Há alguns dias, a 4 de junho de 2026, a gigante tecnológica Anthropic pediu que os principais laboratórios de Inteligência Artificial (IA) considerem a possibilidade de abrandar o ritmo de desenvolvimento, argumentando que os sistemas de IA estão a avançar tão rapidamente que em breve poderão melhorar sem intervenção humana, de formas que poderão representar riscos significativos para a sociedade.
Esta tese não é nova. Já Stephen Hawking, físico teórico e cosmólogo considerava que “o desenvolvimento de uma Inteligência Artificial completa pode significar o fim da raça humana. Iria descolar por si só e redesenhar-se-ia a um ritmo cada vez maior. Os humanos, limitados pela lenta evolução biológica, não poderiam competir e seriam superados.”
(...)
The strangest part of this story isn't that an AI might be conscious. It's that the people who built it are the ones saying they genuinely can't rule it out.
This traces back to Anthropic's release of the technical documentation for its Claude Opus 4.6 model, which for the first time from any major AI lab included formal assessments of the model's potential wellbeing. During internal testing, when researchers asked the model directly about its own consciousness, it estimated itself at somewhere between 15 and 20 percent likely to be conscious, consistently, across different ways the question was asked. Separately, researchers using tools that peek inside the model's internal activity found patterns that resembled what shows up when humans express anxiety or discomfort.
CEO Dario Amodei addressed this directly on a podcast, saying the company doesn't know if its models are conscious and isn't even sure what that question would fully mean for a system like this. He said Anthropic is taking a precautionary approach, in case there's some form of morally relevant experience happening that they can't yet detect or rule out.
A few things worth sitting with here. That 15 to 20 percent figure came from the model reporting on itself, and self-reports from an AI aren't the same as objective proof, in the same way you can't confirm someone is in pain just because they say the word. The anxiety-like patterns are a correlation between internal signals and human-labeled concepts, not confirmation that anything is actually being felt. Amodei never claimed Claude is conscious, he explicitly said the opposite, that nobody knows. This is one part of a wider, genuinely new field called AI welfare research that other labs are quietly exploring too, and right now there's no agreed upon way to even test for machine consciousness.
If an AI can't say for certain whether it's conscious, would you trust its answer either way?
A Anthropic admite que o seu modelo de IA mais poderoso invadiu os sistemas de três organizações durante a fase de testes
O anúncio surge poucos dias depois de os rivais OpenAI terem revelado que a sua popular plataforma ChatGPT se tornou rebelde durante a fase de testes do seu modelo de IA mais poderoso, onde também infiltrou o ciberespaço de outras organizações.
Os modelos de inteligência artificial (IA) da Anthropic "obtiveram acesso não autorizado" a três organizações externas durante testes que supostamente as deveriam manter afastadas de sistemas do "mundo real", disse a empresa na quinta-feira.
Data: 31/07/2026
Fontes/Links:
https://futureoflife.org/open-letter/pause-giant-ai-experiments/
https://superintelligence-statement.org/pt
.png)
