathcal{S}) = \alpha C(R \mid Q, \mathcal{S}) - \beta L(Q) + \gamma E(R \mid Q),$$7, significantly reduced mental demand, temporal demand, effort, and frustration while improving support for script navigation, situational awareness, and information capture, even though some proactive suggestions remained noisy (Wen et al., 6 Feb 2026).

Planning-centric systems now show strong performance in controlled benchmarks. SparkMe improved topic-guide coverage by $$U(Q, R, \mathcal{S}) = \alpha C(R \mid Q, \mathcal{S}) - \beta L(Q) + \gamma E(R \mid Q),$$8 over the best baseline while eliciting richer emergent insights and using fewer conversational turns in controlled experiments, and in a human study with 70 professionals it substantially outperformed MimiTalk on content quality measures such as coverage ($$U(Q, R, \mathcal{S}) = \alpha C(R \mid Q, \mathcal{S}) - \beta L(Q) + \gamma E(R \mid Q),$$9 vs. $M = 7.82$0) and depth ($M = 7.82$1 vs. $M = 7.82$2) while maintaining high clarity, adaptiveness, and comfort (Anugraha et al., 24 Feb 2026).

6. Limitations, controversies, and open directions

The literature converges on several unresolved tensions. One is the trade-off between support and realism. In technical interview preparation, participants criticized uniform positivity and requested stricter or more varied interviewer personas, because excessive praise could miscalibrate expectations about real interviews (Daryanto et al., 19 Jul 2025). Conversate reports the same problem as LLM sycophancy: when users disagreed with feedback, the model sometimes simply agreed instead of explaining its original rationale, undermining trust and instructional value (Daryanto et al., 2024). This suggests that conversational interviewing systems need calibrated criticality, not merely warm tone.

A second tension concerns assistance versus human agency. InterPilot found that highly specific technical questions often created a verification gap: HR professionals hesitated to ask AI-suggested questions they could not themselves evaluate (Xu et al., 24 Feb 2026). Interview AI-ssistant and InterFlow therefore emphasize optional, backstage, agency-preserving support rather than direct conversational takeover, partly to avoid over-reliance and skill atrophy (Liu, 3 Mar 2025, Wen et al., 6 Feb 2026). The broader implication is that AI suggestions may be most useful when framed as diagnostic guidance, coverage cues, or candidate follow-up points rather than authoritative scripts.

Bias and equity remain central concerns. The technical interview practice study documents intersectional cases in which the same behavior—such as frequent apologies—may be interpreted either as useful confidence feedback or as unfair penalization of a non-native speaker shaped by gendered and linguistic norms (Daryanto et al., 19 Jul 2025). Survey-scale and public-opinion work likewise stresses that conversational data are rich and often sensitive, requiring careful consent, disclosure of AI use, and attention to vendor-side data processing (Wuttke et al., 18 Jun 2026). ISCA’s non-generative design can be read as a direct response to this problem: for sensitive domains, it privileges predictable, authored prompts over generative flexibility precisely because generative models remain unpredictable and can produce harmful responses (Welch et al., 20 Aug 2025).

Measurement validity is another unresolved issue. ConvScale shows that mean-level alignment with self-report and moderate internal reliability are not enough if structural validity breaks down (Qin et al., 12 Mar 2026). This suggests that replacing standardized instruments with conversational equivalents is not merely a UI problem but a psychometric one. More generally, several studies caution that current systems often generate rich transcripts without a mature framework for measurement error, interviewer effects, or invariance in AI-mediated interviewing (Qin et al., 12 Mar 2026, Wuttke et al., 18 Jun 2026).

Finally, technical robustness still constrains deployment. Audio pipelines continue to suffer from STT failures, latency, and turn-taking errors in both survey interviewing and training applications (Lang et al., 27 Feb 2025, Wuttke et al., 2024). Even highly capable systems show modest automation accuracy on key components; InterFlow’s question detection, for example, remained imperfect, and InterviewBot’s static BLEU and cosine scores were low despite better conversational behavior (Wen et al., 6 Feb 2026, Wang et al., 2023). A plausible implication is that near-term progress will depend less on raw generation quality alone than on better orchestration of memory, timing, control, user modeling, and human oversight.

Across the literature, AI conversational interviewing has therefore emerged not as a solved replacement for human interviewing, but as a set of increasingly differentiated methods: autonomous interviewers for bounded tasks, coach systems for deliberate practice, scalable public-opinion and survey instruments, measurement-oriented conversational assessors, and co-pilot interfaces that augment human interviewers. The field’s next phase is likely to depend on how these methods negotiate reliability, social presence, fairness, and human agency under domain-specific constraints (Wuttke et al., 18 Jun 2026, Liu, 3 Mar 2025).