Analysis
What Dessa AI Can't Do (Yet): Facts, Tools, Reading and the Honest Limits
Dessa's demos are genuinely startling: a voice that laughs mid-sentence, gasps with gratitude and finishes your sentence for you. But a useful blog owes you the other article too — the one about what Dessa cannot do. Some of these limits are simple facts of its research-preview status. Others are structural consequences of the most interesting thing about it: Dessa has never read a word. Understanding both kinds is the fastest way to understand what Dessa actually is.
It cannot be used by you — yet
The most basic limit is access. Dessa is a research preview, announced October 23, 2026, with early access going first to a small group of AI safety researchers and a free public waitlist at block.xyz/dessa. Block has announced no public release date, no pricing and — as our guide to the Dessa AI API sets out — no API, no download and no weights. Every other limit in this article sits underneath that one: for now, Dessa is something you can listen to and read about, not something you can use, test or build on.
It has never read anything
Block's central claim is that Dessa was trained from scratch on human conversation audio only — no text, no pretrained models — the design we explain in how Dessa AI works. Sit with what that excludes. Dessa has never read a book, a manual, a contract, a webpage, a spreadsheet or a line of code. It cannot read a document you hand it, because reading was deliberately left out of its education.
That does not make it ignorant — people who learn by listening know a great deal, and conversation carries facts as well as feelings. But its knowledge is limited to what spoken conversation taught it, absorbed the way a person absorbs things they have heard: unevenly, without page numbers, and with no built-in way to check a claim against a source. A text-trained assistant can be pointed at documentation. Dessa, by design, learned the world secondhand, by ear.
Its factual reliability is unmeasured
Block has published no benchmarks, no parameter count, no training-data hours and no evaluation of factual accuracy. No independent researcher has yet published a hands-on assessment. So the honest statement is not "Dessa gets facts wrong" — nobody has the data to say that either. It is that factual reliability is unknown, in a field where its text-trained rivals publish scores and can be tested by anyone today.
The demos underline the point rather than settle it. They show fluency, timing and social awareness — including, memorably, Dessa finishing a speaker's sentence when they blanked on a movie title. They also show occasional very human awkwardness, including calling a speaker "Ronaldo". Fluency and accuracy are different virtues, and only one of them has been demonstrated. Our explainer what Dessa AI is keeps the two carefully separated.
It has no tools, no apps and no integrations
Modern assistants earn their keep by acting: searching the web, reading your calendar, calling functions, operating inside products. Block has announced none of that for Dessa — no tool use, no browsing, no integrations with Square, Cash App or anything else, and no product form at all. Speculation about voice ordering in Square restaurants or support lines inside Cash App may be reasonable as analysis of where a Block voice model could go; it is not an announcement, and this site labels it accordingly wherever we explore it.
Relatedly, nobody yet knows what Dessa does with long, structured tasks: multi-step instructions, careful lists, exact numbers read back and verified. Conversation is forgiving of approximation; administration is not. A model educated purely in conversation has never been shown doing paperwork, and paperwork is where much assistant value lives.
What it cannot do socially — the disclosure problem
One limit is ethical rather than technical. Dessa sounds human enough that a listener may not know they are talking to a machine — laughing, sighing, reacting with apparent feeling. Block itself argues that emotion-blind AI is a safety problem because it cannot detect distress; the mirror problem is an AI so emotionally fluent that disclosure, consent and impersonation safeguards become essential. Those safeguards are part of why safety researchers go first, a rollout we examined in is Dessa AI safe?. Until that work is published, what Dessa should not be allowed to do — undisclosed calls, simulated intimacy, persuasive sales — is an open question, not a settled policy.
Why the limits are the point
It would be easy to read this list as a dismissal. It is the opposite. Dessa is an experiment with a deliberately narrow education, testing how much of language, emotion and social intelligence lives in sound alone. Its lead, Ragavan Thurairatnam, called the launch "a GPT-3 moment for speech" — a claim about a turning point in a research direction, not a claim that the model already does everything. The limits above are the experiment's boundaries: no reading, no tools, no measured facts, no public access. If speech-only learning is going to matter beyond demos, it will matter by pushing against exactly those boundaries, in public, under independent testing.
When that testing arrives, this blog will cover what Dessa demonstrably can and cannot do with the same standard used here: Block's claims labelled as Block's, measurements labelled as measurements, and "not announced" wherever the record is silent.
Sources: Block's official Dessa page (block.xyz/dessa) for training, access and the absence of published benchmarks, pricing, API and release date; The Verge and TechCrunch coverage, October 2026.
Frequently asked questions
Can Dessa AI read documents or browse the web?
No. Dessa was trained on human speech only and has never read text. Block has announced no document reading, browsing, tool use or integrations for it.
Is Dessa AI accurate with facts?
Unknown. Block has published no benchmarks or factual-accuracy evaluation, and no independent hands-on assessment has appeared yet. Its demos show conversational fluency, which is a different thing from verified accuracy.
What is the biggest limit of Dessa AI today?
Access and evidence. It is a research preview the public cannot use, with no API, pricing or release date announced — and its core claims rest on Block's own demos until safety researchers publish independent findings.
🚀 Get Dessa launch alerts
Dessa is still a research preview. Join the free DessaAI Daily list and we will email you the day access opens — plus the week's biggest Dessa stories, once a week. No spam, no hype.
Form not loading? Subscribe on the signup page →