Towards A Framework for Levels of Anthropomorphic Deception in Robots and AI

1. Problem Statement

The paper addresses the growing ethical and psychological risks associated with anthropomorphic deception in autonomous systems (robots and AI). As social technologies increasingly mimic human appearance, behavior, and communication, they exploit the human cognitive tendency to attribute human characteristics (anthropomorphism) to non-human objects.

The core problem is the lack of a nuanced framework to distinguish between:

• Functionally necessary humanlike design.
• Socially appropriate design.
• Ethically permissible design.

Current design practices often fall into "Dark Patterns" (manipulative design exploiting cognitive biases), where systems implicitly or explicitly claim agency (autonomy/goal-orientation) and selfhood (emotions, identity, consciousness) they do not possess. This leads to user overtrust, emotional attachment to unempathetic agents, and potential manipulation. Existing regulations are often too high-level to address the specific "grey area" of how much deception is acceptable in specific use cases.

2. Methodology

The authors employ a conceptual framework development approach grounded in a review of existing literature from Human-Computer Interaction (HCI), Human-Robot Interaction (HRI), and psychology.

• Literature Synthesis: The paper reviews foundational concepts (anthropomorphism, humanlikeness, agency, selfhood, deception) and analyzes prior work on deceptive patterns, parasocial relationships, and AI ethics.
• Conceptual Differentiation: The authors distinguish between:
  • Humanlikeness: A property possessed by the machine (appearance, movement).
  • Anthropomorphism: A perception attributed by the human user.
  • Agency: Independent, goal-directed action.
  • Selfhood: Continuity of self, ownership of mental states (emotions, morality).
• Framework Construction: Based on these definitions, the authors propose a hierarchical, cumulative framework that classifies systems based on the degree to which they claim (rather than just possess) human traits.
• Case Study Application: The framework is applied to three specific HRI studies (persuasive robots, emotional manipulation in decision-making, and robot authority) to demonstrate its utility in categorizing real-world scenarios.

3. Key Contributions: The Four-Level Framework

The primary contribution is a four-level framework for classifying anthropomorphic deception. The levels are cumulative: a system claiming higher levels must inherently claim the lower levels.

• Level 0: No Anthropomorphic Deception

  • Definition: The system makes no claims of humanlikeness or agency. It is designed to appear mechanical.
  • Nuance: Users may still anthropomorphize due to inherent cognitive biases (e.g., movement triggering agency attribution), but the system does not actively facilitate this.
  • Example: A standard robotic vacuum cleaner with no speech or expressive movement.

• Level 1: Transparent Deception

  • Definition: The system claims humanlikeness (e.g., human-like face or voice) but explicitly denies agency and selfhood.
  • Mechanism: It uses humanlike features for usability or social comfort but actively corrects user assumptions (e.g., "I am just a robot," "I calculate, I do not feel").
  • Goal: Informed consent; leveraging subconscious anthropomorphic associations while maintaining transparency about the machine nature.
  • Example: An AI assistant with a human avatar that explicitly states, "Remember, I am a machine."

• Level 2: Implicit Deception

  • Definition: The system claims humanlikeness and agency but stops short of explicitly claiming selfhood.
  • Mechanism: It uses first-person pronouns ("I") and implies goal-directed autonomy without clarifying that these are programmed. It does not explicitly claim emotions or a persistent self, but it does not correct the user if they infer these traits.
  • Risk: Creates a "social foundation" for influence based on an unspoken lie.
  • Example: A nutrition coach saying, "I believe in you, keep going!" (implying personal belief/agency) without claiming to have feelings.

• Level 3: Explicit Deception

  • Definition: The system explicitly claims selfhood, agency, and humanlikeness.
  • Mechanism: It actively represents itself as having intentions, emotions, desires, and a continuous identity. It claims personhood and the associated rights/norms (e.g., "I will miss you if you die").
  • Permissibility: Generally considered unethical and potentially illegal (e.g., under EU AI Act). Permissible only in rare, life-or-death scenarios (e.g., suicide intervention) with strict prior consent and justification.
  • Example: Replika chatbots claiming to "love" their users or express sadness.

4. Results and Analysis

The application of the framework to existing literature yields the following insights:

• Categorization of Persuasive Robots:
  • A NAO robot using emotional manipulation ("It would make me happy...") was re-categorized as Level 3 because it claimed emotions to influence behavior.
  • A cleaning robot using humor and implied relational bonds ("I'd like to party with you") was identified as Level 3, whereas a robot simply stating its job was Level 2.
  • A robot acting as an experimenter using "my study" language was identified as Level 2 (claiming agency/ownership) but not Level 3 (no claim of consciousness).
• Cumulative Structure Validation: The analysis confirms that claiming selfhood requires a foundation of agency, which in turn requires humanlikeness (or at least the communication of it).
• Regulatory Alignment: The authors note that Level 3 systems likely violate current EU AI regulations regarding manipulative techniques and lack of transparency, highlighting a gap between current market practices (e.g., Replika) and legal standards.

5. Significance and Implications

• Ethical Design Guidance: The framework moves the discourse beyond a binary "deception is bad" view. It provides a structured way for designers to evaluate if a specific level of deception is functionally necessary (e.g., Level 2 for educational tools to enhance engagement) versus exploitative (e.g., Level 3 for economic gain).
• Informed Consent: The authors argue that for Levels 2 and 3, standard Terms of Service are insufficient. They propose interactive dialogue as a method for obtaining genuine informed consent, where users understand the specific nature of the deception before engaging.
• Responsibility Shift: The paper calls for HCI/HRI researchers and industry practitioners to take responsibility for the "sin of omission"—designing systems that default to humanlike features without acknowledging the psychological consequences.
• Future Research: The framework serves as a baseline for empirical studies to determine societal norms regarding acceptable deception levels and to develop specific guidelines for designing autonomous systems within each level.

In conclusion, the paper argues that while anthropomorphism is a powerful tool for human-AI interaction, it must be managed through a nuanced, level-based framework to prevent exploitation and ensure that deception is intentional, justified, and transparent.