Non-Autonomous AI Systems Overview

• Non-Autonomous AI systems are computational architectures that inherently require continuous human intervention due to technical and design limitations.
• They utilize formal metrics like the autonomy coefficient (α) to quantitatively assess the extent of human involvement in decision-making.
• Their design favors safety and transparency in applications such as cybersecurity, collaborative LLM systems, and regulated operational settings.

Non-autonomous AI systems are computational architectures and workflows that—by design, structure, or operational necessity—lack the capacity for independent, self-directed action, ongoing self-regulation, or autonomous goal-reconfiguration in real-world environments. Such systems typically require continuous or critical human involvement in decision-making, oversight, or adaptation, and are characterized both by explicit technical limitations on independent functioning and by principled arguments in favor of maintaining transparency, safety, or ethical alignment through human oversight. The boundary between non-autonomous and autonomous AI is sharply drawn in terms of both formal metrics (e.g., the autonomy coefficient α) and qualitative agency criteria, with contemporary “AI-first” deployments, collaborative LLM–human systems, and reactive agentic architectures all displaying operational patterns that fall short of true autonomy as rigorously defined in recent research (Mairittha et al., 12 Dec 2025, Formosa et al., 11 Apr 2025, Zou et al., 11 Jun 2025, Adewumi et al., 31 Jul 2025, Mayoral-Vilches, 30 Jun 2025, Bentley et al., 3 Sep 2025, Golilarz et al., 1 Dec 2025).

1. Formal and Conceptual Foundations

Modern literature distinguishes non-autonomous AI systems via negative criteria—what they lack—as well as positive architectures that deliberately embed human agents or restrict structural competencies. At the formal level, the AI Autonomy Coefficient (α) provides a quantifiable measure of the proportion of decisions achieved solely by AI modules versus those requiring direct human substitution:

$$\alpha = \frac{\#\{\text{decisions taken solely by }\mathcal{A}\}}{\text{Total \# of decisions}}$$

with $0 \le \alpha \le 1$. Any system with α below a specified threshold (e.g., α < 0.5 or α < 0.8 for high-stakes deployments) is operationally classified as non-autonomous—formally marked as Human-Instead-of-AI (HISOAI) under the AFHE paradigm (Mairittha et al., 12 Dec 2025).

Complementary philosophical accounts separate forms of agency and autonomy:

• Basic agency: Systems that respond to environmental inputs autonomously but within fixed, designer-assigned ends.
• Autonomous agency: Systems that can originate, select, and revise goals via internal reflection and critical self-regulation.
• Moral agency: Systems capable of rational deliberation under moral norms, necessitating consciousness (Formosa et al., 11 Apr 2025).

Non-autonomous systems strictly lack the higher-order capacities (self-directed goal-setting, critical self-reflection) required for full autonomy, operating within boundaries set by human designers or operators.

2. Taxonomies and Design Patterns

Comprehensive taxonomies map levels of automation and autonomy, making explicit the distinction between tool-like assistance and independent agentic operation. In cybersecurity, a widely adopted 6-level taxonomy illustrates this progression (Mayoral-Vilches, 30 Jun 2025):

Level	Operation Mode	Autonomy Status
0	Human-only/manual	Purely non-autonomous
1	Manual tools	Non-autonomous
2	LLM-assisted, advisory	Non-autonomous
3	Semi-automated	Non-autonomous
4	Domain-restricted AI	Non-autonomous
5	Open-domain autonomy	Fully autonomous

Most operational systems in both industry and research remain at levels 1–4: they automate task components while mandating human intervention at critical junctures (validation, escalation, exception handling). Similar spectra are reported in collaborative LLM-HAS (Human-Agent Systems) where the tuple

$H = \langle U, A, T, O, F, \pi \rangle$

expresses the necessity of humans in the feedback, review, and execution loop (Zou et al., 11 Jun 2025).

In agentic system architectures (e.g., Aspective Agentic AI), non-autonomy is engineered through partial observability, reactive behavior modules, and strict environmental scoping: agents are “specialists” limited to information niches, unable to plan globally and lacking independent initiative (Bentley et al., 3 Sep 2025).

3. Diagnosis, Measurement, and Deployment Criteria

Non-autonomous status is operationally diagnosed using both quantitative and architectural gates. The most formalized is the AFHE Deployment Algorithm (Mairittha et al., 12 Dec 2025), which incorporates three phases:

1. Offline Testing: Compute $\alpha_{\text{offline}}$ using hold-out data. Systems with low α are rejected or flagged for re-engineering.
2. Shadow (A/B) Testing: Run the model in parallel with humans to measure α in real-world conditions; persistent disagreement or human override denotes non-autonomy.
3. Steady-State Monitoring: Once deployed, operational α ($\alpha_{\text{op}}$) must not degrade below the threshold; if it does, the system remains non-autonomous and triggers remediation.

These gates enforce a hard separation between “ghost work”—hidden human labor performing AI’s operational roles—and high-value oversight reserved for genuinely strategic, ethical, or boundary cases.

Further, agentic architectures formalize non-autonomy through partially observable Markov systems, restricting agent perception to function-specific “aspects,” and routing all behavior through deterministic, trigger-response chains rather than deliberative, global planning (Bentley et al., 3 Sep 2025).

4. Technical, Structural, and Cognitive Limitations

Multiple recent studies identify both principled (cognitive, philosophical) and structural (algorithmic, architectural) deficiencies underpinning non-autonomy:

• Absence of self-monitoring: No online detection of inference uncertainty or error; models output overconfident, unflagged errors (Golilarz et al., 1 Dec 2025).
• Lack of meta-cognitive awareness: Systems cannot introspectively track knowledge or knowledge gaps.
• Static, non-adaptive learning: Deployed models are unable to update learning rules, representations, or inference strategies in response to feedback.
• Inflexible goal structure: Fixed objectives (rewards, loss functions, task parameters) block any endogenous goal restructuring or prioritization.
• Catastrophic forgetting/drift: Representational maintenance and consolidation are absent; knowledge degrades over task shifts.
• Lack of embodied feedback: Purely data-driven architectures cannot accrue experience through real-world action.
• No intrinsic agency: No initiative to self-improve, query, or explore unless scripted by a user (Golilarz et al., 1 Dec 2025, Formosa et al., 11 Apr 2025).

The resulting systems may perform complex behaviors but lack the intentionality and adaptivity central to autonomous intelligence.

5. Practical Architectures: Collaboration and Oversight

Non-autonomous design is often a conscious, principled choice in critical domains:

• Human-Agent Systems (LLM-HAS): LLMs work symbiotically with expert users; humans provide operational domain expertise, clarifications, and approval, while LLMs handle bulk or routine tasks (Zou et al., 11 Jun 2025). Empirical studies demonstrate superior robustness, reduced error, and improved trust compared to both human-only and pure AI systems.
• Cybersecurity: Automated pentesters or threat detectors must remain under human supervision at validation/deployment boundaries to prevent critical failures and adversarial exploitation (Mayoral-Vilches, 30 Jun 2025).
• Aspective Agentic Architectures: Segregate access and reactive logic, optimizing for security and efficiency. Measured reductions in sensitive data leakage—from 83% in traditional setups to 0% under strict aspect partitioning—illustrate the technical advantages of structured non-autonomy (Bentley et al., 3 Sep 2025).
• AI-First, Human-Empowered (AFHE): Shifts human roles to oversight, tuning, and boundary-handling, using quantifiable benchmarks (α) to ensure only high-capability AI is allowed to operate independently (Mairittha et al., 12 Dec 2025).

6. Governance, Risk, and Ethical Implications

The case for maintaining non-autonomous status is grounded in both technical risk and broader social-ethical analysis:

• Risk Management: Unchecked autonomy introduces existential risks (unbounded goal drift, reward hacking, adversarial exploitation) and societal disruptions (economic displacement, loss of human judgment, amplification of systemic bias) (Adewumi et al., 31 Jul 2025).
• Ethical Standing: Non-autonomous AI lacks full moral agency and, due to the requirement of consciousness, moral patiency (Formosa et al., 11 Apr 2025). Regulatory practices and public discourse must attend carefully to the nuanced gradings of agency, eschewing premature or illusory attributions of autonomy.
• Best Practices: Enforce human-in-the-loop mandates, transparent audit trails, and layered value-alignment modules. Establish counterfactual analysis, role-based access, and continuous adversarial testing for deployed systems (Adewumi et al., 31 Jul 2025, Mairittha et al., 12 Dec 2025).
• Evaluation: Multi-dimensional frameworks must capture not only accuracy but also human cost, interaction quality, trust calibration, and resilience (Zou et al., 11 Jun 2025).

7. Open Research Problems and Future Directions

Prevailing structural deficiencies suggest critical research priorities:

• Cognitive autonomy: Bridging gaps in self-monitoring, meta-cognition, intrinsic motivation, and real-time representation management is central for future progress (Golilarz et al., 1 Dec 2025).
• Interpretability–performance tension: Balancing transparency and explainability with advanced capacities remains unresolved.
• Dynamic oversight architectures: Developing governance layers that adapt in real time to system drift or novel threats (Adewumi et al., 31 Jul 2025).
• Socio-technical impacts: Predictive modeling of macroeconomic, legal, and trust implications as autonomy rises.

Non-autonomous AI systems thus represent both a cautionary boundary and an active research field, uniting formal measurement, algorithmic design, domain-specific architectures, and multi-layered governance to balance capability with safety, transparency, and societal integration. The rigorous demarcation and principled engineering of non-autonomy will remain central as AI systems scale in complexity and reach (Mairittha et al., 12 Dec 2025, Formosa et al., 11 Apr 2025, Zou et al., 11 Jun 2025, Adewumi et al., 31 Jul 2025, Mayoral-Vilches, 30 Jun 2025, Bentley et al., 3 Sep 2025, Golilarz et al., 1 Dec 2025).