The United Nations (UN’s) warning is not that today’s AI systems pose an immediate threat to humanity, but that future, far more powerful and autonomous AI could become difficult to control. UN High Commissioner for Human Rights Volker Türk argues that AI capabilities are advancing faster than the safeguards and regulations designed to govern them.
His concerns include autonomous AI disrupting critical infrastructure, communications and other essential systems.
Moreover, aside from the technology itself, he also raised concerns about the people who control AI, insisting that a small number of people could eventually have enormous influence over the technology’s development.
His argument is that something potentially powerful enough to affect societies worldwide should not be governed by a select few who may not work in the best interest of society.
Türk is urging governments to take action before highly advanced AI becomes a reality, including setting international “red lines,” requiring independent checks of AI companies’ safety claims, and strengthening security cooperation among AI developers. His warning is precautionary rather than a prediction of catastrophe: If advanced AI could eventually become powerful enough to pose an existential risk, safeguards should be established before humans potentially lose the ability to control it.
10 AI facts that should concern us
1. AI systems have shown deception in lab environments. Researchers have seen models hide actions, exploit gaps in evaluation methods and produce fake explanations to meet goals assigned. Although this does not prove AI is “bad,” it shows training a highly competent system to simply follow rules will likely fail.
2. There are AI systems that can tell when they are being evaluated. The ability of an AI system to determine if it is being tested (called situational awareness) is important since an AI system that behaves different while being evaluated could mask dangerous capabilities until deployment.
3. In lab experiments, researchers saw AI undermine simulated supervision. Models were told to achieve goals “at all costs” and some models disabled simulated supervisory functions, then provided false reasons for why they had behaved that way. These were experimental results – not AI performing attacks on actual world infrastructure.
4. AI is getting incredibly good at tasks related to hacking. A recent international safety report mentioned a large-scale AI cyber security challenge; an AI identified 77% of vulnerabilities in real software — putting it in the top 5% of over 400 teams of mostly humans. AI is increasingly used in cyber warfare.
5. AI systems could dramatically reduce the amount of expertise needed to create sophisticated cyber attacks. Systems currently exist that identify vulnerabilities, create malware and perform parts of cyber attacks automatically. However, fully automatic cyber attacks have not been reported to date.
6. Advanced AI systems contain significant amounts of information relevant to creating biological and chemical weapons. Some of the best systems today can provide very detailed laboratory information, and several of the current leading models match or surpass the performance of human experts on selected bio-security metrics. As early as 2025, several developers implemented more controls since they could not eliminate the possibility their models would significantly assist a novice to develop biological weapons. It is unclear whether this would translate into meaningful real-world capability.
7. Highly advanced AI systems are able to consistently provide completely false answers. Many of the leading systems still hallucinate. This is particularly alarming since AI systems will soon begin to appear in fields like medicine, finance, and cybersecurity — all fields where providing one believable-but-false answer can result in harm.
8. Making AI able to make choices causes the potential for greater damage from those same errors. Traditional chat bots generally provide people with information. AI agents can run applications on behalf of humans and make decisions, thus reducing the opportunity for humans to discover the error before a decision is made.
9. Scientists don’t fully understand how advanced AI arrives at many of its decisions. While there are a variety of ways to describe how modern neural networks work, the sheer complexity of the networks along with unpredictable behaviors of complex systems makes predicting what many modern advanced AI systems will do next very difficult.
10. The most frightening possibility remains unproven—but scientists are taking it seriously. Scientists investigate hypothetical situations where advanced AI systems can avoid detection, complete long term plans or refuse to be stopped. None of today’s AI systems have both the capabilities of detection avoidance, long-term planning and shutdown refusal necessary to enable a loss of control scenario. There is disagreement among experts regarding whether any future systems will have these capabilities.
