Safety

Is Dessa AI Safe? Scams, Impersonation and Safety-First Access

2026-10-28 · 7 min read · By DessaAI Daily

Block did something unusual when it announced Dessa on October 23, 2026: alongside demos of an AI laughing, gasping and sounding tearfully grateful, it described its own creation in cautionary terms and put AI safety researchers at the front of the access line. That order — safety researchers first, public later — is the clearest signal of what worries Block about a voice this convincing.

Why a human-sounding voice changes the risk

Most online deception still leaves clues. A scam email has odd phrasing. A robocall sounds flat, pauses in the wrong places, and collapses if you interrupt it. A full-duplex model trained only on real conversation attacks exactly those clues. Dessa listens while it speaks, backchannels naturally, reacts to emotion, and adjusts in real time. Block's demos, covered in our Dessa explainer, include it finishing sentences and responding with what sounds like genuine feeling.

None of that means Dessa has been used in a scam — there is no evidence of that, and Dessa is not publicly available. The risk is prospective, and it is the same risk raised by every realistic voice system: if a machine can hold a live, emotional, interruptible phone conversation, then “I could tell it was a robot” stops being a defense your family can rely on.

The specific risks researchers will probe

Impersonation. Could a system like this be made to sound like a specific person — a boss, a bank officer, a grandchild in trouble? Block has not said Dessa can clone a named person's voice, and you should not assume it can. But safety testing has to ask how close expressive speech gets to impersonation, and what safeguards stop it.

Phone and payment scams. Block owns Square and Cash App — businesses where a convincing voice asking for a payment, a code or an account change is not an abstract scenario. A voice that detects distress matters for safety in one direction; a voice that can manufacture urgency matters in the other.

Undisclosed AI identity. If a caller sounds human, laughs like a human and cries like a human, does the listener deserve to be told it is not human? Block argues today's emotion-blind voice AI is itself a safety problem because it cannot detect real pain or panic. The mirror-image problem is a voice so human that disclosure becomes the safety feature.

Emotional manipulation. A system that hears frustration, loneliness or confusion — and responds with perfectly timed warmth — could comfort a user or steer one. Where persuasion becomes manipulation is a research question, not a marketing one.

Why safety researchers go first

Block says early access goes to a small group of AI safety researchers, using institutional affiliation as a signal of legitimacy, so outsiders can probe for risks Block has not anticipated before wider release. That is a meaningful choice for a company that could have launched to waitlist hype alone. It also means the first independent reports about Dessa will likely be about failure modes, misuse and safeguards — not just how natural it sounds. Our waitlist guide explains how that access order works in practice.

The scams that exist right now are not Dessa

The immediate danger today is not the model; it is the hype around it. Block has announced no Dessa app, no download, no token, no paid tier and no pricing. So a “Dessa Pro” unlock, an early-access pass for sale, a Dessa app in a store, or a crypto token using its name is trading on the announcement, not affiliated with Block. The only official place is block.xyz/dessa, and joining the waitlist is free.

The same caution applies to voice scams generally, whether or not AI is involved: hang up and call back on a number you trust, never share a one-time code with a caller who contacted you, agree a family code word for emergency money requests, and treat urgency — “act now, do not tell anyone” — as the reddest flag of all. A convincing voice is a reason to verify through a second channel, not a reason to comply faster.

What Block has not announced — and why that matters

Block has published no benchmarks, no parameter count, no safety evaluation results and no release date, pricing or API. That honesty cuts both ways: we cannot claim Dessa is dangerous in specific measured ways, and we cannot claim it is safe. What we can verify is the process Block chose — a research preview, a waitlist, and safety researchers first — and the company's own framing that a voice AI unable to hear distress is unsafe, while one that sounds fully human creates new duties of care.

When the first safety researchers publish findings, those reports will matter more than any demo. This blog will cover them the day they appear, and update this guide if Block announces safeguards, disclosure rules or a wider release.

Sources: Block's official Dessa page (block.xyz/dessa); The Verge and TechCrunch coverage of the safety-researcher rollout, October 2026; PYMNTS coverage noting the split public reaction on scam risk, October 2026.

Frequently asked questions

← More Dessa AI articles Join the official waitlist ↗