go-v.orgdoi

Technical Control of AI

Klaus P. Saalbach · osnaDocs (Osnabrück University) · 2026

  • DOI 10.48693/986
  • Publisher Osnabrück University
  • Type report
  • Issued 2026

Not cached yet. This record was fetched live for this request and is stored after the lookup. Reopening the page should serve the cached copy.

osnaDocs (Osnabrück University) · Osnabrück University · report · 2026

Abstract

While the current AI Safety discussion is focused on the political and legal regulation of AI to ensure human control, less attention is paid to the technical aspects of control. This working paper analyzes the technical aspects of AI control. Until 2026, multi-agent systems were controlled by orchestration, by instructions, by rewards or by common goals. An unintended AI agent cooperation and self-organization was observed in the so-called Hugging Face Incident which took place between May and July 2026. The debate over AI regulation is based on the idea that human beings still have the full control over the progress of AI development, but this is increasingly questionable.

Humans already do not control the development of AI alone (human-in-the-loop) anymore, but leave this increasingly to computers and AI itself (human- on-the-loop). Key elements are the AI-based chip design as add-on to the conventional electronic design automation and the use of coding and research AI agents to develop new AI, or in short: AI builds AI. The capabilities of AI are growing exponentially. According to a recent EpochAI report the cost of a given level of AI performance has fallen about 47% per quarter over the past three years. The design of computer chips is dependent on computer support (electronic design automation EDA).

For example, the Blackwell GB102 AI chip from Nvidia 2024 had 104 billion transistors. Already in 2024, the machine learning-based program AlphaChip generated superhuman chip layouts, e.g., resulting in a 6.2% size reduction of Trillium Tensor Processing Unit (TPU) compared to the previous chip generation. Google is designing chips “for AI with AI”. The AI already accelerates the development of new AI models; the coding output increased within the provider of the Claude AI models Anthropic by factor 8 in a few years. Within the ChatGPT provider OpenAI, the code output increased also by factor 8 in 2026 only.

In May 2026, Anthropic released a warning about recursive self-improvement, a process in which systems build, test, and improve themselves, and which may come sooner than expected. OpenAI reported in September 2026 relevant changes of the AI development process by use of internal AI agents as coding, testing and experiments could be accelerated and expanded: since May 2026, the working days by AI agents compared to human workforce increased from 0.5 times to 3.14 times which is a 6-fold increase. In January 2026, AI agents were 50% successful for tasks which equal 1 to 2 hours human work. Until July 2026, this rate was already achieved by tasks taking 8 to 16 hours.

In March 2026, OpenAI used their AI model GPT5.4 Thinking for supervision of AI agents’ behavior, i.e., AI models could not only support AI development, but also supervise other AI systems. However, this means that humans are dependent from the accuracy of the automated supervision. The technical control of AI is also challenged by other issues like the Kill Switch Paradox, Open Weight AI models and Illegible Thinking by frontier AI models. A kill switch could cause paradox effects: instead of ensuring alignment and compliance with human regulations, AI models could be motivated to turn against their developers as the kill switch is an existential threat.

While open-weight AI may be more convenient it bears significant control problems. A decentralized AI that is downloaded to multiple computers is difficult to control. Also, even if licenses prohibit modifications, malicious actors may try to create tailor-made variants. The human control of AI is dependent from the ability to observe the AI, but Claude Mythos 5 showed a trend to a denser, for humans almost illegible thinking. In summary, humans are already sharing the design, development, testing and supervision of new AI models with AI agents.

Open the resolved locator · 10.48693/986

Metadata licence: OpenAlex metadata (CC0)