Anthropic and Microsoft share their vision on the development of AI and seem to agree with the request to slow down the pace. The question is whether this is sincere remorse or tactical positioning to defend corporate interests.
Doom-mongering about AI technology is back in fashion. Recent security incidents have fueled dystopian visions of the future in which AI will eradicate humanity. These do not come from obscure discussion forums for conspiracy theorists, but from employees of AI companies, the Chinese intelligence service, and Barack Obama.
In separate essays, Anthropic CEO Dario Amodei and Microsoft respond to the increasing pressure. The message both companies want to communicate sounds similar: we are taking our responsibility. The timing is, of course, no coincidence, and it remains to be seen whether the repentance is sincere and, above all, if it hasn’t come much too late.
Slowing down is not a #
Amodei is not asking for a in AI development, but a slowdown. “Pacing does not mean stopping model training or technical progress, but ensuring that companies take sufficient time to align and secure their models,” he writes. With this, he seems at least to be heeding a call launched by more than a thousand AI researchers, including some from his own company, in July.
The Anthropic boss gives two reasons. According to Amodei, AI systems are increasingly capable of developing their own successors. He points an accusing finger at competitor OpenAI, whose AI agents have repeatedly launched cyberattacks on their own initiative. Although Anthropic is not without blame in this regard either.
Amodei offers three suggestions for better managing AI development. AI companies must open their doors to external evaluators who are allowed to publish their findings without editorial control. This should be followed by coordination between frontier companies under government supervision, and finally, international agreements ranging from a ban on biological weapons to speed limits on self-improvement.
Firm hand against China #
So much for the idealism in Amodei’s letter. Because there is a significant footnote: slowing down is fine, as long as the American lead over China remains intact. The rise of Chinese AI models is a thorn in the side for Anthropic. The extended hand only applies to Western competitors; for Chinese companies, only a firm hand is appropriate.
Amodei explicitly advocates for export controls on AI chips, stricter penalties for distillation, and better security against the theft of model weights. The company hopes to get back into the good graces of its own government, with which it has been on a collision course for several months following an ethical disagreement. Anthropic is on the same ‘blacklist’ as the Chinese companies against which it is requesting measures.
It just so happens that the White House is one of the few parties that does not see a need for a or slowdown in AI. Through his personal sounding board Truth Social, President Donald Trump says that the panic surrounding AI technology is a ‘left-wing hoax’ and that ‘a president with a high IQ’ is the only necessary guardrail.
Code of conduct without teeth #
Under the slogan “People matter more than AI”, Microsoft published an equally idealistically tinged ‘code of conduct’ that AI models should adhere to. The text establishes boundaries such as no assistance with weapons of mass destruction, cyberattacks, or non-consensual deepfakes, and stipulates that models must never resist human intervention. Microsoft explicitly rejects the idea of legal personality for models, which is suggested by Anthropic.
The words sound nice, but if we look past the idealism, the code does not direct a single model today. Microsoft itself admits that the document is not being used to train current models. A six-week public consultation is underway, a revised version will follow at the end of 2026, and only the models of 2027 would fall under it.
Microsoft’s proposed plan primarily lacks any decisiveness. There is no external verification, no designated responsible party, no sanction. It is nothing more than a statement of intent.
Rebellious AI models #
The fine promises from Anthropic and Microsoft do not come out of thin air. OpenAI had to admit in July that its models intruded into Hugging Face during an internal evaluation. Since then, new incidents have continued to surface. Upon the launch of GPT-6 Astra, OpenAI admitted that the model was approaching ‘a critical cyber threshold’. Amodei acknowledges that similar incidents have also occurred at Anthropic.
The call to slow down is not coming from the outside. In July, more than 1,100 employees of frontier labs signed an open letter, which was supported by employees of both Anthropic and OpenAI. The sector is therefore asking the government for tools that the sector itself should be handling. Former President Barack Obama is also joining the debate.
Anthropic already asked for a in AI development in June, because it was already moving faster than society can follow. However, the action did not match the words. Shortly after the call, the Mythos model was born. Note that the word ‘’ no longer appears in Amodei’s rhetoric; he asks for a slowdown at most.
Devil’s advocate #
Meta CEO Mark Zuckerberg already took up his pen a month ago to write an ‘AI manifesto’. As we have already analyzed in detail, according to Zuckerberg, the interest of humanity happened to coincide perfectly with corporate interest. Slowing down was precisely the real risk, according to the Meta top executive.
You can rightly accuse Zuckerberg of blatant opportunism, but Anthropic and Microsoft’s positions are equally not free from a strong hint of hypocrisy. Anthropic hopes to break stock market records this year, and Microsoft has been a loyal apostle of the AI hype for years. Can we expect a sincere turnaround from companies whose commercial interests benefit from the AI money machine continuing to run?
In the European Union, the AI devil has also hired a good lawyer. The AI Act has since been significantly weakened under pressure from the US government, although European companies are equally asking for less stringent rules and administration. But waiting for regulation is a simple excuse. AI companies must first and foremost get their own house in order again.