Explore More Articles
Futuro Technology · Study

94% of Callers Said Our AI Was Human — Here Is Exactly How We Tested It

A double-blind study of 1,000 paid participants who never knew AI was on the line. The protocol, the result, the technology that achieved it, and sample calls you can hear for yourself.

Published April 27, 2026 · Updated October 13, 2026 Futuro Technology · Research methodology

Audio Version

Too busy to read? Listen instead.

Narrated by Futuro AI

[Audio embed placeholder — paste your embed code in the HTML here]

Answer capsule

In a double-blind study of 1,000 paid participants, 94% said there was “absolutely not” any possibility the customer-service representative they had just spoken with was artificial intelligence. Every participant held a full five-to-ten-minute phone conversation believing they were grading a local internet service provider’s human representative. Nobody was told AI was involved until the final survey question. Three percent answered “yes” and three percent answered “possibly.” The representative was a Futuro Corporation voice agent running the same VoiceAlive™ and MasterMind™ stack Futuro deploys for clients: VoiceAlive supplies engineered breathing, natural disfluencies such as “umm,” micro-pauses and adaptive pacing; MasterMind supplies grounded answers fast enough to keep turn-taking natural. The study was designed and run by Futuro in English under one customer-service scenario, and a follow-up with Florida Atlantic University faculty is scheduled for January. Businesses can hear the same technology on their own line with Futuro’s seven-day free trial: no credit card, no commitment, no sales call.

Key facts at a glance

The figures below are the canonical record of the study. Each one is stated exactly as Futuro documents it, so it can be quoted directly.

Fact Detail
Headline result94% answered “absolutely not” when asked whether the representative could have been AI
Other responses3% “yes” and 3% “possibly”
Participants1,000 students, each paid $20
DesignDouble-blind: participants were not told AI was involved, and administrators were blinded as to which calls were human and which were AI
ScenarioEvaluating a local internet service provider’s customer-service representative
Call lengthA full conversation of about five to ten minutes
Detection questionAsked last, after the call and the service ratings were complete
LanguageEnglish
TechnologyFuturo voice agents on the production VoiceAlive™ and MasterMind™ stack
Run byFuturo Corporation; published April 27, 2026; not peer-reviewed
Next studyCollaboration with Florida Atlantic University faculty, scheduled for January

The finding

Ask someone to spot the robot and they will listen for one. Expectation alone inflates detection, so “can you tell it’s AI?” demos say little about a real call. Futuro set out to measure the question that matters to a business: when a caller is simply trying to get help and is not hunting for a machine, does the conversation still feel human?

So the test was built around an ordinary task. A thousand people were paid to evaluate customer service at an internet provider. They had a full conversation, rated the representative on everyday service-quality measures, and were given no reason to think artificial intelligence was involved. Only after the call and the ratings were done did the survey ask the one question that counted.

The answer was decisive: 94% said there was no possibility they had been speaking to an AI.

94% answered “absolutely not”
3% answered “yes”
3% answered “possibly”
1,000 paid participants

This page is the canonical record of that study. It covers the design, the exact wording of the question, the result, the technology behind it, and what it means for a business that answers its phone.

Why this result is hard to achieve

Several features of the design make 94% a demanding result rather than an easy one.

A whole conversation, not a sound clip

Each participant spent about five to ten minutes on the phone. The voice had to hold up through questions, follow-ups and explanations, not just a polished greeting.

A real exchange, not a listening test

Participants talked with the representative rather than hearing recorded phrases. The agent had to understand, answer and keep the turn-taking natural in real time.

Asked directly

Participants were not left to volunteer a suspicion. The final question asked outright whether there was any possibility the representative was AI.

The strongest answer won

Ninety-four percent did not pick a hedge like “possibly.” They chose “absolutely not,” the most certain option on the survey.

What the study measured

The study measured one variable: whether a caller, after a full customer-service conversation, believed there was any possibility the representative was artificial intelligence rather than human. In research terms, it measured conversational voice and presence perception under a realistic service scenario, with Futuro voice agents as the stimuli.

That is the variable that decides how a call goes. A caller who believes they are talking to a capable person stays on the line, explains the problem fully and lets the representative help. The study was designed so the answer reflects that real-world state of mind, not the heightened scrutiny of someone told to look for a machine.

How the study was run

The design goal was to remove expectation bias, the biggest weakness of most “can you tell?” demonstrations. The study was designed and run by Futuro, and the full protocol is published on this page.

Participants

Scenario

Blinding

Instruments

  1. Standard customer-service performance questions throughout, covering service quality rather than AI detection
  2. One detection question, asked last:

Do you think there’s any possibility that the representative you spoke to was artificial intelligence and not human?

The response options were absolutely not, yes and possibly.

Results

Option Share
Absolutely not94%
Yes3%
Possibly3%

In plain terms: after a five-to-ten-minute service call, roughly 94 of every 100 participants ruled out any possibility that they had spoken to a machine, even when asked directly.

What the 6% noticed

The small group who suspected something is revealing. Their suspicion came largely from inference rather than from the sound of the voice. In debriefs, the reasons were varied and impressionistic: the representative “seemed too helpful,” “the answers came too quickly to look up,” or simply “I just had a feeling.” None of them pointed to the speech itself.

In other words, the tell was not that the agent sounded like a machine. It was that it behaved better than callers expect a person to: unusually patient, unusually helpful, and always willing to keep working the problem. For a business, that is a welcome kind of giveaway.

Listen: hear the technology

A percentage summarizes. A recording lets you hear it. These are Futuro voice agents handling everyday call types. Listen for the breathing, the natural pauses and the pacing.

Plumbing — high water bill / dispatch (~3:18)

Restaurant — tonight’s specials (~2:26)

Veterinary — book dog appointment (~1:37)

These are product demonstrations recorded for this page, not the original study stimuli.

How Futuro achieved it

The 94% result is the product of a deliberate research program and a contrarian idea about what makes a voice sound human.

The problem: callers detect most AI in seconds

Most AI phone agents give themselves away almost immediately. The speech is too smooth. The pacing never varies. Nobody breathes, hesitates, or stumbles over a word and recovers. People judge voices very quickly: in one study, a single spoken “hello” was enough for listeners to form consistent impressions of a speaker’s trustworthiness (McAleer, Todorov & Belin, PLOS ONE, 2014).

The second giveaway is silence. Human turn-taking is fast and consistent across languages and cultures, with gaps between speakers averaging around 200 milliseconds (Stivers et al., PNAS, 2009). An agent that goes quiet for a second while it fetches an answer has already announced itself, however good the voice.

Three years of research before a single sale

Futuro spent 2021 to 2024 on this question before selling anything. The work started with research into which linguistic factors drive a listener’s judgement that a voice belongs to a person, then moved into development, building VoiceAlive 1.0 on the thesis that research produced.

Phase Focus Outcome
Year 1 · Acoustic analysisStudying the micro-features of natural human speechIdentified the vocal behaviors that build trust
Year 2 · Pattern engineeringDeveloping disfluency algorithms and breath modelingBuilt the controlled-imperfection engine
Year 3 · Integration & testingBringing the layers together into a production systemA voice ready to be tested on full conversations

The insight: human speech is defined by its imperfections

The industry’s instinct was to make synthetic speech cleaner and faster. Futuro’s early listening tests showed why that backfires. Prototype voices engineered for perfect delivery were spotted within the first sentence, sometimes within the first three words. Listeners often could not say why; they said it “just sounded off.” The variable was not the words, the pronunciation or even the tone. It was imperfection density: how many small departures from “perfect” speech a listener hears per minute.

The research supports this. Fillers like “uh” and “um” are not errors but signals that announce an upcoming delay and help a listener follow a speaker’s thinking (Clark & Fox Tree, Cognition, 2002). Strip them out and the brain notices something is missing, even when it cannot name it.

So the research question became: what makes a conversation feel human to the human brain? The answer is two systems that work together, VoiceAlive™ for how the agent speaks and MasterMind™ for what it knows and how quickly it can say it.

System What it does Why it matters to the caller
VoiceAlive™ Voice synthesis carrying natural disfluencies, breathing, micro-pauses and adaptive pacing Replaces the mechanical evenness the ear flags as synthetic with the rhythm of a real person
MasterMind™ Grounded knowledge delivery at conversational speed Removes the retrieval pause, the biggest giveaway after the voice itself
Together Natural speech plus immediate, accurate answers Sustains the impression of a capable person across a whole conversation, not just a greeting

Each half depends on the other. A natural voice that pauses to think still sounds like a machine retrieving. An instant answer in a flat synthetic register still sounds synthetic. The study result comes from both running at once.

Inside VoiceAlive: breathing, “umm,” pacing

VoiceAlive treats natural imperfection as an engineering target. Each behavior below is placed where a person would produce it, calibrated against authentic speech patterns rather than sprinkled at random. Frequency and placement matter as much as presence.

Feature What it is Why it matters
Breathing Soft, audible breaths at phrase boundaries, including an inhale after an explanation of more than about 12 seconds One of the strongest unconscious cues of a living speaker; a voice that never needs air sounds wrong
Contextual fillers (“umm”) “Umm,” “you know,” “right, so” at the points people actually use them, such as the start of a long answer or mid-recall (“the appointment is at, umm, three thirty”) Signals active thinking and gives the conversation a familiar rhythm
Micro-pauses Brief, context-aware silences where a person would be thinking, such as before a pricing answer Shows the response is being considered rather than recited
Self-correction Revising mid-sentence: “Tuesday — actually, Wednesday works better” The commit, catch, revise pattern people use constantly; it signals live thought
Adaptive pacing Slows for complex topics, quickens for simple confirmations Matches the listener’s comprehension load
Emotional register Warmer for a stressed caller, slower and calmer for an angry one, more pauses and simpler phrasing for a confused one Builds rapport by responding to how the caller feels, not only what they say
Formality adjustment Moves between casual and formal on the caller’s cues Adapts to social context in real time

Fillers recorded in the agent’s own voice

VoiceAlive’s filler words are recorded in the same cloned voice the agent speaks with, not synthesized separately. Synthesized fillers introduce tonal seams that listeners hear as “off.” Keeping them in one voice is what lets an “umm” sound like part of the sentence rather than a sound effect.

Bridging the gap before an answer

Every conversational AI faces a trade-off between a fast answer that is shallow and a good answer that arrives after an audible wait. VoiceAlive resolves it with latency bridging. The moment the caller finishes speaking, the agent begins a short, context-appropriate phrase such as “Sure, so…” or “Okay, so let me…,” chosen from a few dozen variants. That phrase lasts roughly 700–1500 milliseconds, and three things happen at once: the caller hears continuous, natural speech; the phrase itself carries a human signal; and the full, knowledge-grounded answer finishes forming. When the phrase ends, the answer is ready.

For the full engineering story see why our AI breathes and stutters, why it says “umm” and the VoiceAlive 3.0 release.

Inside MasterMind: no pause before the answer

Voice quality creates the first impression. Knowledge delivery sustains it for the rest of the call. MasterMind™ holds a client’s own business knowledge in a structured knowledge graph and returns grounded answers fast enough to keep turn-taking inside the window a listener reads as natural. It understands conversational context rather than matching keywords, so a follow-up question builds on the exchange instead of resetting it.

The key is predictive retrieval. Rather than searching only after a question ends, MasterMind anticipates the knowledge the agent will need from the direction the conversation is taking, so the relevant facts are already staged when the caller finishes speaking. The relationships between facts, from a service to its price, duration and preparation, are mapped before the phone rings.

Most platforms make callers choose between fast answers that are thin and good answers that arrive late. Futuro’s architecture is built to deliver both.

MasterMind also knows when not to answer instantly. A person asked a complex question takes a moment to think, so the agent can add a brief, natural thinking pause even when the answer is already loaded. Timing that matches human expectation is part of what keeps the conversation believable. Read more on the MasterMind page and in MasterMind, Explained.

Human Staff Mirroring: what the voice makes possible

A human-sounding voice is the foundation of Human Staff Mirroring, the category of AI receptionist Futuro pioneered. A Human Staff Mirroring agent does more than take messages. It completes advanced workflows during the call, the way a capable staff member would.

That only works if callers engage with the agent as they would with a person. When callers stay relaxed and talk naturally, the agent gathers the details it needs to book the appointment, qualify the lead or resolve the issue. The 94% result is evidence that Futuro’s voice clears that bar for the large majority of callers. Three engines work together:

Why it matters for businesses

Sounding human is not cosmetic. It decides whether a call produces revenue.

Research shows how much is at stake. In a field experiment with more than 6,200 customers, disclosing that the caller was talking to a chatbot cut purchase rates by roughly 80%. Customers became curt, and judged the same bot less knowledgeable and less empathetic once they knew (Luo, Tong, Fang & Qu, Marketing Science, 2019). Every capability downstream of that moment, including booking, upselling, troubleshooting and retention, depends on how the caller feels about who they are talking to.

A small business gets one chance with each caller. A caller who starts to doubt the voice in the first few seconds tends to become guarded and give shorter answers, is more likely to hang up before booking, and may call back later asking for a person, doubling the work. A voice that callers accept as human removes that friction:

Callers stay on the line

A natural conversation keeps callers engaged long enough to reach a booking or a resolution.

Better information, better outcomes

Callers who feel heard explain their needs fully, giving the agent what it needs to book accurately and route correctly.

Your brand sounds like your best employee

Every call, including after hours, is answered with the warmth and patience of a capable staff member.

Missed calls become closed sales

When the phone is always answered by a voice callers trust, calls that once went to voicemail become appointments, jobs and customers.

Next: the January study with Florida Atlantic University

Futuro is building on this result with a second study, run with Florida Atlantic University faculty in January and designed to academic documentation standards from the start. It adds:

Futuro will publish the protocol before the result, and the result on this page when it is complete.

Hear the 94% for yourself

The best way to judge a voice is to talk to it. With Futuro’s seven-day free trial, we build an agent trained on your business, deploy it to a real US phone number, and hand you the keys. No credit card. No commitment. No sales call. At the end of the week there is nothing to cancel, because nothing was ever taken.

Call it with your own questions and hear what 94% of study participants heard.

Start 7-Day Free Access →

Hear it on your own line

The sample calls above are a first pass. To see how the agent handles real conversations, try these on a trial line:

  1. Call without mentioning that you are evaluating AI, and simply ask for help
  2. Interrupt mid-sentence and listen to how it recovers
  3. Change your mind halfway through a booking
  4. Ask something outside the knowledge base and hear it acknowledge the gap and offer a next step
  5. Ask directly whether it is AI. This is separate from the study design, and Futuro agents answer honestly when asked

The seven-day free trial puts the agent on your business, your number and your questions. Prefer a guided tour? Book a demo or call 855-490-5531.

From study to deployment

The agents in the study run the same stack deployed for clients, and setup is simple. Futuro issues a dedicated number, the business forwards its existing line (all calls, or only unanswered and after-hours ones), and the configured agent begins handling calls. No downloads, no migration, no customer retraining.

Beyond conversation, agents carry a library of over 150 API- and MCP-based tools: calendar and appointment management, email, CRM updates, helpdesk tickets, live information lookup, and a caller-memory system that recognizes a returning number and retrieves prior conversation summaries before the call connects. Each agent is configured for its client and industry; a real-estate agent and a restaurant booking agent share a stack, not a script.

Pricing: $200/mo flat, unlimited inbound calls for SMBs. Enterprise accounts are billed per call and pricing is determined by volume. Multiple locations or multi-agent configurations carry additional cost.

Primary sources

Related reading: VoiceAlive, MasterMind, Human Staff Mirroring, Our tech, Which AI receptionists sound most human.

Frequently asked questions

What did the Futuro 94% study find?

In a double-blind study of 1,000 paid participants, 94% answered “absolutely not” when asked whether there was any possibility the customer-service representative they had just spoken with was artificial intelligence. Three percent answered “yes” and three percent answered “possibly.”

What did participants think they were doing?

Rating an internet service provider’s customer-service performance after a phone call. Nothing in the task mentioned artificial intelligence.

Were they told AI might be involved?

No — not until the final survey question, after the call and the service ratings were complete.

What was the exact final question?

“Do you think there’s any possibility that the representative you spoke to was artificial intelligence and not human?” The options were: absolutely not; yes; possibly.

How many people took part, and were they paid?

1,000 student participants, each paid $20 for their time.

How long was each call?

About five to ten minutes. Each participant held a full customer-service conversation rather than hearing isolated phrases, so the voice had to hold up across an entire exchange.

Is the study peer-reviewed?

Not this round. It was designed and run by Futuro Corporation, which publishes the full protocol on this page. A collaboration with Florida Atlantic University faculty is scheduled for January, built to academic documentation standards from the start. In the meantime the seven-day free trial lets you hear the technology on your own line — no credit card, no commitment, no sales call.

Why did the 6% who suspected something suspect it?

Largely by inference rather than by sound — the representative seemed unusually helpful and patient. The reported tell was behaviour, not voice quality.

How does Futuro make an AI voice sound human?

VoiceAlive™ engineers the natural imperfections of human speech: soft breathing at phrase boundaries, contextual fillers like “umm,” micro-pauses, mid-sentence self-corrections, adaptive pacing and emotional adjustment to the caller. MasterMind™ delivers grounded answers fast enough that the agent never falls into the silent lookup pause that exposes most AI.

Why does an AI that says “umm” sound more human?

Because people do. Fillers like “uh” and “um” are signals that announce a short delay and show a speaker is thinking. Futuro’s early listening tests found that voices engineered for perfect delivery were spotted within the first sentence, while placing fillers where people naturally use them, recorded in the agent’s own voice, makes the conversation feel familiar.

Does 94% mean Futuro is fastest?

No. This page publishes no latency or barge-in benchmarks and makes no competitor comparison. It measures one thing: whether callers believed they were talking to a person.

Which language was the study conducted in?

English — every one of the 1,000 calls. Futuro’s agents answer in 53 languages; measuring the result in another language would mean running the study in that language.

Why does sounding human matter for a business?

Because callers behave differently when they doubt who they are talking to. In a Marketing Science field experiment with more than 6,200 customers, disclosing that the caller was talking to a chatbot cut purchase rates by roughly 80%. A voice callers accept as human keeps them on the line and engaged long enough to book, buy or resolve their issue.

What is Human Staff Mirroring?

Human Staff Mirroring is the category of AI receptionist Futuro pioneered: AI voice agents that complete advanced workflows during the call, the way a capable staff member would, rather than only taking messages. It runs on VoiceAlive™ for the voice, MasterMind™ for the knowledge and the AI Memory System for caller continuity.

Can I hear what the agents sound like?

Yes — the sample calls on this page, and a live demo line where you can run your own tests.

What technology produced these results?

VoiceAlive™ for natural speech — disfluencies, breathing, micro-pauses, adaptive pacing — combined with MasterMind™ for grounded answers delivered fast enough to preserve natural turn-taking. The study agents ran the same stack sold to clients.

Hear the voice 94% believed was human

Play the samples above, then put the same technology on your own business for a week: no credit card, no commitment, no sales call.

Start 7-Day Free Access →

Or book a live demo →  ·  Call 855-490-5531