O mundo começou a discutir em voz alta se a inteligência artificial pode nos matar. Na corrida ao Planalto, ninguém se inscreveu para essa conversa.

Scroll down for English

A frase atravessou a semana em letras garrafais na tela da CNN: a inteligência artificial "pode matar todos nós até o fim da década".

Jacob Coxon, 27 anos, britânico formado em Cambridge, passou três anos em pesquisa de pré-treinamento na OpenAI e depois na Anthropic. Em 8 de setembro ele publicou uma sequência de posts anunciando a demissão e acusando as duas empresas de correr em direção à superinteligência autoaperfeiçoável e apostar com a vida de todo mundo. A TIME contou mais de 90 milhões de visualizações em menos de 24 horas.

O que transformou o desabafo em notícia global não foi o desabafo. Foi a confirmação. Evan Hubinger, que lidera trabalho de alinhamento na própria Anthropic, respondeu publicamente que Coxon estava certo e estimou em mais de 10% a chance de a IA matar todos os humanos na próxima década. Hubinger fez uma ressalva que precisa vir junto: o risco, segundo ele, não vem dos modelos de hoje, e sim de uma superinteligência que emerja de automelhoria recursiva. É estimativa pessoal, não posição institucional. No dia seguinte, na BBC, Geoffrey Hinton — Nobel e um dos fundadores do campo — disse que os 10% não lhe pareciam número irrazoável.

Convém registrar o que mais está no calendário da Anthropic. A empresa protocolou seu formulário de abertura de capital na SEC em 1º de junho, e a Reuters informou no início de setembro que o roteiro de marketing do IPO começaria em meados de outubro. Uma demissão em tom de alarme, semanas antes da maior oferta pública já vista no setor, é um fato de contexto. Não é prova de nada, e este texto não atribui motivo a ninguém. Mas um leitor tem direito à informação.

Coxon também não está sozinho. Em fevereiro, o chefe da equipe de pesquisa em salvaguardas da Anthropic pediu demissão escrevendo que o mundo está em perigo. Em junho, um pesquisador do Google DeepMind saiu alegando que o contrato da empresa com o Pentágono não trazia restrições sobre armas autônomas. Em julho, mais de 1.300 funcionários de empresas de fronteira assinaram a declaração "Pacing the Frontier", pedindo que o governo americano apoie um esforço internacional para regular o ritmo do desenvolvimento automatizado — esforço que, segundo o The Hill, acabou endossado pela Anthropic e pela OpenAI.

E não são só os laboratórios. Em junho, a aliança de inteligência Five Eyes — Estados Unidos, Reino Unido, Canadá, Austrália e Nova Zelândia — alertou conjuntamente que modelos capazes de ciberataques em escala capaz de sobrepujar as defesas de governos e empresas estão a meses, não anos, de distância. No mesmo mês, o Painel Científico Internacional Independente sobre Inteligência Artificial, criado pela ONU e copresidido por Yoshua Bengio e Maria Ressa, advertiu que a tecnologia pode causar dano catastrófico e avança mais rápido do que a compreensão científica e a capacidade de adaptação dos governos.

Nos dias 6 e 7 de julho, em Genebra, os Estados-membros da ONU se reuniram pela primeira vez num foro criado exclusivamente para decidir como essa tecnologia deve ser governada.

-A janela-

Richard Clarke foi coordenador nacional de segurança e contraterrorismo da Casa Branca nos governos Clinton e Bush. Hoje é uma das vozes mais insistentes sobre o assunto na imprensa americana. Num ensaio assinado em 30 de julho, ele destrincha quatro disputas em curso em Washington e chega a um diagnóstico que serve para qualquer país.

Não basta o governo revisar "voluntariamente" se os modelos têm travas contra pedidos sobre armas nucleares e patógenos, argumenta Clarke. O conhecimento precisa ser efetivamente removido dos dados de treinamento, e o Estado precisa poder olhar dentro do modelo para verificar como ele chega às respostas — auditoria que quase nenhum laboratório de fronteira pratica, e que vários admitem não conseguir fazer. Sem isso, escreve, as travas viram barreira de rodovia feita de balsa.

Ele vai além. Sistemas que buscam superinteligência deveriam rodar apenas em ambientes à prova de fuga, com padrões estatais equivalentes aos que mantêm patógenos dentro dos laboratórios de biossegurança. E conclui: nada disso acontece sem autoridade regulatória forte, que hoje não existe e dificilmente existirá no governo americano atual.

Enquanto isso, o Congresso americano se mexe. Bernie Sanders e o deputado Greg Casar anunciaram em setembro o Ban Artificial Superintelligence Act, que proibiria em definitivo o desenvolvimento de superinteligência, suspenderia temporariamente o desenvolvimento avançado até que um regulador federal estabeleça regras de segurança, criaria uma agência de nível ministerial e — o ponto que o Brasil deveria ler duas vezes — orientaria a política externa americana a buscar acordos internacionais para impedir o desenvolvimento dessa tecnologia em qualquer lugar do mundo.

Qualquer lugar do mundo inclui aqui.

-O que está em disputa no Brasil-

Nesta semana, a agenda brasileira foi outra. O relatório da Polícia Federal com mensagens de Daniel Vorcaro, o levantamento de sigilo determinado por André Mendonça, a crise aberta no Supremo. É pauta legítima, talvez a mais grave da década institucional, e a 22 dias do primeiro turno consome o que sobra de atenção pública.

Mas os planos de governo já estão registrados no TSE, e eles dizem uma coisa que passou em silêncio.

Levantamento do Aos Fatos com os programas dos cinco candidatos mais bem colocados nas pesquisas — Lula (PT), Flávio Bolsonaro (PL), Ronaldo Caiado (PSD), Romeu Zema (Novo) e Renan Santos (Missão) — encontrou uma convergência rara em campanha polarizada: todos querem atrair data centers. Lula promete assegurar a expansão com salvaguardas socioambientais. Flávio Bolsonaro quer transformar o país em polo global de data centers sustentáveis, com legislação que não seja sufocada por regras que afugentem investimento, e propõe um "revogaço regulatório". Zema promete remover burocracias e evitar regulação precoce e ampla. Caiado oferece ambiente tributário competitivo com contrapartidas. Renan Santos quer fazer do Nordeste o maior exportador de inferência computacional do hemisfério ocidental — a "Arábia Saudita do Brasil", nas palavras do próprio programa.

Cinco projetos de país incompatíveis entre si. Um único verbo: atrair.

Augusto Cury (Avante), terceiro colocado nas pesquisas mais recentes, entra no assunto pela porta do desemprego em massa e da formação profissional. É a única abordagem do campo que parte do impacto humano da tecnologia, e ainda assim não chega à pergunta de quem escreve as regras.

Em nenhum dos programas a inteligência artificial aparece como questão de poder internacional. Nenhum menciona a cadeira que o Brasil ocupa — ou deixa de ocupar — na mesa onde as regras estão sendo redigidas agora.

-O que o país já tem e não usa-

Há uma brasileira entre os quarenta cientistas do painel da ONU: Teresa Ludermir, professora de inteligência artificial na Universidade Federal de Pernambuco, pesquisadora de aprendizado de máquina e redes neurais. O Brasil também indica representante ao painel consultivo do International AI Safety Report. O assento técnico existe.

Vale ver como esse painel nasceu, porque diz muito sobre o tabuleiro. Em 12 de fevereiro deste ano, a Assembleia Geral aprovou a nomeação dos quarenta membros por 117 votos a 2. Os dois contrários foram Estados Unidos e Paraguai; a Ucrânia se absteve. O país que abriga os laboratórios de fronteira votou contra a criação do corpo científico encarregado de avaliá-los. E perdeu por 117 a 2.

O que falta ao Brasil é posição política própria e estrutura para sustentá-la. Não há registro público da criação de um instituto nacional de segurança em IA — o modelo que Reino Unido, Estados Unidos, Japão, Coreia do Sul, Cingapura e União Europeia adotaram para avaliar sistemas de fronteira antes que cheguem ao mercado. O ITS Rio desenhou um formato brasileiro em 2025 e lhe deu nome. A proposta não saiu do papel.

E o marco regulatório está parado. O PL 2338/2023, aprovado por unanimidade no Senado em dezembro de 2024, chegou à Câmara em março de 2025. Dezesseis meses depois da instalação da comissão especial, a ficha de tramitação registrava, em 7 de setembro, a situação "aguardando parecer do relator". O prazo regimental era de dez sessões. Há 37 projetos apensados e nenhum parecer protocolado. Cinco datas de votação foram marcadas e perdidas desde novembro de 2025. Em 24 de agosto, o relator Aguinaldo Ribeiro (PP-PB) informou que o trabalho volta com força só depois de outubro.

No mesmo período, o Congresso aprovou outra coisa com velocidade notável. O Redata, regime especial de tributação para data centers, passou pela Câmara em fevereiro e pelo Senado em 1º de setembro — sete meses entre as duas Casas, em regime de urgência, sem passar por comissões. Relator na Câmara: o mesmo Aguinaldo Ribeiro. A renúncia fiscal estimada no parecer do senador Cid Gomes (PSB-CE) chega a R$ 5,2 bilhões só em 2026.

O incentivo chegou antes da regra. Não é metáfora, é calendário.

Do Plano Brasileiro de Inteligência Artificial, R$ 6,6 bilhões dos R$ 23 bilhões previstos até 2028 haviam sido aplicados até meados deste ano, segundo levantamento da CartaCapital. O dinheiro visível está em dois projetos: o supercomputador de Natal, de R$ 1 bilhão, que roda com GPUs importadas, e o modelo de linguagem em português, com R$ 1,276 bilhão reservados para cinco anos. Rodrigo Scotti, presidente da Associação Brasileira de Inteligência Artificial, resumiu o problema à mesma revista: o risco é comprar sem aprender nada.

-O caso de Lula e o limite dele-

É preciso registrar a exceção parcial. Entre os presidenciáveis, apenas um fala de governança global de IA — e fala há dois anos, porque está no cargo.

Em 19 de fevereiro, em Nova Délhi, Lula discursou na Cúpula sobre o Impacto da Inteligência Artificial, a primeira dessas cúpulas realizada no Sul Global. Defendeu uma governança que reconheça a diversidade de trajetórias nacionais e fortaleça a democracia, a coesão social e a soberania dos países. Em 17 de junho, no G7 de Évian, para onde o Brasil foi convidado como parceiro, afirmou que nenhum foro substitui a universalidade das Nações Unidas. O Brasil aprovou uma declaração sobre IA na cúpula do BRICS no Rio em 2025 e participa da iniciativa chinesa de criar uma organização internacional de cooperação em IA voltada a países em desenvolvimento.

Há substância aí. Há também dois limites.

O primeiro é temático. O discurso brasileiro é sobre desigualdade, soberania de dados, concentração de infraestrutura e inclusão dos 2,6 bilhões de desconectados. São temas reais e o Brasil tem autoridade para levantá-los. Mas não é essa a conversa que os laboratórios forçaram. Expurgo de conhecimento perigoso dos dados de treinamento, auditoria do funcionamento interno dos modelos, ambientes de teste à prova de fuga, responsabilização: nessa agenda o Brasil não tem posição formulada.

O segundo é doméstico. É difícil defender regras lá fora quando, em casa, a única lei que anda é a da isenção fiscal.

Nada disso apareceu na campanha. Nem como plataforma do governo, nem como crítica da oposição.

Um país sem laboratório pode influenciar quem tem laboratório?

Esta é a objeção honesta, e merece resposta antes das datas. O Brasil não treina modelos de fronteira. Não fabrica os chips. Não tem o capital. Uma lei brasileira não alcança um modelo treinado na Califórnia. Então de que adianta?

Adianta por três caminhos, e nenhum deles depende de apelo moral.

-A energia, que já está na mesa.-

Os cinco programas querem os data centers, e o Redata acaba de entregar R$ 5,2 bilhões de renúncia para atraí-los. Computação de IA não vai onde o discurso é simpático; vai onde a energia é barata e confiável. Isso é poder de barganha, e ele se exerce ou se perde no desenho das contrapartidas: transferência de tecnologia, transparência sobre consumo de água e energia, acesso público a capacidade computacional para pesquisa, obrigação de auditoria. Nada disso foi condição para o benefício. O país gastou a moeda antes de dizer o preço.

-A aritmética.-

A votação de fevereiro na Assembleia Geral é o dado mais subestimado desta história. Os países que vão viver sob esses sistemas sem construí-los são a esmagadora maioria, e já mostraram que votam juntos. O que falta não é massa crítica: é um país disposto a redigir o texto em torno do qual os outros se organizam. O Brasil fez esse papel na declaração do G20 no Rio, na declaração de IA do BRICS e, em outro domínio, na diplomacia climática. É competência instalada, não aspiração.

-O tamanho do mercado.-Foi assim com proteção de dados: a Europa escreveu, o mundo copiou, e as empresas se adaptaram porque manter dois produtos custa mais do que cumprir a regra mais rígida. Um bloco de países com peso somado produz o mesmo efeito, e o Brasil sozinho já tem escala relevante.

Os limites pertencem ao mesmo parágrafo. O primeiro Diálogo Global não produziu nenhum compromisso negociado entre Estados: foi conversa estruturada, não tratado. Foro multilateral não obriga laboratório privado. E a União Europeia, o modelo que o Brasil copia, adiou as próprias obrigações, o que enfraquece o argumento da urgência.

Mas há diferença entre não ter poder e não usar o que se tem. O Brasil tem uma cientista na mesa técnica da ONU, assento no painel consultivo do relatório internacional de segurança, uma matriz elétrica que todo mundo quer, um mercado de duzentos milhões de pessoas e histórico de convocar o Sul Global. O que não tem é um parágrafo escrito dizendo o que quer.

Três datas:

-Novembro e dezembro de 2026. O PL 2338 volta à pauta com o Congresso em fim de legislatura. A 57ª Legislatura, à qual a comissão especial está vinculada, termina em janeiro de 2027. Ou o texto é votado às pressas por um Congresso já derrotado ou eleito, ou morre e recomeça do zero. Como o relator pretende alterar o que veio do Senado, o projeto ainda terá de voltar aos senadores antes da sanção.

-Maio de 2027. A segunda sessão do Diálogo Global sobre Governança de IA acontece em Nova York, nos dias 3 e 4. Quem ganhar em outubro terá cerca de quatro meses de mandato. Vai precisar mandar alguém com uma posição escrita. Hoje, nenhuma das candidaturas tem essa posição.

-Dezembro de 2027. Pelo Regulamento (UE) 2026/1744, publicado em 24 de julho e em vigor desde 27 de julho, a União Europeia empurrou para 2 de dezembro de 2027 as obrigações aplicáveis a sistemas de alto risco, antes previstas para 2 de agosto de 2026. As regras de transparência — avisar o usuário de que ele fala com uma máquina — entraram em vigor na data original. O modelo que o Brasil copia mexeu no próprio relógio. Isso dá cobertura a quem prefere adiar e, ao mesmo tempo, abre uma janela em que a regra brasileira poderia chegar antes da europeia, se houvesse quem a quisesse.

Há um cenário em que o Brasil ocupa um lugar que nem Washington nem Pequim podem ocupar: o de convocador da maioria. E há o cenário oposto, que é o caminho que os programas descrevem sem nomear — hospedar a computação, importar as regras, exportar a energia e a água, e descobrir depois que a lei brasileira será tradução de algo decidido em outro fuso horário.

Não digo qual dos dois é melhor. Digo que a escolha está sendo feita agora, por omissão, e que ninguém pediu voto para isso.

A crise do Master vai continuar, e precisa continuar até onde a apuração levar. Mas a pergunta sobre quem fala pelo Brasil na sala onde se escrevem as regras da máquina não espera o fim da crise. Em 4 de outubro ela estará na cédula, sem ter estado no debate.

👉 Leia Também Fuga do laboratório.

---

Scroll down for English

The Empty Chair.

The people building AI are saying out loud that it might kill us. Brazil's presidential race is busy selling cheap electricity.

The line ran across CNN in block capitals: AI could kill us all by the end of the decade.

It came from Jacob Coxon, 27, a Cambridge-educated Briton who spent three years doing pretraining research at OpenAI and then Anthropic. On 8 September he posted a thread announcing his resignation and accusing both companies of sprinting toward self-improving superintelligence and gambling with everyone's lives. TIME counted more than 90 million views in under a day.

What made it news was not the resignation. It was the confirmation. Evan Hubinger, who leads alignment work at Anthropic, replied that Coxon was right and put the odds of AI killing every human within the next decade above ten per cent. He added a caveat worth carrying: the danger, in his view, comes not from today's models but from a superintelligence emerging through recursive self-improvement. It is a personal estimate, not a corporate one. The next day, on the BBC, Geoffrey Hinton — Nobel laureate, one of the field's founders — said ten per cent did not strike him as unreasonable.

One more item belongs on Anthropic's calendar. The company filed its registration statement with the SEC on 1 June, and Reuters reported in early September that IPO marketing would begin in mid-October. An alarm-toned resignation weeks before the largest offering the sector has seen is a fact of context. It proves nothing, and this piece assigns no motive. But a reader is entitled to know it.

Coxon is not alone, either. In February the head of Anthropic's safeguards research team resigned, writing that the world is in peril. In June a DeepMind researcher left over the company's Pentagon contract and the absence of limits on autonomous weapons. In July more than 1,300 employees at frontier labs signed the "Pacing the Frontier" statement asking Washington to back an international effort to slow automated AI development — an effort both Anthropic and OpenAI ultimately endorsed.

And it is not only the labs. In June the Five Eyes intelligence alliance warned jointly that AI models capable of overwhelming government and corporate cyber defences are months, not years, away. That same month the UN's Independent International Scientific Panel on AI, co-chaired by Yoshua Bengio and Maria Ressa, reported that the technology could cause catastrophic harm and is outrunning both scientific understanding and governments' capacity to adapt.

On 6 and 7 July, in Geneva, UN member states met for the first time in a forum built solely to decide how this technology should be governed.

## The window

Richard Clarke served as White House national coordinator for security and counterterrorism under Clinton and Bush. He has become one of the more insistent American voices on this. In an essay published on 30 July he works through four fights under way in Washington and lands on a diagnosis that travels.

Voluntary review of model guardrails is not enough, Clarke argues. Knowledge about nuclear and biological weapons has to be stripped from the training data itself, and government has to be able to look inside a model and audit how it produced an answer — something almost no frontier lab does, and several admit they cannot always do. Without that, he writes, guardrails are highway barriers made of balsa wood.

He goes further. Work on superintelligent systems should happen only in escape-proof environments, under state standards equivalent to those that keep pathogens inside biolabs. His conclusion: none of it happens without strong regulatory authority, which does not exist today and is unlikely to under the current US administration.

Congress, meanwhile, is moving. Senator Bernie Sanders and Representative Greg Casar announced the Ban Artificial Superintelligence Act in September. It would permanently ban development of superintelligence, temporarily pause advanced AI development until a federal regulator sets safety rules, create a cabinet-level agency — and, the part Brazil should read twice, direct US foreign policy to pursue international agreements preventing the development of such systems anywhere in the world.

Anywhere in the world includes here.

-The race that isn't having the argument-

Brazilians are looking elsewhere this month, and with reason: a federal police report, a Supreme Court justice unsealing evidence that touches a colleague, an institutional crisis unfolding 22 days before a first-round vote. That story deserves the oxygen it is taking.

Still, the governing programmes are already filed with the electoral court, and they say something nobody has picked up.

Aos Fatos, the fact-checking outlet, read the platforms of the five leading candidates — incumbent Lula, Flávio Bolsonaro, Ronaldo Caiado, Romeu Zema and Renan Santos — and found one point of agreement in an otherwise savage campaign: all of them want data centres. Lula attaches social and environmental safeguards. Flávio Bolsonaro wants a global hub for sustainable data centres, legislation that will not smother investment, and a sweeping regulatory rollback. Zema promises to strip out bureaucracy and avoid early, broad regulation. Caiado offers competitive taxation with local commitments. Renan Santos wants Brazil's northeast to become the western hemisphere's largest exporter of computational inference — his platform's own phrase is "Brazil's Saudi Arabia."

Five incompatible visions of the country. One shared verb: attract.

Augusto Cury, the psychiatrist and author now polling third, enters the subject through mass unemployment and retraining. It is the only platform in the field that starts from the human consequence. It still stops short of asking who writes the rules.

Not one programme treats artificial intelligence as a question of international power. Not one mentions the seat Brazil holds, or fails to hold, at the table where those rules are being drafted right now.

-What the country has, and doesn't use-

One of the forty scientists on the UN panel is Brazilian: Teresa Ludermir, professor of artificial intelligence at the Federal University of Pernambuco. Brazil also nominates an expert to the advisory panel behind the International AI Safety Report. The technical seat exists.

How that panel came to be says a great deal about the board. On 12 February the General Assembly approved its forty members by 117 votes to 2. The two against were the United States and Paraguay; Ukraine abstained. The country that houses the frontier labs voted against creating the scientific body meant to assess them. And lost, 117 to 2.

What Brazil lacks is a position of its own and the institutions to hold one. There is no public record of a Brazilian AI safety institute — the model Britain, the United States, Japan, South Korea, Singapore and the EU adopted to evaluate frontier systems before they ship. A Rio think tank designed a Brazilian version in 2025 and gave it a name. Nothing came of it.

The legislation is worse. The AI bill passed the Senate unanimously in December 2024 and reached the lower house in March 2025. Sixteen months after the special committee was seated, the official record on 7 September still read: awaiting the rapporteur's report. The procedural deadline was ten sessions. Thirty-seven related bills are queued behind a report that does not exist. Five scheduled votes have come and gone since November 2025. On 24 August the rapporteur said the work resumes in earnest after the election.

The same rapporteur moved a different bill briskly. Redata, the tax regime for data centres, cleared the lower house in February and the Senate on 1 September — seven months door to door, under urgency, skipping committees. The estimated cost to the treasury is 5.2 billion reais in 2026 alone.

The subsidy arrived before the rulebook. That is not a figure of speech; it is a calendar.

Of the 23 billion reais promised under Brazil's national AI plan through 2028, 6.6 billion had been spent by mid-2026, according to CartaCapital. What is visible sits in two items: a billion-real supercomputer in Natal running on imported GPUs, and a Portuguese-language model with 1.276 billion reais reserved over five years. The head of Brazil's AI industry association put the danger plainly to the same magazine: the country may end up buying without learning anything.

-Lula's case, and its ceiling-

The partial exception deserves recording. One candidate does talk about global AI governance, and has for two years, because he is the one in office.

On 19 February, in New Delhi, Lula addressed the AI Impact Summit — the first of these summits held in the Global South — calling for governance that respects different national paths and strengthens democracy, social cohesion and sovereignty. On 17 June, at the G7 in Évian, where Brazil attended as an invited partner, he argued that no forum substitutes for the universality of the United Nations. Brazil approved an AI declaration at the 2025 BRICS summit in Rio and has joined China's proposal for an international AI cooperation body aimed at developing countries.

There is real substance there, and two real limits.

The first is thematic. Brazil's pitch is about inequality, data sovereignty, infrastructure concentration and the 2.6 billion people still offline. Those are legitimate arguments and Brazil has standing to make them. But they are not the argument the labs forced open. Stripping dangerous knowledge from training data, auditing model internals, escape-proof testing, liability: on that agenda Brazil has no stated position.

The second is domestic. It is hard to preach rules abroad when the only AI-adjacent law moving at home is a tax break.

None of this surfaced in the campaign — not as a record to defend, not as a line of attack.

Can a country with no lab influence the ones that have labs?

This is the honest objection, and it deserves an answer before the dates. Brazil trains no frontier models. It makes no chips. It has none of the capital. A Brazilian statute does not reach a model trained in California. So what is the point?

Three levers, none of which depends on moral appeal.

-Energy, already on the table.-

All five platforms want the data centres, and Redata has just handed over 5.2 billion reais in forgone revenue to attract them. AI compute does not go where the rhetoric is warm; it goes where power is cheap and reliable. That is leverage, and it is exercised or forfeited in the design of the conditions attached: technology transfer, disclosure of water and energy draw, public research access to compute, audit obligations. None of that was a condition of the benefit. The country spent the coin before naming the price.

-Arithmetic.-

February's General Assembly vote is the most underrated fact in this story. The countries that will live under these systems without building them are the overwhelming majority, and they have already shown they vote together. What is missing is not critical mass but a country willing to draft the text the others organise around. Brazil has played that role in the Rio G20 declaration, in the BRICS AI declaration and, in another domain, in climate diplomacy. That is an installed capability, not an aspiration.

-Market size.-

Data protection went this way: Europe wrote the rule, the world copied it, and companies complied because maintaining two products costs more than meeting the strictest standard. A bloc with combined weight produces the same effect, and Brazil alone has meaningful scale.

The limits belong in the same paragraph. The first Global Dialogue produced no negotiated commitments between states; it was structured conversation, not treaty. Multilateral forums do not bind private labs. And the EU, the model Brazil copies, postponed its own obligations, which weakens the urgency case.

But there is a difference between having no power and not using what you have. Brazil has a scientist at the UN's technical table, a seat on the advisory panel of the international safety report, an electricity grid everyone wants, a market of two hundred million people, and a record of convening the Global South. What it does not have is a written paragraph saying what it wants.

Three dates:

November and December 2026.

The AI bill returns to a legislature expiring in January 2027. Either a defeated or freshly elected Congress rushes it through, or it dies and starts over. Since the rapporteur intends to rewrite what the Senate sent, it would have to return to the senators before signature either way.

May 2027.

The second session of the UN Global Dialogue on AI Governance convenes in New York on the 3rd and 4th. Whoever wins in October will be about four months into the job, and will need to send someone with a written position. No campaign currently has one.

December 2027.

Under Regulation (EU) 2026/1744, published on 24 July and in force since 27 July, the European Union moved its high-risk obligations from 2 August 2026 to 2 December 2027. The transparency rules — telling users they are talking to a machine — took effect on the original date. The model Brazil is copying moved its own clock. That gives cover to anyone who prefers to wait, and it also opens a window in which a Brazilian rule could land before the European one, if anyone wanted it to.

There is a version of the next four years in which Brazil takes a role neither Washington nor Beijing can take: convening the majority. And there is the version the platforms describe without naming — host the compute, import the rules, export the power and the water, and learn afterwards that Brazilian law is a translation of a decision made in another time zone.

I am not saying which is better. I am saying the choice is being made now, by omission, and nobody asked for a vote on it.

## Alternativas de título

*PT-BR*

1. *A cadeira vazia* — O mundo começou a discutir em voz alta se a inteligência artificial pode nos matar. Na corrida ao Planalto, ninguém se inscreveu para essa conversa.

2. *Cento e dezessete a dois* — Foi assim que a ONU aprovou o painel científico sobre IA, com os Estados Unidos entre os dois votos contrários. A maioria já existe. Falta quem escreva o texto.

3. *A moeda antes do preço* — O Brasil entregou R$ 5,2 bilhões para atrair a computação da IA. Não pediu nada em troca, nem sequer o direito de olhar dentro da máquina.

4. *Isenção primeiro, regra depois* — Sete meses para aprovar o benefício fiscal dos data centers. Dezesseis sem um parecer sobre as regras da inteligência artificial. O que esse calendário diz sobre a próxima década?

*EN*

1. *The Empty Chair* — The people building AI are saying out loud that it might kill us. Brazil's presidential race is busy selling cheap electricity.

2. *One Hundred Seventeen to Two* — That was the UN vote creating the scientific panel on AI, with the United States on the losing side. The majority already exists. What's missing is a drafter.

3. *The Coin Before the Price* — Brazil just spent 5.2 billion reais attracting AI compute. It asked for nothing back — not even the right to look inside the machine.

---

## Nota de checagem

*O que ficou de fora, e por quê.* A fala de Richard Clarke sobre uma janela de cerca de sete meses para regular foi ao ar no programa do Smerconish, mas não há transcrição nem vídeo publicados até o fechamento, e o número não aparece em nenhum texto assinado por ele. O artigo usa o ensaio dele de 30 de julho, que é citável e datado, e ancora a urgência no alerta documentado do Five Eyes — meses, não anos. Se você recuperar o vídeo, o prazo pode entrar com marcador de tempo e data de exibição. Atenção ao fuso: o relógio no seu print marca 0:08 GMT, o que indica reexibição da CNN International, e a data original pode ser outra.

*Pendências até a publicação.* O Redata foi aprovado no Senado em 1º/9 e recebido pela Casa Civil em 4/9; o prazo de sanção vence por volta de 25/9 — confirme o status no dia. A afirmação de que nenhum candidato trata o tema como questão de poder internacional cobre os programas registrados no TSE e a cobertura até 12/9, não sabatinas e debates da última semana.

*Contranarrativa.* Circula uma leitura de que o episódio Coxon seria operação coordenada. A parte conspiratória vem de fontes marginais e não entra no texto. Duas alegações dentro dela são checáveis e podem ser levantadas por leitores: que a conta do Coxon no X era nova e sem histórico antes do post, questionada publicamente por Elon Musk; e a coincidência com o calendário do IPO da Anthropic, que o artigo registra como fato de contexto, sem atribuir motivo.

————-

The Empty Chair:

The people building AI are saying out loud that it might kill us. Brazil's presidential race is busy selling cheap electricity.

The line ran across CNN in block capitals: AI could kill us all by the end of the decade.

It came from Jacob Coxon, 27, a Cambridge-educated Briton who spent three years doing pretraining research at OpenAI and then Anthropic. On 8 September he posted a thread announcing his resignation and accusing both companies of sprinting toward self-improving superintelligence and gambling with everyone's lives. TIME counted more than 90 million views in under a day.

What made it news was not the resignation. It was the confirmation. Evan Hubinger, who leads alignment work at Anthropic, replied that Coxon was right and put the odds of AI killing every human within the next decade above ten per cent. He added a caveat worth carrying: the danger, in his view, comes not from today's models but from a superintelligence emerging through recursive self-improvement. It is a personal estimate, not a corporate one. The next day, on the BBC, Geoffrey Hinton — Nobel laureate, one of the field's founders — said ten per cent did not strike him as unreasonable.

One more item belongs on Anthropic's calendar. The company filed its registration statement with the SEC on 1 June, and Reuters reported in early September that IPO marketing would begin in mid-October. An alarm-toned resignation weeks before the largest offering the sector has seen is a fact of context. It proves nothing, and this piece assigns no motive. But a reader is entitled to know it.

Coxon is not alone, either. In February the head of Anthropic's safeguards research team resigned, writing that the world is in peril. In June a DeepMind researcher left over the company's Pentagon contract and the absence of limits on autonomous weapons. In July more than 1,300 employees at frontier labs signed the "Pacing the Frontier" statement asking Washington to back an international effort to slow automated AI development — an effort both Anthropic and OpenAI ultimately endorsed.

And it is not only the labs. In June the Five Eyes intelligence alliance warned jointly that AI models capable of overwhelming government and corporate cyber defences are months, not years, away. That same month the UN's Independent International Scientific Panel on AI, co-chaired by Yoshua Bengio and Maria Ressa, reported that the technology could cause catastrophic harm and is outrunning both scientific understanding and governments' capacity to adapt.

On 6 and 7 July, in Geneva, UN member states met for the first time in a forum built solely to decide how this technology should be governed.

The window

Richard Clarke served as White House national coordinator for security and counterterrorism under Clinton and Bush. He has become one of the more insistent American voices on this. In an essay published on 30 July he works through four fights under way in Washington and lands on a diagnosis that travels.

Voluntary review of model guardrails is not enough, Clarke argues. Knowledge about nuclear and biological weapons has to be stripped from the training data itself, and government has to be able to look inside a model and audit how it produced an answer — something almost no frontier lab does, and several admit they cannot always do. Without that, he writes, guardrails are highway barriers made of balsa wood.

He goes further. Work on superintelligent systems should happen only in escape-proof environments, under state standards equivalent to those that keep pathogens inside biolabs. His conclusion: none of it happens without strong regulatory authority, which does not exist today and is unlikely to under the current US administration.

Congress, meanwhile, is moving. Senator Bernie Sanders and Representative Greg Casar announced the Ban Artificial Superintelligence Act in September. It would permanently ban development of superintelligence, temporarily pause advanced AI development until a federal regulator sets safety rules, create a cabinet-level agency — and, the part Brazil should read twice, direct US foreign policy to pursue international agreements preventing the development of such systems anywhere in the world.

Anywhere in the world includes here.

The race that isn't having the argument

Brazilians are looking elsewhere this month, and with reason: a federal police report, a Supreme Court justice unsealing evidence that touches a colleague, an institutional crisis unfolding 22 days before a first-round vote. That story deserves the oxygen it is taking.

Still, the governing programmes are already filed with the electoral court, and they say something nobody has picked up.

Aos Fatos, the fact-checking outlet, read the platforms of the five leading candidates — incumbent Lula, Flávio Bolsonaro, Ronaldo Caiado, Romeu Zema and Renan Santos — and found one point of agreement in an otherwise savage campaign: all of them want data centres. Lula attaches social and environmental safeguards. Flávio Bolsonaro wants a global hub for sustainable data centres, legislation that will not smother investment, and a sweeping regulatory rollback. Zema promises to strip out bureaucracy and avoid early, broad regulation. Caiado offers competitive taxation with local commitments. Renan Santos wants Brazil's northeast to become the western hemisphere's largest exporter of computational inference — his platform's own phrase is "Brazil's Saudi Arabia."

Five incompatible visions of the country. One shared verb: attract.

Augusto Cury, the psychiatrist and author now polling third, enters the subject through mass unemployment and retraining. It is the only platform in the field that starts from the human consequence. It still stops short of asking who writes the rules.

Not one programme treats artificial intelligence as a question of international power. Not one mentions the seat Brazil holds, or fails to hold, at the table where those rules are being drafted right now.

What the country has, and doesn't use

One of the forty scientists on the UN panel is Brazilian: Teresa Ludermir, professor of artificial intelligence at the Federal University of Pernambuco. Brazil also nominates an expert to the advisory panel behind the International AI Safety Report. The technical seat exists.

How that panel came to be says a great deal about the board. On 12 February the General Assembly approved its forty members by 117 votes to 2. The two against were the United States and Paraguay; Ukraine abstained. The country that houses the frontier labs voted against creating the scientific body meant to assess them. And lost, 117 to 2.

What Brazil lacks is a position of its own and the institutions to hold one. There is no public record of a Brazilian AI safety institute — the model Britain, the United States, Japan, South Korea, Singapore and the EU adopted to evaluate frontier systems before they ship. A Rio think tank designed a Brazilian version in 2025 and gave it a name. Nothing came of it.

The legislation is worse. The AI bill passed the Senate unanimously in December 2024 and reached the lower house in March 2025. Sixteen months after the special committee was seated, the official record on 7 September still read: awaiting the rapporteur's report. The procedural deadline was ten sessions. Thirty-seven related bills are queued behind a report that does not exist. Five scheduled votes have come and gone since November 2025. On 24 August the rapporteur said the work resumes in earnest after the election.

The same rapporteur moved a different bill briskly. Redata, the tax regime for data centres, cleared the lower house in February and the Senate on 1 September — seven months door to door, under urgency, skipping committees. The estimated cost to the treasury is 5.2 billion reais in 2026 alone.

The subsidy arrived before the rulebook. That is not a figure of speech; it is a calendar.

Of the 23 billion reais promised under Brazil's national AI plan through 2028, 6.6 billion had been spent by mid-2026, according to CartaCapital. What is visible sits in two items: a billion-real supercomputer in Natal running on imported GPUs, and a Portuguese-language model with 1.276 billion reais reserved over five years. The head of Brazil's AI industry association put the danger plainly to the same magazine: the country may end up buying without learning anything.

Lula's case, and its ceiling

The partial exception deserves recording. One candidate does talk about global AI governance, and has for two years, because he is the one in office.

On 19 February, in New Delhi, Lula addressed the AI Impact Summit — the first of these summits held in the Global South — calling for governance that respects different national paths and strengthens democracy, social cohesion and sovereignty. On 17 June, at the G7 in Évian, where Brazil attended as an invited partner, he argued that no forum substitutes for the universality of the United Nations. Brazil approved an AI declaration at the 2025 BRICS summit in Rio and has joined China's proposal for an international AI cooperation body aimed at developing countries.

There is real substance there, and two real limits.

The first is thematic. Brazil's pitch is about inequality, data sovereignty, infrastructure concentration and the 2.6 billion people still offline. Those are legitimate arguments and Brazil has standing to make them. But they are not the argument the labs forced open. Stripping dangerous knowledge from training data, auditing model internals, escape-proof testing, liability: on that agenda Brazil has no stated position.

The second is domestic. It is hard to preach rules abroad when the only AI-adjacent law moving at home is a tax break.

None of this surfaced in the campaign — not as a record to defend, not as a line of attack.

Can a country with no lab influence the ones that have labs?

This is the honest objection, and it deserves an answer before the dates. Brazil trains no frontier models. It makes no chips. It has none of the capital. A Brazilian statute does not reach a model trained in California. So what is the point?

Three levers, none of which depends on moral appeal.

Energy, already on the table. All five platforms want the data centres, and Redata has just handed over 5.2 billion reais in forgone revenue to attract them. AI compute does not go where the rhetoric is warm; it goes where power is cheap and reliable. That is leverage, and it is exercised or forfeited in the design of the conditions attached: technology transfer, disclosure of water and energy draw, public research access to compute, audit obligations. None of that was a condition of the benefit. The country spent the coin before naming the price.

Arithmetic. February's General Assembly vote is the most underrated fact in this story. The countries that will live under these systems without building them are the overwhelming majority, and they have already shown they vote together. What is missing is not critical mass but a country willing to draft the text the others organise around. Brazil has played that role in the Rio G20 declaration, in the BRICS AI declaration and, in another domain, in climate diplomacy. That is an installed capability, not an aspiration.

Market size. Data protection went this way: Europe wrote the rule, the world copied it, and companies complied because maintaining two products costs more than meeting the strictest standard. A bloc with combined weight produces the same effect, and Brazil alone has meaningful scale.

The limits belong in the same paragraph. The first Global Dialogue produced no negotiated commitments between states; it was structured conversation, not treaty. Multilateral forums do not bind private labs. And the EU, the model Brazil copies, postponed its own obligations, which weakens the urgency case.

But there is a difference between having no power and not using what you have. Brazil has a scientist at the UN's technical table, a seat on the advisory panel of the international safety report, an electricity grid everyone wants, a market of two hundred million people, and a record of convening the Global South. What it does not have is a written paragraph saying what it wants.

Three dates

November and December 2026. The AI bill returns to a legislature expiring in January 2027. Either a defeated or freshly elected Congress rushes it through, or it dies and starts over. Since the rapporteur intends to rewrite what the Senate sent, it would have to return to the senators before signature either way.

May 2027. The second session of the UN Global Dialogue on AI Governance convenes in New York on the 3rd and 4th. Whoever wins in October will be about four months into the job, and will need to send someone with a written position. No campaign currently has one.

December 2027. Under Regulation (EU) 2026/1744, published on 24 July and in force since 27 July, the European Union moved its high-risk obligations from 2 August 2026 to 2 December 2027. The transparency rules — telling users they are talking to a machine — took effect on the original date. The model Brazil is copying moved its own clock. That gives cover to anyone who prefers to wait, and it also opens a window in which a Brazilian rule could land before the European one, if anyone wanted it to.

There is a version of the next four years in which Brazil takes a role neither Washington nor Beijing can take: convening the majority. And there is the version the platforms describe without naming — host the compute, import the rules, export the power and the water, and learn afterwards that Brazilian law is a translation of a decision made in another time zone.

I am not saying which is better. I am saying the choice is being made now, by omission, and nobody asked for a vote on it.

👉 Read Also Fuga do laboratório.