Block’s Dessa AI talks like a person, but not always comfortably

Block, the company led by Jack Dorsey, has introduced a new AI model called Dessa. According to The Verge, Dessa was trained exclusively on human conversation, not on books or web text, and the company is promoting the model’s ability to chat in a way that feels more like talking to a real person, awkward pauses and all.

Why it matters

Most AI chatbots today are trained on large, general data sets that include books, articles, code and web pages. Block says Dessa is different. It was built from scratch using only actual human conversations. The goal: make interactions more natural and relatable, even if that means responses sometimes feel awkward or stilted. This is a deliberate break from the polished but sometimes robotic tone common to tools like ChatGPT.

According to Mashable, this approach is meant to close the gap between AI-generated dialogue and the unpredictable feel of real conversations. Block’s own demo chats show Dessa stumbling, hesitating, and occasionally responding in ways that feel slightly off, more like a new acquaintance than a customer service bot. For businesses, this signals a move toward AI that can handle messier, more realistic exchanges.

How Dessa’s training changes conversational AI

Block’s Dessa didn’t use written text for training. The company says it focused on audio from real conversations, aiming for a model that could capture the subtle patterns and hesitations of spoken language. This is different from how most models are built, and it shows. In example transcripts, Dessa’s replies can sound hesitant, with pauses and filler language that’s typical in everyday speech.

The Verge reports that Ragavan Thurairatnam, Dessa’s lead, shared examples where the AI’s responses felt a bit uncanny, sometimes friendly, sometimes awkward. This is intentional, not a flaw. By leaning into the quirks of human speech, Block wants Dessa to be more approachable, even at the risk of some odd moments. For companies, this could mean AI support and sales bots that feel less scripted and more adaptable in live conversation.

What this means if you run a business

For business owners, Dessa’s approach is a double-edged sword. On one hand, more natural AI could improve customer engagement, especially for support or sales roles where canned responses undermine trust. Customers may be more likely to stay on the line or finish a chat if the AI sounds like a patient, if slightly awkward, human.

But training AI only on real conversation introduces new risks. Awkward or unpredictable replies might frustrate users or reduce confidence in the service. Businesses will need to test these systems carefully to see if the benefits of realism outweigh the downsides of unpredictability. There is also the question of oversight: if the model mimics human awkwardness, it might also pick up on less desirable conversational habits unless monitored closely.

This release from Block signals a shift in the market. Companies relying on AI for customer contact will need to decide whether to adopt more human-like, imperfect bots or stick with the smoother, less realistic style of current mainstream models. For more on AI models and risk, see our recent coverage on Anthropic’s warnings about business risk.