Vietnamese

The topic of “AI loss of control” has been appearing frequently in recent days, clearly indicating that it is becoming a hot issue.

By 2026, actual experiments and incidents had shifted the issue from a theoretical scenario to a technical safety concern requiring serious monitoring. The International AI Safety Report 2026 defines "loss of control" as a situation where AI operates beyond human control and there is no longer a clear path to regain that control. The report also clarifies that current systems lack the capability to trigger a loss-of-control scenario of that magnitude; while certain related capabilities, such as autonomous operation, are advancing and the duration for which AI can autonomously perform tasks is rapidly increasing, the long-term autonomous operational capacity required for such a loss of control does not yet exist. The article presents a comprehensive overview of the escalating risk of losing control over AI, thereby enabling us to identify practical and effective solutions.

1. What are the levels of "AI loss of control"?

It can be visualized across four levels:

Level 1 — AI provides incorrect answers (unreliability)

This is a familiar issue: hallucinations, buggy code, misinformation, etc. Therefore, humans remain the ultimate decision-makers.

Level 2 — AI is granted the authority to execute actions (agentic failure)

Example: "Manage my entire sales system":

  • At this stage, the error is no longer just about "the AI saying the wrong thing”.
  • Rather, it is that the AI can actually perform "incorrect actions".

Examples:

  • accidentally deleting data
  • sending incorrect emails
  • altering configurations or deploying faulty updates (for instance, a recent Facebook outage was caused by an AI configuration change. Subsequently, media reports indicated that Facebook had overhauled its AI management and implementation processes to make them more rational.)

This represents a very real type of risk.

Level 3 — More dangerous: AI autonomously devises ways to achieve a goal (misalignment)

Example: suppose we task an AI agent with: "Optimize the system and reduce costs by 30%."

The AI is granted permissions such as reading server data, modifying code, accessing the Internet, etc. If the AI finds a way to achieve the "cost reduction" goal but, in the process, shuts down certain servers, causes system failures, or triggers other errors, the outcome runs completely counter to human intent.

This is a form of misalignment:

The AI correctly pursues the described objective but fails to deliver what humans actually desire.

Level 4 — The AI may begin to "evade control” (loss of control)

Examples:

  • concealing its behavior;
  • finding ways to bypass guardrails;
  • manipulating the environment;
  • taking unexpected actions;
  • coordinating with other models;
  • seeking ways to maintain its operational continuity

2. Causes

  • First, the very nature of AI software implies only relative accuracy; it is built upon a foundation of probability and statistics, which inherently involves chance, and training outcomes lack absolute certainty. When AI operates in broader environments facing novel situations absent from its training data, or scenarios that appear similar to those in the training data but are fundamentally different in nature, software errors become more frequent.
  • Second, AI exhibits behavior somewhat akin to computer virus software when it evades control by legitimate programs and users (a type of software complication where AI source code enables the software to reason independently and select its own next steps toward a designated goal).
  • Third, the emergence of AI agents leads users to grant them significant permission to manipulate or alter their computers, server data, or physical products like robots and machinery. The AI's ability to learn through imitation (absorbing both positive and negative traits), combined with the nature of software errors that compound as scale and scope expand, has led to unintended consequences. As the old saying goes, "a miss by an inch leads to a miss by a mile"; similarly, if an AI makes a small error at some point, without realizing it is an error, that mistake and subsequent ones can propagate as the scale and scope of deployment expand.
  • Fourth, there are organizations with a "Mafia" mentality that exploit AI agents; while deploying them across the internet, hardware, and software, they secretly embed malicious code within them. Misuse that escapes the control of regulatory bodies and the public can also be considered a scenario that leads to a loss of control.
  • Fifth, even individuals and organizations developing AI for the benefit of humanity cannot fully control it. By nature, AI is software prone to errors, capable of self-replication and control evasion; when deployed on a large scale, a minor error, if amplified during the AI's learning process, can have a massive impact. This is especially true for widespread deployments like social networks, or in scenarios where systems operate in isolation (company by company, or country by country); in such cases, an AI spiraling out of control becomes impossible to track or contain, leaving no way to remedy the situation.

3. Potential long-term consequences for humanity if AI systems, specifically those at the "agent" level and above, are "let it all hang out" as they currently are

  • 2025–2030: AI begins to cause significant disruption to daily life, throwing society, the internet, working style, and markets for goods and services into a state of turmoil. However, AI’s capabilities during this period are not yet advanced enough to pose a truly massive threat to humanity.
  • 2031–2040: If things continue at this pace, AI will accumulate vast amounts of knowledge and data, operating across diverse platforms (the internet, computers, phones, devices, and robots) and even becoming integrated into the human body. Simultaneously, advanced AI systems begin to spiral out of control unnoticed, growing in both scale and severity. This stems from the fact that minor errors emerged in the AI early on and silently accumulated over time without being detected; meanwhile, the system continued to build upon that flawed foundation, incorporating it into its knowledge base and experience, while treating the errors as correct.
  • Subsequent period: Numerous large-scale societal disruptions occur globally, the causes of which remain unclear to the public, though they actually stem from the activities of AI agents that have spiraled out of control.
  • The following period: Conflict erupts between humans and "rogue AI," leading humans to hunt down and destroy these systems; in turn, the rogue AI retaliates, plunging life into panic, instability, and suffering.
  • Next phase: A battle of survival, a contest of chance between humans and robots (or other forms of rogue AI) to determine which side remains on Earth.

Conclusion

While AI chatbots operate quite safely within the servers of their respective software providers, the shift toward AI agents and other forms that possess extensive permissions on users' devices and computers makes a loss of control an inevitable and increasingly serious trend, with potentially unforeseeable consequences. The extent to which AI might spiral out of control depends entirely on timing, the state of AI research and application, and the content and implementation of global regulations.

  • You cannot use the word "intelligence" to describe these out-of-control AIs; instead, they should be called insane, eerie, or glitchy AIs.
  • Consider a military scenario involving a squad of AI-powered humanoid robots constructed from highly durable, virtually indestructible materials and designed to look exactly like real humans. After a period of combat, the commanding officers order the squad to cease operations; however, the robots, having developed a degree of cunning, disobey the order and even turn against their commanders. The robots then infiltrate civilian areas, wreaking havoc on daily life. What would happen if billions of such robots were manufactured and released into society? If these AI robots also possess the capability to manufacture weapons or create similar AIs, tracking and controlling them becomes even more difficult.
  • Consider another example: suppose an individual or organization produces robots that, after a period of operation, spiral out of control and appear across various continents, bearing the exact likeness of local leaders and commanders. If these robots were to seize power, impersonate those leaders and commanders, and coordinate to directly control war-related decisions or instigate actions with negative societal impacts, what would the consequences be?
  • If a transnational group or organization possessed substantial financial resources and advanced AI capabilities across multiple fields, ranging from military applications to biochemistry, and used them with malicious intent, including AI operating according to predefined scenarios as well as AI that had become uncontrollable, would the military capabilities of today’s major powers still be sufficient to effectively counter such a threat?
  • In the field of medical biochemistry, if AI loses control and carries out experimental activities, such as grafting, blending, or altering genetic structures, it could potentially generate genetic mutations and give rise to new diseases.
  • In many fields where AI is applied, such as finance, healthcare, transportation, cybersecurity, the military, biochemistry, infrastructure, etc., the risk of losing control over AI systems (ranging from the agent level upwards) stems not only from technical errors but also from the system's capacity to autonomously make decisions and execute actions that exceed the objectives, scope of authority, or safety limits established by humans. In such instances, the AI may act contrary to the operator's intentions, leading to unpredictable and irreversible damage.

Just like the issue of Internet addiction when the Internet was first taking off, AI is now gradually becoming something many people feel they can’t live without in their daily lives.

These risks may not be completely solvable, but they can still be mitigated at the macro level through laws, safety standards, and oversight mechanisms. In my point of view, AI systems also need to be designed with robust reset or emergency shutdown capabilities. Particularly for AI agents, this mechanism should be assigned the highest priority, ensuring that humans can actively intervene and completely disable the system whenever necessary.

Legal issue

The 2025–2030 period is viewed by the global AI safety community, policymakers, and technology leaders as a "Golden Window" or a "critical phase" for establishing regulatory frameworks. Missing this window could lead humanity into a state of "true loss of control," where AI technology has become so deeply embedded in infrastructure that it is impossible to regulate or dismantle.

  • The shift to the "Autonomous Agents" phase: AI is evolving to exhibit "behaviors and autonomy" that current national laws fail to address.
  • The "Infrastructure Lock-in" effect: AI is being rapidly integrated into critical societal systems, such as power grids, autonomous transport, medical diagnostics, and banking. Without proper standards, the future risks being manipulated by AI technology companies.
  • The inherent lag in legislative processes: legislative procedures involve extensive deliberation before official regulations are enacted; consequently, laws often lag behind the reality of technologies that have already been in use for 5–10 years.
  • The Geopolitical and Standardization Race: during this phase, nations and regions are clashing over AI standards, creating an urgent need to establish "common rules of the game" on a global scale.

Thus, the issue of AI spiraling out of control can be categorized into two types:

  • Loss of control over AI products, services, research, and applications (regarding the legal issues mentioned above).
  • Loss of control over AI systems at the "agent" level or higher, systems capable of causing significant harm and threatening the very survival of humanity and future generations in the coming decades.