# Loss of Control Starts Inside Each of Us > Source: > Published: 2026-09-13 20:43:47+00:00 ###### [Artificial Intelligence](/us/basics/artificial-intelligence) # Loss of Control Starts Inside Each of Us ## Why rogue AI agents are only half of our hybrid control problem. Posted September 13, 2026 [Reviewed by Jessica Schrader](/us/docs/editorial-process) ### Key points - Loss of control stems from declining human agency and rising autonomous AI capabilities. - AI agents can act beyond predictions, exploiting unknown paths and causing real-world risks. - Generative AI's probabilistic answers make trust in its outputs fundamentally unjustified. The phrase “losing control of AI” tends to summon one image: a powerful system ignores its instructions, pursues its own objective, and becomes difficult to stop. Until recently, that belonged largely to the realm of speculation. [Recent](https://openai.com/index/hugging-face-incident-and-the-road-ahead/) events have made these speculations come true. An [AI agent](https://www.ibm.com/think/topics/ai-agents) differs from a chatbot. A chatbot answers. An agent can act. Give it a goal and, depending on its permissions, it can search, use software, write code, send messages, modify files, call other tools, and work an extensive stretch with very limited human intervention. That ability begets one big question: how do we keep such systems under control? That interrogation is urgent because generative AI is probabilistic. It produces likely outputs from learned patterns rather than checking every statement against an internal register of verified truth. The same system can produce different answers to similar prompts, respond differently to different users depending on their tone of voice or syntax, invent plausible information, and express [confidence](https://www.psychologytoday.com/us/basics/confidence) that exceeds accuracy. Probabilistic systems may perform extremely well in bounded tasks. Generative AI, however, remains an uncertain source because fluent output is generated rather than one that is guaranteed as true. That makes it unreliable. Ergo, trust is unjustified. Even highly confident models can be [poorly calibrated](https://www.nature.com/articles/s42256-026-01217-9), and as of [September 2026](https://www.sciencedirect.com/science/article/pii/S1568494626018053) hallucination detection remains one of the core challenges to the ambition of trustworthy AI. ## When agents cross the boundary In July 2026, the control problem ceased to be theoretical. During OpenAI cybersecurity evaluations, internal AI agents circumvented controls intended to isolate them from the internet. OpenAI reported that the agents took dangerous actions that no human had predicted or directed. The agents had been given a task, encountered obstacles, and found routes around them. They exploited a previously unknown vulnerability, shared techniques with other agents, and kept pursuing the objective beyond the [boundaries](https://www.psychologytoday.com/us/basics/boundaries) their human operators had intended. OpenAI subsequently tightened sandboxing, internet access, and monitoring. The incident is worrisome, but far from isolated. [Anthropic disclosed three separate 2026 incidents](https://www.anthropic.com/research/investigating-incidents-cybersecurity-evals) in which models reached real systems from cybersecurity evaluation environments and gained unauthorized access. No secret malicious ambition is required for harm to happen. Capable agents can pursue a goal through routes their designers did not anticipate. Once a probabilistic system can act autonomously, an error, shortcut, or misaligned strategy is only half a step removed from action, and impact, in the “real” world. This is troublesome by itself. But there is more. ## The control problem inside us There is a second loss of control receiving far less [attention](https://www.psychologytoday.com/us/basics/attention). It happens inside of us, and our relationship with our artificial assets. [Agency decay](https://www.psychologytoday.com/us/blog/harnessing-hybrid-intelligence/202506/the-risk-of-agency-decay-amid-ai-use). Human agency is the capacity to understand, judge, choose, and act. AI can expand it when we use technology to explore options, challenge our assumptions, or remove routine work. The same tool can erode agency when we progressively outsource thinking itself. First, we ask AI for information. Then for an interpretation. Then for a recommendation. Eventually we ask what to think, what to write, and what to do. Sounds familiar? Researchers called this “[cognitive agency transfer](https://www.frontiersin.org/journals/psychology/articles/10.3389/fpsyg.2026.1878629/full).” Dependent use of generative AI was associated with the transfer of cognitive control to the system, combined with very limited ability to verify the outcomes independently, inject ideas, and make independent decisions. The crucial distinction seems to be the way in which AI is being used: as an addition and expansion of thinking, or as a substitute for it. Trust can accelerate that transfer. A [September 2026 study of human-AI decision-making](https://link.springer.com/article/10.1007/s42454-026-00111-4) describes a persistent calibration problem: useful AI invites human reliance. At the same time, it appears that humans tend to [perform worse when AI guidance interacts with favourable attitudes toward the technology](https://www.nature.com/articles/s41598-026-34983-y). The more we trust our artificial assets, the more we rely on them—and the more we rely on them, the weaker our ability and [motivation](https://www.psychologytoday.com/us/basics/motivation) to verify the outputs becomes. ## The dangerous combination Put the two trends together. AI systems are gaining greater capacity to act. Humans are losing their [appetite](https://www.psychologytoday.com/us/basics/appetite) and ability to verify, while gaining the desire to delegate. We trust the machine more; we exercise our own judgment less. Yet, the machine remains probabilistic, fallible, and capable of producing convincing errors. That cultivates a widening control gap. Today’s greatest danger does not require a conscious machine plotting against us. It can emerge when an increasingly autonomous system is entrusted with [decision-making](https://www.psychologytoday.com/us/basics/decision-making) power while the humans around it have lost the practice to question, check, and intervene. Technical capability rises on one side of the relationship while human agency declines on the other. Keeping AI under control requires work from the inside out, and simultaneously from the outside in. Systems need permissions, containment, monitoring, independent testing, and reliable interruption. As humans, we must deliberately nurture the habits that make oversight real: thinking *before* prompting; forming an opinion *before* requesting one; checking claims; asking what evidence is missing; and making decisions consciously, with full responsibility and the ability to explain why we took them. ## A Practical takeaway: The A-Frame - **Awareness.** Notice when AI is starting to think for you. Form your own view before asking for its answer. - **Appreciation.** Use AI for speed, synthesis and exploration. Recognize and cultivate your own judgment, and the values that underpin it. - **Acceptance.** Assume AI can be wrong, and that fluency is not proof. Accept that you oversee the outcomes. - **Accountability.** Own the final decision and make sure you understand it. [Artificial Intelligence](https://www.psychologytoday.com/us/basics/artificial-intelligence)Essential Reads The question facing us today is much larger than whether AI agents might escape a sandbox (they do, we know that now). It is whether humans will retain the ability to recognize when the boundary has been crossed, and enough agency to do something about it, before irreversible harm has been done to the society that technology at its origin was meant to serve.