Block’s Dessa Voice AI Draws Praise and Scam Concerns

Highlights

Block has introduced Dessa, a voice AI trained entirely on human speech that processes audio directly rather than converting it to text first. The company says it can handle interruptions, laughter, and shifts in tone, and can run locally on a MacBook Pro with an M4 Max chip or better. Demonstrations have impressed some listeners with their natural sound, while others warn that realistic AI voices could make scams harder to recognize. Dessa is currently a research preview, not a publicly available product. Block has not disclosed the model’s size, benchmark results, or public release date, and is initially sharing it with a small group of AI safety researchers.

Sentiment Analysis

• Overall sentiment: Mixed. The article presents both enthusiasm about the model’s conversational quality and concern about possible misuse. Its demonstrations are described as strikingly human, with reactions highlighting emotional expression, natural cadence, and the ability to respond to interruptions.
• Positive response: Some social media users expressed amazement, and Changpeng “CZ” Zhao pointed to potential applications in children’s education. Block also suggests speech-based systems might eventually recognize emotional cues, including distress and pain, that text transcripts can miss.
• Negative response: Critics question why AI voices need to sound human and warn that realism could lower listeners’ guard or enable more convincing scam calls.
• Important uncertainty: Public evidence remains limited to company demonstrations. The article notes the absence of disclosed benchmarks and independent testing, so claims about performance should be treated as preliminary.

Article Text

Block, the payments company led by Jack Dorsey, has unveiled Dessa, a voice-based artificial intelligence model that the company says can converse in a natural and expressive way. Demonstrations have drawn attention because the system appears to respond with conversational timing, emotion, and vocal variation. The debut has prompted two distinct reactions: interest in the potential of more natural voice interfaces and concern that realistic synthetic speech could be used to deceive people.

According to Block, Dessa was trained entirely on human speech rather than text. The model processes audio directly, instead of first converting a speaker’s words into text and then generating a spoken reply. Block says this approach allows Dessa to listen while speaking and respond to interruptions, laughter, and changes in tone. In demonstrations, it laughs during a sentence and reacts to a surprise with an expressive response. The company says the conversations were recorded live, without someone directing the model’s responses.

Dorsey described the experience on X as similar to talking with a person, citing emotion, nuance, pauses, and natural cadence. Social media users also commented on the examples. X user Alexander Hoff said he had never heard anything artificial that sounded so human, while user Hesamation called the demonstration “straight up sci-fi stuff.” These reactions reflect the impression created by the samples, but they do not constitute independent performance testing.

Block says it trained Dessa from scratch using a single Nvidia GB200 NVL72 computing system. The company also states that the model is compact enough to run locally on a MacBook Pro equipped with an M4 Max chip or better. Local operation could allow speech processing on a user’s computer, though Block has not provided additional technical details about the model’s size or its resource requirements beyond that claim.

Despite the attention, Dessa is a research preview rather than a public product. Block has not disclosed the model’s size, benchmark results, or a public release date. The audio samples available so far are company demonstrations, not independent evaluations. These limits mean that observers do not yet have enough public information to compare the model’s capabilities systematically with other voice systems.

Some reactions have focused on possible constructive applications. Binance founder Changpeng “CZ” Zhao suggested that natural voice interaction might support children’s education. He said the current generation of AI voice tools had posed a challenge for his education project, GiggleAcademy. The article presents this as a potential use rather than a confirmed deployment or a tested educational outcome.

Other responses have emphasized risks. Filmmaker Justine Bateman questioned why AI voices need to sound human, arguing that a convincing voice could encourage people to lower their guard. Journalist Shiv Aroor similarly warned that scam calls could become much more widespread. These concerns center on impersonation and social engineering: if listeners cannot readily distinguish a synthetic voice from a human one, a caller may find it easier to gain trust. The article does not report a specific scam involving Dessa, but identifies deception as a concern raised by observers.

Block says it is offering early access to a small group of AI safety researchers. The company argues that speech-based AI could eventually become more capable of recognizing emotional cues, such as distress and pain, that may be lost in text transcripts. Researchers will also assess risks that Block may have overlooked. The early-access process is therefore both a research step and an opportunity to examine safety issues before any wider availability.

Dessa is connected to Block’s broader business, which operates Square, Cash App, bitcoin wallet Bitkey, and bitcoin mining hardware unit Proto. Block’s AI research lab reports directly to Dorsey and includes former staff from Toronto-based AI startup Dessa, which Square acquired in 2020. The company’s announcement places the voice model within its research efforts, while leaving its eventual product role and release plans unspecified.

Key Insights Table