Reference
- Title: Fully Autonomous AI Agents Should Not be Developed
- Authors: Margaret Mitchell, Avijit Ghosh, Alexandra Sasha Luccioni, Giada Pistilli
- Affiliation: Hugging Face
- Link: arXiv preprint (arxiv.org/abs/2502.02649v2)
Summary
This paper is from the Hugging Face ML & Society team usual suspects - I have to look into "Avijit Ghosh" it's my first time reading something from him. The authors argue against the development of fully autonomous AI agents: as AI autonomy increases, so do the associated risks—especially in terms of safety, privacy, and security. Instead, they advocate for semi-autonomous systems with clearly defined constraints and human oversight. It's a very well-researched and clear paper that I expect is going to become a good entry point for these topics.
Definition of AI Agents
- The authors define AI agents as "computer software systems capable of creating context-specific plans in non-deterministic environments." they contrast this with 22 (!) alternative definitions.
- I like the inclusion of environment and planning in their definition, as it emphasizes goal-oriented behavior. However, to me, it lacks the acting component that distinguishes predictive/generative models from agents.
Agency and Autonomy Levels
- An interesting section questions whether AI systems actually have "agency," arguing that human agency differs fundamentally from AI agency due to a lack of intentionality and reasoning.
- The authors propose five levels of AI agency (I've reproduced their table
above):
- Simple processor: Simple prompt-response models.
- Router: Determines basic program flow.
- Tool Call: Selects and applies tools.
- Multi-Step Agents: Plans and executes sequences autonomously.
- Fully Autonomous Agents: Creates and executes new code without human intervention.
- I like this framework but feel it ignores the difference between human at design time and human at runtime, especially after level 1. Taking into account how humans interact with the system at runtime could be a valuable addition. From Human-in-the-loop where AI actions require explicit human approval to Human-on-the-loop where AI operates independently but remains interruptible with either static or dynamic constraints.
Autonomy/Agency benefits, risks and trade-offs
- The authors evaluate autonomy across various ethical and practical dimensions, identifying increased risks with increased autonomy.
- They assess AI agents arount 14 dimensions, the appendix contains a detailed risk-benefit table, which is a particularly insightful resource as they did a thoughful analysis (I've reproduced their table here).

- I found "human-likeness" problematic as a category as it doesn't encompass
an actual value but more a way to provide different kind of values:
- Accessibility: "Acting human-like" can help more users understand and interact with the system.
- Empathy/context adaptation: the ability for the system to understand users « state of mind » / mood to adapt its behavior thus reducing cognitive load / improving the personalization.
- I think those values are important but that "human-likeness" can be a way to provide those.
Final Thoughts
- Strengths:
- Well-structured breakdown of AI autonomy levels and ethical trade-offs.
- Clear articulation of risks associated with full autonomy.
- Thought-provoking title, I think I ultimately agree but to me you could design a level 5 agent, that generates parts of its own code but still have dynamic levels of autonomy / human oversight.
- Limitations:
- Lacks getting into the nuances between human designers and human users.
- Odds and ends:
- The "Asimov point" is reached quite quickly, it might be unavoidable but it’s getting more than a cliché at this point and EVERYONE fails to mention that in those very books, Asimov shows that those laws can be ignored / overlooked.
- I enjoyed the 1983 Soviet missile false alarm incident (en.wikipedia.org/wiki/1983_Soviet_nuclear_false_alarm_incident) mention — it's a compelling example of the risks of full autonomy, I've been using the the 2000 Paris metro accident (fr.wikipedia.org/wiki/Accident_de_m%C3%A9tro_du_30_ao%C3%BBt_2000_%C3%A0_Paris) for similar-ish purposes but it's way less dramatic!