Most Americans Say AI Is Bad for Society, boost.ai Ships Built-In Model Testing, and Block Previews a Speech Model Trained Without Text

Three items from the past week cover three layers of a chatbot product: how the public feels about it, how you prove it still works after a model change, and how it will sound if it speaks. Details come from press coverage and vendor announcements, so treat product claims as the vendors' own and note where a detail rests on secondary reports.

1. The Public Is Skeptical, and Your Users Are the Public

An ABC News/Washington Post/Ipsos poll conducted October 9-13 found that 57% of Americans say AI is bad for society, against 42% who say it is good. Only 2% call it extremely good, while 17% call it extremely bad. Per coverage, about two-thirds oppose building new data centers in the U.S. Opinion skews by age: 65% of people under 30 and 59% of those 30 to 39 say AI is bad for society. A separate Reuters/Ipsos poll from September asked a different question, so the numbers are not comparable.

A national poll is not a measurement of your own users, who chose to open your chatbot. But it sets the context they arrive in. Expect more questions about where answers come from, what is stored, and whether a human is available. Cheap trust signals are product work: visible citations, a clear memory page, and an obvious way to reach a person.

2. boost.ai v14: Testing Models Becomes a Platform Feature

On October 21, conversational AI vendor boost.ai announced version 14 of its platform, adding LLM model testing, AI-assisted integration building and a fully configurable chat panel, with rollout to customers over the coming weeks. The announcement is a press release, and I could not verify how the testing works in practice.

The direction is the useful part. Models are swapped, deprecated and repriced often, and every swap can quietly change behavior. Teams that treat a model change like a dependency upgrade need a regression suite they can run before switching. A minimal version is a fixed set of questions with checks on the answers:

const cases = [
  { q: 'How do I reset my password?', mustInclude: ['reset link'] },
  { q: 'Can I talk to a person?', mustInclude: ['agent', 'human'], anyOf: true },
];

for (const c of cases) {
  const answer = await ask(candidateModel, c.q);
  const hits = c.mustInclude.filter((s) => answer.toLowerCase().includes(s));
  const pass = c.anyOf ? hits.length > 0 : hits.length === c.mustInclude.length;
  console.log(pass ? 'PASS' : 'FAIL', c.q);
}

String checks are crude, and an LLM judge can score what they miss, but even this catches a model that stops offering a human handoff.

3. Block Previews Dessa, a Speech Model Trained on Conversation

Per secondary reports, Block released Dessa on October 23 as a research preview of a speech model it says was trained from scratch on human conversation alone, with no text and no pretrained models. Early access goes to a small group of AI safety researchers, and a public waitlist is open. There is no pricing or general release date, and I could not find independent verification of the training, hardware or performance claims. Block's own page does not clearly state the date.

If the claims hold, a model that learns timing, overlap and backchannels directly from conversational audio is a different approach from stitching speech-to-text, a language model and text-to-speech together. Until there is access and benchmarks, the practical step is to keep your voice pipeline modular so a native speech model can replace it later.

What Connects the Three

Skeptical users raise the bar for evidence, model churn raises the need for tests, and new modalities raise the number of things to test. Reliable, explainable behavior is the common thread.

A chart of the poll split by age group would make a good visual for a follow-up.

Conclusion

Build for an audience that is wary by default: show your sources, test every model change before shipping it, and stay skeptical of research previews until someone outside the lab can measure them.

— Maya