Why A Human-Sounding AI Voice Is No Longer Proof Of Anything

🔍 Read the full analysis: Why A Human-Sounding AI Voice Is No Longer Proof Of Anything on ThorstenMeyerAI.com

Before you orderOffer from Amazon

Get hardware and tech essentials delivered free with Prime

• Fast, free delivery on millions of items
• Prime Video, Amazon Music and more included
• Member-only deals all year

Start your free Prime trial →Free trial for eligible customers · Cancel anytime

As an affiliate, we earn on qualifying purchases.Hide

Block announced Dessa, a speech-to-speech research model that it says can listen and talk at the same time and reproduce cues such as laughter, pauses and interruptions. The company’s demonstrations have drawn attention to a practical risk: a human-sounding voice is no longer reliable proof of identity, even though outside researchers have not tested the model.

Block, the company behind Square and Cash App, posted a project page for Dessa on October 23. The company says the model was trained from scratch on recorded human conversation, without text or pretrained models. It also says Dessa processes speech in a full-duplex manner, meaning it can listen while speaking and respond with backchannels or interruptions.

Block’s examples include a model laughing mid-sentence, helping a speaker recall a film title and responding to an anniversary surprise. These are company-selected recordings, not independently verified tests. Block says the recordings are live conversations and that nobody steered them. Co-founder Jack Dorsey described talking with the model as “like talking with a person,” while Dessa lead Ragavan Thurairatnam called the work “a GPT3 moment for speech.”

The evidence available to outsiders is limited. As of October 26, Block had published no technical paper, model card, benchmark results, weights or API, and had not announced a public release date. The company says inference runs faster than real time on a MacBook Pro M4 Max or better, but has not provided a latency figure. Access is limited to a waitlist and a small group of AI safety researchers, according to the source material.

Why Voice Checks Need a Backup

More natural speech could make voice interfaces easier to use, but the same qualities can make impersonation harder to spot. A cloned voice can imitate conversational rhythms that callers may associate with a real person. That does not mean Dessa has been used in a scam; there is no evidence in the source material that the model has been publicly released or used for fraud. The concern is that improvements in conversational realism could reduce one informal warning sign people rely on.

The FBI’s 2025 Internet Crime Report, published in April 2026, counted 22,364 complaints mentioning AI, with $893 million in reported losses. It also recorded more than $5 million in losses to distress scams using cloned voices of loved ones. Those figures describe reported complaints and losses, not proof that Dessa caused or contributed to them. They show that voice-based deception is already a concern before this model is available.

For employers and families, the practical implication is to avoid treating a voice alone as authorization. A second channel, such as calling back using a number already on file or confirming a payment request in writing, can help check identity. Families can agree on a code word for urgent situations. These steps address impersonation risk generally; they are not evidence that any one product can prevent fraud.

What Block Has Demonstrated

Dessa comes from a small research lab at Block, which the company says was founded by former members of Toronto-based deep-learning startup Dessa. Square acquired that startup in 2020, according to CoinDesk. Block frames the project as an exploration of speech models that retain emotion and conversational timing rather than converting speech into text and then generating a response.

Outside coverage has been more cautious than the launch language. The Verge described some conversations as “a little awkward and uncanny,” while finding Dessa more natural than another model in the comparison. TechCrunch reported that it was unclear whether emotion in the demonstrations might be steered by tags. CoinDesk said Block had not disclosed the model’s size, benchmark results or a public release date. These reports concern the demonstrations and missing technical details; they do not establish how the model performs across users or settings.

Block says the model was trained using one NVIDIA GB200 NVL72 and learned from “hundreds of times fewer words” than large language models need. Those are Block’s descriptions, and the source material provides no independent documentation to assess the training claim or compare performance. The company also presents language learning as an open question.

““Talking with it feels like talking with a person…full of emotion, nuance, pauses, and a natural cadence.””

— Jack Dorsey, Block co-founder, as quoted in Block’s project announcement

What Has Not Been Independently Tested

There is no public model access or independent evaluation in the material available as of October 26. It is not yet clear how Dessa performs beyond Block’s chosen examples, how reliably it handles different speakers and environments, or how it behaves across languages. The company has not published a model card, benchmarks, technical paper or release date.

Other open questions include whether the model can imitate a specific person, what safeguards or consent checks it uses, and whether its output is watermarked. Block says Dessa chooses “its own voice and personality,” but the source material does not document the limits of that capability. It also remains unclear whether the emotion heard in the demos is produced entirely by the model or influenced by other controls.

The source material says the EU AI Act’s disclosure rule under Article 50 has applied since August 2, 2026, but it does not provide a legal analysis of how that rule applies to a particular Dessa use. No claim should be made from these demos alone about legal compliance or the model’s suitability for deployment.

Awaiting Access and Technical Evidence

The next meaningful steps are publication of technical documentation, independent access and testing, and clearer information from Block about release plans, safeguards and consent. Until those details are available, the demos can show what Block wants viewers to hear, but cannot establish the model’s general performance or risk profile.

Meanwhile, organizations handling payments or sensitive requests can review whether their procedures rely too heavily on a caller’s voice. Families can set a verification phrase and confirm urgent requests through a trusted number. These precautions apply to voice impersonation broadly, not specifically to Dessa. A familiar voice should prompt verification, not replace it.

Key Questions

What is Block’s Dessa?

Dessa is a speech-to-speech research model announced by Block on October 23, 2026. Block says it was trained on recorded human conversation and can listen while it speaks.

Can the public use or test Dessa?

Not based on the information available as of October 26. Block had not announced public access, released model weights or published an API. The source material says a small group of AI safety researchers had early access.

Has Dessa been shown to commit fraud?

No such use is established in the source material. The concern is that realistic voice AI may make impersonation harder to detect; the reported FBI fraud figures are not attributed to Dessa.

How can people check an urgent voice request?

Hang up and call the person back using a number you already know, and confirm money requests through a separate channel. Families may also agree on a code word. A voice alone should not authorize a payment or disclosure of sensitive information.

Source: ThorstenMeyerAI.com