Could AI Voices Like Dessa Make Phone Scams Worse?
Dessa's full-duplex AI voice sounds fully human. Here's why that raises the risk of phone scams and why detection tools struggle to keep up.
Could AI voices like Dessa make phone scams worse?
Yes, almost certainly, if the technology spreads beyond controlled demos. Dessa, the full-duplex speech model from Jack Dorsey’s company Block, produces voices that laugh mid-sentence, interrupt naturally, and react in real time the way a real person on a phone call does. Scam calls currently fail often because the voice sounds robotic or can’t hold a natural conversation. A voice model that removes those tells, combined with existing voice-cloning tools, closes the gap that currently protects a lot of potential victims, especially older adults who have never heard an AI sound like this.
TL;DR
- Dessa is a full-duplex speech-to-speech model built by Block, meaning it takes audio in and sends audio out directly with no text step in between, and it can listen and talk at the same time like a person on a call.
- It was trained only on recordings of real human conversation, not text, which is why it picked up the messy, natural parts of speech, like pauses, laughs, interruptions, and overlapping talk, that make a voice feel human rather than scripted.
- Phone scams rely heavily on a convincing voice and natural back-and-forth, and a model that can sound emotional, respond instantly, and run many calls at once removes one of the main reasons scam calls currently fail.
- Detection tools exist but face a structural problem: audio watermarking and AI-voice detectors are in a constant arms race with better models, and most real phone calls, like a landline or a small business line, have no detector running at all. The human ear is the only filter.
- Block has not released Dessa publicly. It’s in a waitlist-gated early access phase aimed at AI safety researchers, with no public API, weights, or pricing, so the model itself isn’t yet usable by scammers.
- The underlying capability is already proven, and in AI, once a capability is demonstrated, other groups tend to replicate it, which means the scam risk is about where this technology heads next, not just what Dessa itself does today.
What makes Dessa different from older AI voice assistants?
Most AI voice assistants work in three separate steps: they convert your speech to text, run that text through a language model to generate a reply, then convert the reply back into speech. That pipeline works, but it strips out tone, emotion, and timing along the way, and it forces a strict turn-taking rhythm, like a walkie-talkie. You talk, then it talks, then you talk again.
Dessa skips the text step entirely. It’s a speech-to-speech model, so sound goes in and sound comes out directly. It’s also full duplex, meaning it can listen and generate speech at the same time. That’s why it can say “mhm” or laugh while someone is still mid-sentence, interrupt at a natural point, or get louder if asked to speak up. According to Block, the model was trained from scratch using only recordings of human conversation, with no text and no existing language model involved, and it reportedly needed far fewer words of training data than a typical large language model.
That training choice matters because spoken language and written language are genuinely different. People don’t talk the way they write. Pauses, false starts, overlapping speech, and little involuntary laughs only show up in actual conversation, never in text. A model trained purely on speech ends up fluent in exactly the cues humans unconsciously use to judge whether they’re talking to a real person.
Why do these ultra-realistic voices matter for scam calls specifically?
Phone scams depend on two things: a voice that sounds trustworthy and a scammer who can keep a natural conversation going under pressure. Right now, a lot of scam attempts collapse because the voice sounds off, there’s an unnatural delay, or the caller can’t react fluidly to what the target says. Those failure points are exactly what Dessa-style models remove.
A classic example is the “grandparent scam,” where someone calls an older person pretending to be a grandchild in trouble, claiming they’ve been in an accident or arrested and need money fast. These calls succeed by exploiting urgency and emotional trust built on voice recognition. A voice AI that sounds naturally distressed, laughs or stumbles convincingly, and responds instantly to whatever the target says takes away one of the few remaining tells that something is wrong. Add voice cloning, which can replicate a specific person’s voice from a short audio sample, and the scam becomes far more targeted.
The risk isn’t limited to family emergency scams. Fake bank calls, fake delivery notifications, fake tech support, and fake government officials all depend on a voice sounding legitimate. Any scam built around a convincing phone call becomes more effective once the voice itself stops giving anything away. And because this type of model can run many simultaneous conversations without fatigue, the economics of running scam operations at scale shift too.
Why can’t detection tools just catch AI-generated voices?
Other agents start typing. Remy starts asking.
Scoping, trade-offs, edge cases — the real work. Before a line of code.
People are building detection systems, and some companies add inaudible watermark-style signals to AI-generated audio so it can be identified later. But two practical problems limit how much this actually protects anyone right now.
First, detection is an arms race. Every time a detection method improves, voice models improve to get past it, and there’s no reason to expect that cycle to stop favoring one side permanently. Second, and more important, detection tools only work if someone is actually running them at the moment of the call. A grandmother’s landline has no AI detector. The person answering the phone at a small business doesn’t have one either. For the overwhelming majority of real-world phone calls, the only thing standing between a listener and an AI voice is their own judgment, and that’s precisely the judgment these models are now good enough to fool.
Is Dessa itself available for scammers to use right now?
No. Block has not released Dessa publicly. Access is limited through a waitlist, and the company has said it’s prioritizing early access for AI safety researchers specifically to identify risks before any wider release. There’s no public API, no published pricing, and the model’s weights haven’t been released. So nobody can currently plug Dessa into a scam operation.
That said, the capability has been demonstrated clearly and publicly, with demo clips viewed millions of times. History in AI suggests that once a lab shows a capability is achievable, competing groups and open-source projects tend to catch up, sometimes quickly. The concern raised by people who watched the Dessa demos isn’t really about this one model being misused. It’s about what happens once similar full-duplex, conversation-trained voice models become widely available, whether through Block, competitors, or open alternatives.
Is this risk actually new, or just an extension of existing voice-cloning scams?
Voice cloning scams already exist and already cause real financial harm. What’s new with full-duplex, conversation-trained models like Dessa is the behavior layer on top of the voice itself. Older voice-cloning scams could replicate how someone sounds but often struggled to sustain a natural, responsive conversation, which gave alert listeners a chance to notice something was wrong. A model that reacts instantly, overlaps speech naturally, and improvises believable personal details on the spot (as Dessa did in one demo where it invented a backstory about being human) removes that safety margin. The risk isn’t a new category of scam so much as a significant upgrade to how convincing the oldest trick in the scam playbook, a trustworthy-sounding voice, can become.
Frequently Asked Questions
What is Dessa, exactly?
Dessa is a full-duplex, speech-to-speech AI voice model built by a research lab inside Jack Dorsey’s company Block. It converts audio directly to audio without a text intermediate step and was trained entirely on recordings of human conversation rather than text data.
Can Dessa be used by scammers today?
Not currently. Block has kept Dessa in a waitlist-gated early access program focused on AI safety researchers, with no public API, model weights, or pricing released.
Why do elderly people face higher risk from AI voice scams?
- ✕a coding agent
- ✕no-code
- ✕vibe coding
- ✕a faster Cursor
The one that tells the coding agents what to build.
Older adults are often less exposed to recent advances in AI voice technology, so they’re less likely to suspect that a completely natural-sounding caller could be artificial. Scams targeting them, like the grandparent emergency scam, rely on emotional urgency and voice trust, both of which become more effective with a highly realistic AI voice.
Can AI detection tools reliably catch these scam calls?
Not reliably in real-world settings. Detection methods exist, including audio watermarking, but they’re in a continuous arms race with improving voice models, and most everyday phone lines don’t run any detection system at all.
What makes a voice trained only on speech sound more human than text-based AI voices?
Spoken conversation contains pauses, false starts, laughter, and overlapping speech that never appear in written text. A model trained exclusively on real conversations learns these natural, imperfect patterns directly, which is what makes it harder to distinguish from an actual person compared to older voice assistants built around text-based language models.