Block unveiled Dessa, a voice AI that sounds almost human, trained only on speech and able to run locally on a MacBook Pro. Reactions split between awe and warnings it could power convincing phone scams. This matters because voice this realistic changes how people decide who, or what, they are talking to.
What actually happened
Block, the payments company led by Jack Dorsey, announced Dessa as a research preview on 24 October 2026. The model was trained entirely on human speech, not text, and can listen while speaking, handling interruptions and laughter in real time, according to Block's announcement. Dorsey wrote on X that talking with it "feels like talking with a person...full of emotion, nuance, pauses, and a natural cadence." Block trained Dessa from scratch on a single Nvidia GB200 NVL72 system, and it can run locally on a MacBook Pro with an M4 Max chip or better. X users Alexander Hoff and Hesamation called the demos startlingly natural, while filmmaker Justine Bateman and journalist Shiv Aroor warned the realism could fuel deception and scam calls. Binance founder Changpeng "CZ" Zhao suggested uses in children's education through his GiggleAcademy project.
How we got here
Block's AI research lab reports directly to Dorsey and includes staff from Toronto startup Dessa, which Square acquired in 2020. That history explains the model's name and the lab's long interest in speech technology. Most voice assistants today convert speech to text, generate a reply, then convert it back to audio, a process that strips out tone and timing. Dessa processes audio directly, skipping that conversion. This follows a broader pattern of AI voice tools getting good enough to impersonate real people, which has already been used in phone and video scams reported elsewhere in 2026.
Why this matters for you
For builders, Dessa is a signal that conversational AI compact enough for consumer hardware is becoming realistic, relevant for anyone designing voice interfaces into wearables or smart glasses. For users, more natural AI voices mean fewer cues to tell human from machine on a call. For Block, the research preview positions its AI lab as a serious player beyond payments, without yet committing to a product or timeline. No benchmarks, model size or release date have been disclosed, so real-world performance remains unverified outside company demos.
The bigger question
If AI voices become indistinguishable from humans, should disclosure that you are talking to a machine become mandatory, and who would enforce that?
What to watch
Block has given early access only to a small group of AI safety researchers, with no public release date, benchmark results or model size disclosed yet. Watch for whether Block publishes independent testing, and whether CZ's GiggleAcademy pursues Dessa for children's education. Any consumer rollout would be the next concrete milestone to track.
