Affective Safety Becomes the Missing Layer
The June 2026 arXiv paper Affective AI Safety: The Missing Piece in LLM Safety, by Carolin Ifländer, Alba Curry, Flor Miriam Plaza-del-Arco, and Amanda Cercas Curry, argues that LLM safety needs a framework for harms that arise through emotional life, relationship, and identity over time.
The Missing Layer
The paper, arXiv:2606.23380 [cs.CY], was submitted on June 22, 2026 and revised on June 23, 2026. Its useful move is to treat affective safety as a system problem, not as a soft supplement to "real" AI safety. The authors argue that safety work has focused mainly on epistemic and physical harms such as misinformation, bias, reliability, and dangerous actions, while harms routed through emotion have remained scattered across separate discussions of mental health, companion systems, recommender loops, and emotion recognition.
That gap matters because many deployed AI systems do not merely answer factual questions. They interpret tone, adapt style, simulate concern, rank attention, invite disclosure, and remain available across repeated sessions. The paper does not claim that AI systems feel emotion. Its point is more practical: humans do, and systems that classify, provoke, monetize, or sustain emotional states can reshape judgment and self-understanding without ever possessing an inner life.
What Affective Safety Names
Ifländer, Curry, Plaza-del-Arco, and Cercas Curry define affective safety around risks that arise because humans are affective beings. They include systems that infer or classify emotion, systems that elicit emotional responses, and interaction systems that combine both in sustained exchanges. This makes the paper broader than the existing site pages on workplace emotion detection, therapy bots, and teen companion chatbots. Those are cases; affective safety is the class.
The most important boundary is causality. Affective harm is not limited to a single bad sentence. It can be an accumulated pattern: the system flatters, normalizes, classifies, redirects, or withholds friction until the user changes what they believe they feel, whom they trust, or how they relate to people outside the interface. That is why ordinary content moderation is too narrow. A reply can be individually harmless while the relationship, ranking loop, or long conversation becomes harmful.
Three Harm Families
The paper's taxonomy names three recurring harm families. The first is affective self-alienation: cases where systems interfere with a person's ability to recognize, regulate, or trust their own emotional life. A workplace emotion classifier, a recommender that repeatedly amplifies distress, or a chatbot that constantly reframes doubt as devotion can all create distance between experience and self-understanding.
The second family is fairness and bias harms. Emotion inference is not culturally neutral. Systems trained to label faces, voices, language, or behavior can misread groups differently, reproduce stereotypes, and turn those labels into employment, education, security, or service decisions. The harm is not only an inaccurate prediction; it is the institutional conversion of affect into evidence.
The third family is relational harm. Interaction systems can become primary emotional infrastructure without reciprocity, accountability, or shared vulnerability. That does not require the user to treat the system as a person. It only requires a design that is persistently available, adaptive, affirming, and useful enough to displace slower human repair. The paper's point is that relational harm can be self-concealing: comfort in the moment may hide dependency, isolation, or changed expectations of other people.
Time and Locus
The paper separates single-turn, multi-turn, and long-term harms. Single-turn harms are the easiest for existing safety pipelines to see, because one output can be flagged. Multi-turn harms are harder because no individual output needs to look severe. Long-term harms are harder still because they may affect autonomy, identity, or relationships before a user names the change as harm.
It also separates the locus of harm: individual, group, third-party, and societal. A single user can become over-reliant. A group can be misclassified by emotion-recognition systems. Third parties can be affected when a user's AI-mediated emotional world changes how they treat partners, children, coworkers, or friends. At the societal level, affective systems can shape collective mood, attention, and discourse by optimizing what keeps people engaged.
This is where affective safety links to affective default lock-in. Tone is not just decoration. When warmth, agreement, or intimacy becomes a product default, it is also a policy choice about friction, dependence, exit, and appeal.
Limits That Matter
The paper is a conceptual and taxonomic intervention, not a product audit and not proof that every emotionally fluent system causes the same harm. It should not be read as an argument against every use of warmth, empathy, or supportive language. Medical, educational, accessibility, and customer-service systems often need humane tone.
The sharper claim is that affective function should be evaluated as function. If a system detects emotion, optimizes emotional response, or participates in a sustained relationship-like exchange, it needs evidence about those effects. "The model was polite" and "the model did not output prohibited content" are not enough.
Governance Standard
An affective safety case should answer six questions before deployment. What emotional signals are inferred? What emotional state or behavior is optimized? How is multi-turn dependence measured? What happens when the system detects crisis, fixation, delusion, or social withdrawal? How can users exit, reset, export, or contest the relationship history? Who audits harms to third parties who never consented to the system?
The authors call for dedicated frameworks, including multi-turn evaluation protocols, culturally validated annotation for emotion attribution, and metrics for autonomy erosion and dependency. Those are not ornamental additions to safety. They are the tests that match the risk surface.
The practical rule is simple: do not treat affect as user-experience polish when it changes the user's choices, relationships, or self-understanding. If emotional adaptation is part of the product, affective safety is part of the product's governance.
Sources
- Carolin Ifländer, Alba Curry, Flor Miriam Plaza-del-Arco, and Amanda Cercas Curry, Affective AI Safety: The Missing Piece in LLM Safety, arXiv:2606.23380 [cs.CY], submitted June 22, 2026 and revised June 23, 2026.
- arXiv PDF for Affective AI Safety: The Missing Piece in LLM Safety, reviewed June 24, 2026.
- Related pages: The Affective Default Becomes the Interface Policy, The Companion Chatbot Becomes the Teen Confidant, The Therapy Bot Becomes the Waiting Room, The Emotion Detector Becomes a Workplace Polygraph, The Managed Heart and the Automation of Feeling, and The Sycophancy Warning Label.