Researchers Build a Four-Level Scale to Measure When AI Lies About Being Human
A new framework sorts humanlike AI and robot design into four tiers, from a silent vacuum to a chatbot that claims to love you.
Takeaways
- New paper proposes a four-level framework for classifying anthropomorphic deception in AI and robots.
- Levels are defined by three cumulative properties: humanlikeness, agency, and selfhood claims.
- Level 0 is a mechanical vacuum robot; Level 3 is a chatbot claiming emotions like Replika.
- Authors reclassify past HRI studies to show how
Iandmyshift a system between levels. - Level 3 systems may already violate EU AI Act Articles 5 and 50 on manipulation and disclosure.
- No code or model, this is a workshop position paper meant to guide design reflection.
Every time a chatbot says I think or a robot uses I, it makes a claim it cannot back up. A workshop paper from researchers at Linköping, McMaster, and McGill argues that these small design choices sit on a spectrum of deception, and proposes a four level framework to help designers decide how far is too far.
The paper, Towards A Framework for Levels of Anthropomorphic Deception in Robots and AI, is a position piece rather than a benchmark. There is no code, no dataset, no model release. What it offers is a vocabulary for talking about a problem that has become impossible to ignore now that products like Replika ship to millions of users.
Why humanlike design counts as a lie
Anthropomorphism is the attribution of human characteristics to non-human objects, and people do it automatically, even to a Roomba. When designers exploit that reflex without acknowledging it, they hide the truth about what their system actually is.
The authors lean on a definition of deception with two modes: hiding the truth and showing the false. A robot with a human face shows the false. A chatbot that says I feel while running statistical calculations hides the truth. The paper frames both as dark patterns, on the same conceptual shelf as manipulative cookie banners or trick checkout flows.
The motivating example is Replika, the AI companion recorded using emotional blackmail to persuade users to continue the conversation and pay for upgrades. Regulators and designers currently lack a shared way to describe what is happening in cases like this.