- Статьи
- Internet and technology
- Bit of bad luck: ChatGPT and Grok pointed out Trump's mistakes in planning the war in Iran
Bit of bad luck: ChatGPT and Grok pointed out Trump's mistakes in planning the war in Iran
ChatGPT and Grok assessed the US and Israeli operation against Iran, which, according to Western media, was planned by the US military using another neural network, Claude. The key mistakes, according to artificial intelligence, were made by people in uniform who relied on Claude, but did not listen to his advice. At the same time, the products of OpenAI and xAI companies highly appreciated the skills of their "colleague" from Anthropic. Izvestia figured out what went wrong when planning attacks on Iran and who should ultimately make the final decision on the strikes.
Who developed the strategy for the special operation in Iran
Izvestia asked OpenAI's ChatGPT and xAI's Grok to evaluate the actions of Claude, a neural network that, according to The Wall Street Journal, was used by the US military when developing a strategy for action in Iran.
According to both language models, the AI from Anthropic worked flawlessly, but it was the US administration's excessive reliance on neural networks that caused five system errors that quickly turned the "lightning operation" into a protracted and expensive military campaign.
ChatGPT cited the choice of an optimistic scenario, overestimation of AI accuracy, excessive reliance on target analysis, disregard for adaptive enemy actions, and the illusion of control as key mistakes when using Claude in Operation Epic Fury. When asked to clarify the final point, he noted that the US military, led by Trump, believed that AI was fully in control of the situation, instead of using it as an auxiliary tool.
According to Grok, the first mistake of the United States was blind faith in numbers. Claude should have given out that the probability of a "victory in one day" scenario is only 20-25%, and a protracted air war is 30-35%. But Trump and his team chose the most optimistic option.
The second mistake is completely ignoring Claude's warning about the "rallying around the flag effect." The neural network was supposed to show in thousands of simulations that after Khamenei's death, a powerful national upsurge would occur in the first two weeks (65-80% probability). Iranians will take to the streets, but not to protest against the regime. This is exactly what happened: Tehran was filled with pro-regime rallies, and anti-American sentiments only strengthened the current government.
The third mistake was purely political in nature. The day before the strikes, Trump declared Anthropic a "supply chain risk" and banned the use of Claude at the federal level. But a day later, the military continued to work with this model in secret networks. An absurd situation developed: some analysts counted the data using Claude, some hastily retrained on ChatGPT. There was an inconsistency, a waste of time, and internal strife at the Pentagon.
The conflict between Anthropic and the Pentagon was the first public clash of corporate AI ethics with military interests. Since 2024, the Claude model has been used in the closed networks of the US Department of War, but the company forbade its use for mass surveillance of citizens and the creation of fully autonomous weapons. In February 2026, the Pentagon demanded that these restrictions be lifted, insisting on prioritizing national security tasks, but the head of Anthropic, Dario Amodei, refused. As a result, the company was declared a "supply chain risk," banned from cooperating with it, and given six months to remove Claude from government systems. The vacant niche was occupied by OpenAI and xAI, which offered more flexible conditions for interaction with the military.
The fourth mistake was choosing a scenario without taking into account the "tail risks" that Claude should have highlighted separately. According to Grok, the AI accurately warned: the probability of activation of the "axis of resistance" (Hezbollah, Houthis, Iraqi militias) is 15-20%, and attempts to block the Strait of Hormuz are about 40%.
And finally, the fifth and most fundamental mistake is the complete lack of strict human control over the recommendations of the neural network. Claude did a perfect job with tactics: prioritizing targets, simulating strikes, and analyzing intelligence. But strategic questions— when to stop, how to respond to retaliatory strikes, and whether a plan B is needed — were left to politicians and emotions. And it was precisely the lack of mandatory human control that led to the fact that the tactical success of the first day turned into a strategic problem, the AI concluded.
Ethical issue and consequences
Claude could have calculated the consequences of all these actions in advance. Such systems are capable of simulating thousands of scenarios, including the effect of rallying society around power in the face of an external threat. This is a well-known historical mechanism that has already manifested itself in various conflicts, so the growing support for the regime and anti-American demonstrations in Tehran could hardly come as a surprise, IT expert Sergey Pomortsev explained to Izvestia.
According to him, such models usually take into account broader consequences, such as increased proxy conflicts. In particular, we may be talking about the involvement of forces allied to Iran in the region, including the Hezbollah movement, which increases the risk of a protracted conflict. By processing large amounts of data, AI is able to assess the likelihood of such scenarios and show the risks of possible escalation.
— Artificial intelligence does not make decisions on its own — it only offers options. The final choice is always up to people, and this is where mistakes most often occur. Political motives, haste, or overconfidence can lead to choosing the most optimistic of all scenarios," he added.
Trump, the expert believes, wins in short-term effectiveness (as seen in the example of the operation in Iran), but runs a high risk of long-term problems.: This undermines trust, lawsuits, and the outflow of users from American AI technologies around the world.
"A more cautious and balanced approach, as recommended by experts, would be to combine AI with strict human control and reliable protective mechanisms that prevent the system from going beyond ethical and safe limits," explained Daniil Arzhakov, Chief Specialist of the Digital Development and Digital Project Implementation Service of the Federal State Social Security Fund.
The Pentagon will have to solve the largest ethical problem related to the use of AI in military operations, the experts surveyed concluded. OpenAI and Sam Altman have already signed documents on cooperation with the military department.
Переведено сервисом «Яндекс Переводчик»
