In 1991, the film "Terminator 2: Judgment Day" showed how humanity creates Skynet - an artificial intelligence (AI) that controls strategic defense. The machine becomes self-aware and unleashes a nuclear war. That image still influences debates about AI: supposedly large language models (LLMs - programs that generate text by responding to prompts) will become a superintelligence and begin pursuing their own goals.
But such a scenario requires many unconfirmed steps. A far more real threat is already taking shape - and it needs neither a superintelligence nor malicious intent.
In August 2026, Ukraine's Ministry of Defense reported that its forces use more than 70 AI and computer vision systems to carry out strikes. The algorithms recognize equipment and people and acquire targets on their own, so that a drone completes an attack even if contact with the operator is lost. All that remains for a human is to approve what the machine has already chosen.
This is where the weak link arises - trust. According to a CNN investigation, in spring 2026, during the war with Iran, a US special operations analyst used an AI chatbot to assess intelligence about a Chinese vessel in the Middle East. The chatbot, relying on open and classified materials, mistakenly identified the cargo as components of a nuclear program. The analyst again used AI to format this into a standard intelligence report. The military was preparing to board the vessel, aircraft were already in the air, but an official noticed the error. A source said the false report "almost started a war."
The AI was neither intelligent nor hostile - it simply made a mistake, but people trusted it and began to act. This is a problem of trust: LLMs produce fluent and confident text, and when formatted as an official document, it gains an authority it does not deserve. In one experiment, non-specialists could not distinguish a convincing but incorrect chatbot answer on physics from genuine expertise.
Policymakers often reassure: the final decision is made by a human. But in the case of the vessel, people were in the chain constantly - they simply became another stage in passing on the error. This matches decades of research on automation bias - the tendency to rely excessively on automation. In 2003, US Patriot batteries shot down two allied aircraft after operators accepted an erroneous classification.
A human in the chain works only if they are capable, equipped and ready to challenge the machine. This requires verification: AI results must be clearly labeled and, before action, checked against independent sources. War leaves no right to an unverified error.
Short-term trust risks do not cancel long-term questions about advanced AI. But public anxiety is focused on what machines may become, not on what they already do. The danger is not that machines will get smarter, but that we will entrust responsibility to systems that are far less reliable than they seem.