This review is part of a larger series of LinkedIn newsletters titled The Human Side of AI: Cutting through the AI noise to show you how AI can be a powerful tool for your creativity, efficiency, and strategy.
Conversational AI is no longer experimental. In HR, platforms like Paradox and HireVue automate parts of screening with voice-based interactions. In customer service, AI agents field large volumes of calls for enterprises, handling everything from billing questions to technical support in a tone that keeps getting more natural. That same capability has now reached market research in earnest.
What’s new since the last review: When this piece first ran in April 2025, AI interviewers were a “what if.” Most of that “what if” has since shipped. OpenAI moved from a browser demo to a production Realtime API and, in July 2026, to GPT-Live, a full-duplex voice experience inside ChatGPT. Hume released EVI 3. Sesame open-sourced its speech model and launched a consumer app. And a whole category of purpose-built AI interview platforms now sells directly to research teams. This update adds a closer look at Sesame with ratings, refreshes the three platforms below, and updates where the ethical questions stand now that they are live rather than hypothetical.
The through-line for a researcher is simple. These tools can hold a fluid conversation and probe when an answer is thin, capturing signal in how something is said, not only what is said. What was a promising demo in early 2025 is now a working product category.
Here is where the three platforms we highlighted stand today.
Sesame open-sourced its Conversational Speech Model (CSM-1B, Apache 2.0) and now ships two things: a research preview that remains the best way to hear high-fidelity voice presence, and a consumer companion app. The short version is that the voice quality is still one of the clearest previews of where AI-moderated interviews could go, and it still is not a research tool.
OpenAI has moved fastest on distribution. The browser demo linked in 2025 became the Realtime API, generally available since August 2025 with production voice tooling, and in July 2026 OpenAI shipped GPT-Live, a full-duplex voice experience that listens and speaks at the same time inside ChatGPT. You can interrupt it and talk over it the way people do on a real call. For interview design, that removes one of the clearest “I’m talking to software” tells, even though GPT-Live is not yet broadly available as a developer API at launch.
Hume AI kept its focus on the emotional layer. EVI 3, its speech-to-speech model, reads prosody and tone and adjusts its delivery in kind, and it lets you specify a voice in plain language, for instance by asking it to sound hesitant or emotionally restrained. For ad testing, concept validation, or interviews where how a respondent says something matters as much as the words, that layer is the differentiator.
What this means
Two things stand out. First, high-scale, high-fidelity voice research is now viable. You can imagine running hundreds or thousands of qualitative interviews conducted by voice agents, then analyzing transcripts alongside tone, pacing, and delivery. Signals that were hard to capture at scale are becoming structured data.
Second, the future this piece pointed toward has arrived as a product category. Since the original version ran, purpose-built platforms like Listen Labs, Strella, Outset, and Voicepanel have emerged to run moderated interviews, export transcripts, and synthesize themes. The gap that Sesame and the raw voice models leave open is exactly the gap these tools fill.
Trust and ethics
The harder part is still trust.
The realism that makes these tools useful also raises the stakes around influence and identity. In espionage, a “legend” is a fully constructed false identity with a backstory, a location, an accent, and documents to support it. Conversational AI is now good enough to wear one. Picture a research participant whose accent, vocabulary, and cultural references all match a plausible professional persona. If that participant were generated to deceive, would a screen catch it?
The same persuasive potential sits on the researcher’s side. A voice agent can match a respondent’s tone and pacing, the mirroring technique familiar from sales. Done well, that yields more natural conversations. Push it too far and it blurs the line between rapport and manipulation. A few specific risks are worth naming:
- Persuasive moderators. An AI that guides a respondent toward certain answers through tone or phrasing rather than probing neutrally.
- Synthetic empathy. Warmth that opens a respondent up, then reads as a betrayal if they learn later it was a machine.
- Bias in tone matching. An agent that mirrors some accents better than others, quietly favoring certain groups over others.
- Synthetic respondents. Models trained to impersonate a respondent type, feeding plausible but fake answers into a high-volume study.
None of these are brand new. Fraud, deception, and response bias have always been part of research. What changed is the realism and the scale, and how hard it now is to tell the sincere from the engineered. The response is the unglamorous one: tighter quality controls, disclosure about when a participant is speaking with AI, and a human in the loop wherever interpretation carries the weight.
A Closer Look: Sesame
We first flagged Sesame in this roundup as a startup founded by Oculus veterans building voice agents natural enough to make you forget you’re talking to software. A year on, Sesame ships two distinct experiences, and both are worth understanding if you’re thinking about the future of AI-moderated interviews.
The Research Preview remains the best place to evaluate the underlying Conversational Speech Model. A short, time-limited session with Maya or Miles needs no account, and a free account buys longer access. It remains the fastest way to hear what voice presence actually sounds like: natural interruptions, mid-sentence tone shifts, and laughter that mostly lands where it should.
The Mobile Preview app is a different product. It repackages the same voice model into four personal agents, Maya, Miles, Simone, and Charlie, built for in-between moments like a commute or a quiet walk. It adds search cards, note-taking, memory that persists across sessions, and an incognito mode for conversations you do not want remembered. It is a consumer companion, not a research tool, and Sesame says so plainly.
Key strengths
- Conversational realism. Sesame generates speech directly instead of reading LLM output aloud, which produces more natural pacing, disfluencies, and interruption handling than most competing experiences. For a preview of what AI-moderated interviews could feel like in a few years, this is one of the clearest ones available.
- Live parallel search. The mobile app can run background searches mid-conversation and weave results in without the stilted pause common elsewhere. That makes it feel more assistant-like than demo-like, even though the output is still aimed at consumers rather than researchers.
- Zero-friction trial. Between the account-free research preview and the free mobile app, there is little barrier to spending an hour listening to where conversational AI is headed.
What could be better
Neither product is built for research yet. There is no interview mode, no transcript export, no panel management, and no synthesis layer. Those are the gaps that purpose-built platforms exist to fill.
The mobile app also cannot ingest documents or show a verbatim transcript of your own conversations, which limits it even for casual note-taking. And because memory persists by design, anyone thinking about pointing it at real interview data should work through consent and data handling first. Sesame’s own materials note that its agents can make mistakes or generate incorrect outputs, which makes that caution worth taking seriously.
Ratings
| Dimension | Rating | Rationale |
|---|---|---|
| Usability | 4.5 / 5 | The voice interaction itself is nearly frictionless: the easiest “just start talking” experience in this category. |
| Power | 3.5 / 5 | The speech model is best-in-class, but the packaged apps do not yet do much a researcher could operationalize in a real workflow. |
| Flexibility | 3 / 5 | The 1B-parameter CSM model is open source under Apache 2.0, so developer teams could build their own moderated-interview tooling on top of it. The consumer apps offer no direct path today. |
| Cost | 4 / 5 | Free during the preview phase on both products, with no research-specific pricing tier yet to evaluate. |
Nobody should point Sesame at a real interview yet. But if you want to hear what AI-moderated qualitative research could become on the fidelity axis, this is still the demo worth running.
In sum
Conversational AI has crossed from supporting research workflows to conducting them. Voice systems already guide conversations and adjust mid-dialogue to a respondent’s tone, and a category of tools now packages that into something a research team can actually run.
The open questions are no longer about capability. They are about judgment: whether an AI moderator surfaces real insight or just moves through a logic tree with good manners, and whether we stay honest with the people on the other end about who, or what, they are talking to.
That is the part no model settles for us. The future here is not human or AI. It is Human + AI, with a researcher deciding what the conversation means and where a machine has no business leading it.
At Cascade Insights®, we’re benchmarking these voice and interviewer tools in real B2B studies as they mature; if you’re deciding where they fit in your own research, that’s a conversation we’re glad to have.
Last updated: 7/14/2026