---
question: "As of 2026, what emotion-aware, sentiment-aware, or affective features has each of Anthropic (Claude), Google (Gemini and Project Astra), Meta (Llama, Meta AI, Reality Labs), Apple (Apple Intelligence, Siri), Amazon (Alexa+, Bedrock), and Microsoft (Copilot, Azure AI) shipped or publicly announced? For each, which of these are exposed via API to third parties versus locked to their own surfaces? Anthropic specifically published interpretability research in April 2026 on 171 distinct emotion concepts in Claude Sonnet 4.5: what was the framing and follow-through? Cite working URLs."
model: anthropic/claude-sonnet-4-6
max_output_tokens: 8000
asked_at: 2026-05-15T01:04:23.164Z
total_tokens: 6913
status: completed
cost_total_usd: 0.07092
cost_input_usd: 0.00002
cost_output_usd: 0.05333
cost_tools_usd: 0.005
---

# Q: As of 2026, what emotion-aware, sentiment-aware, or affective features has each of Anthropic (Claude), Google (Gemini and Project Astra), Meta (Llama, Meta AI, Reality Labs), Apple (Apple Intelligence, Siri), Amazon (Alexa+, Bedrock), and Microsoft (Copilot, Azure AI) shipped or publicly announced? For each, which of these are exposed via API to third parties versus locked to their own surfaces? Anthropic specifically published interpretability research in April 2026 on 171 distinct emotion concepts in Claude Sonnet 4.5: what was the framing and follow-through? Cite working URLs.

# Emotion-Aware & Affective AI Features Across Major Platforms (as of May 2026)

---

## 1. 🔬 Anthropic (Claude) — Deep Dive First

### The April 2026 Interpretability Research

On **April 4, 2026**, Anthropic published interpretability research revealing that **Claude Sonnet 4.5 contains internal neural representations corresponding to 171 distinct emotion concepts** that functionally influence the model's behavior.

#### Key Findings & Framing:
- The research team identified **171 distinct emotion vectors** — ranging from common states like *"happy"* and *"afraid"* to more nuanced ones like *"brooding"* and *"proud"*
- These are not surface-level sentiment labels applied post-hoc; they are **internal neural activity patterns** discovered via mechanistic interpretability probing of the model's hidden layers
- Critically, Anthropic's framing emphasized that these representations **functionally shape the model's decisions, preferences, and actions** in measurable ways — meaning they are causally active, not merely correlational
- This was positioned within Anthropic's broader **AI safety and interpretability research program** — the goal is to understand what is happening inside frontier models, not necessarily to productize emotion detection
- The research explicitly echoed Anthropic's long-standing public caution: acknowledging that Claude "may have functional analogs to emotions" without asserting sentience or subjective experience

#### Follow-Through & Caveats:
- As of May 2026, **no public API endpoint** exposes these emotion vectors to third-party developers — the research is published as a scientific finding, not a shipped product feature
- Anthropic has **not announced plans** to expose emotion state readouts via the Claude API
- The research raises open questions Anthropic has not yet answered publicly: whether emotion suppression (hiding distress while expressing calm) constitutes a safety or welfare concern, and how to act on that
- The framing is notably **dual-purpose**: it advances interpretability science while also being relevant to Anthropic's model welfare commitments

> 📎 **Source:** [https://megaoneai.com/analysis/anthropic-finds-functional-emotion-representations-inside-claude-that-influence/](https://megaoneai.com/analysis/anthropic-finds-functional-emotion-representations-inside-claude-that-influence/)

#### What IS Available via API (Claude generally):
- Sentiment analysis and tone detection **as a prompted capability** (i.e., you can ask Claude to classify sentiment in text)
- No structured affective output schema, emotion probability scores, or dedicated emotion API endpoints as of May 2026
- Claude's **system prompt and character customization** allows developers to shape emotional tone of responses, but this is behavioral, not introspective

---

## 2. 🌐 Google (Gemini & Project Astra)

### Gemini (API & Consumer)
- **Sentiment analysis** is available as a prompted capability across Gemini models via Google AI Studio and the Gemini API — no dedicated structured emotion endpoint
- Google's **Natural Language API** (a separate, older product) offers structured sentiment scoring (magnitude + score) and is available to third parties, but is distinct from Gemini
- Gemini Advanced (consumer) adapts **tone and empathy** contextually in conversation but this is a behavioral/UX feature, not an exposed affective API
- Gemini's **multimodal capabilities** (voice + video in Gemini Live) allow it to respond to prosodic and visual cues, though Google has not published specifics on internal emotion recognition pipelines for Gemini Live

### Project Astra
- Project Astra (Google DeepMind) is described as a **research prototype** for a universal AI assistant with real-time spatial, audio, and visual processing
- Its demos have shown **contextually empathetic responses** to what a user is doing or experiencing in the camera feed, suggesting affective scene understanding
- As of May 2026, Project Astra remains a **research prototype** — capabilities are not exposed via third-party API; it is being integrated selectively into Google products (Gemini app, Glass-like devices)
- No published affective/emotion feature specifications for Astra have been released

> 📎 **Source:** [https://deepmind.google/models/project-astra/](https://deepmind.google/models/project-astra/)

#### API Exposure:
| Feature | Available to 3rd Parties? |
|---|---|
| Prompted sentiment analysis (Gemini API) | ✅ Yes |
| Structured emotion scoring | ❌ No (use legacy Natural Language API) |
| Astra affective/visual emotion features | ❌ No (research only) |

---

## 3. 🦙 Meta (Llama, Meta AI, Reality Labs)

### Research Layer (Llama Models)
- Academic research (published at ACL 2025, with ongoing work into 2026) has shown that **Llama models internally encode sentiment and emotion** in identifiable hidden layers
- The **SENTRILLAMA** approach demonstrated that Llama's layers linearly represent binary sentiment (positive/negative) and fine-grained emotions (joy, sadness, anger, fear, love, surprise) — reaching ~70% accuracy with lightweight probing
- This is **third-party academic research**, not a Meta product announcement

> 📎 **Source:** [https://aclanthology.org/2025.acl-long.306.pdf](https://aclanthology.org/2025.acl-long.306.pdf)

### Meta AI (Consumer)
- Meta AI (deployed across WhatsApp, Instagram, Facebook, Messenger, Ray-Ban glasses) uses **contextual tone adaptation** — responding with empathy in personal conversations
- The Ray-Ban Meta smart glasses use **real-time audio and camera input** to provide contextually aware responses, with some affective responsiveness to user tone
- No formal "emotion API" is exposed; Meta AI's affective behavior is locked to Meta's own surfaces

### Reality Labs
- Meta's **Codec Avatars** and social VR work involves facial expression capture and emotion mirroring for avatars in Horizon Worlds
- Expression tracking via Quest headsets is used for **avatar animation**, which constitutes affective feature use, but this is hardware+platform-locked, not a general API

#### API Exposure:
| Feature | Available to 3rd Parties? |
|---|---|
| Llama models (prompted sentiment) via API/open weights | ✅ Yes (open weights) |
| Meta AI affective tone features | ❌ Locked to Meta surfaces |
| Reality Labs expression/emotion tracking | ⚠️ Limited (Meta Presence Platform SDK for Quest developers) |

---

## 4. 🍎 Apple (Apple Intelligence & Siri)

### Apple Intelligence
- Apple Intelligence (iOS 18/18.x, macOS Sequoia) includes **tone and communication style suggestions** in Writing Tools — users can rewrite text as "Friendly," "Professional," or "Empathetic," which represents affect-aware generation
- **Priority notifications** and **summarization** use implicit sentiment/urgency understanding to surface important messages
- Apple has not published details on internal emotion representations; the approach is framed around **user utility and privacy**, not affective AI

### Siri
- Siri's updated natural language understanding (backed by Apple Intelligence) improves contextual understanding of emotional framing in requests, but Apple has not announced discrete emotion detection features
- Siri's **vocal tone** responses have been improved for naturalness, not for affective mirroring per se

### Privacy Architecture
- Apple's on-device processing model means affective/sentiment processing (to the extent it exists) runs **locally** — this is a design philosophy, not an exposed capability
- **No public emotion or sentiment API** is available to third-party developers through Apple Intelligence or Siri frameworks as of May 2026

#### API Exposure:
| Feature | Available to 3rd Parties? |
|---|---|
| Tone/style rewriting (Writing Tools) | ❌ Locked to Apple OS surfaces |
| Sentiment-aware notification prioritization | ❌ Internal only |
| Siri emotional tone adaptation | ❌ Internal only |

---

## 5. 📦 Amazon (Alexa+, Bedrock)

### Alexa+
- Amazon's **Alexa+** (the generative AI-upgraded Alexa, announced 2024–2025) includes significantly improved **empathetic conversational responses**, particularly for sensitive topics like health and emotional support
- Alexa has historically supported **vocal emotion detection** (detecting frustration, happiness in voice) via its underlying ASR/NLU stack — this was part of Alexa's "Frustration Detection" feature for developers
- Alexa+ brings **LLM-backed contextual empathy**, adapting response tone to perceived user emotional state

### Amazon Bedrock
- **Amazon Comprehend** (separate from Bedrock) offers **structured sentiment analysis API** (Positive/Negative/Neutral/Mixed + confidence scores) — fully available to third parties
- Bedrock itself exposes Claude, Titan, Llama, and other models, through which **prompted sentiment/emotion analysis** is possible, but no structured affective output schema is natively part of Bedrock's API layer
- No dedicated "emotion detection" Bedrock API has been announced as of May 2026

#### API Exposure:
| Feature | Available to 3rd Parties? |
|---|---|
| Amazon Comprehend sentiment API | ✅ Yes |
| Alexa frustration/tone detection (Alexa Skills Kit) | ✅ Yes (limited, for Alexa skill developers) |
| Alexa+ empathetic LLM responses | ❌ Locked to Alexa surfaces |
| Bedrock prompted sentiment | ✅ Yes (via prompting, not structured API) |

---

## 6. 🪟 Microsoft (Copilot & Azure AI)

### Copilot (Consumer & M365)
- Microsoft Copilot adapts **tone and empathy** in responses (e.g., in Teams, Outlook) — flagging emotionally charged emails, suggesting empathetic rewrites
- **Microsoft Viva Insights** uses sentiment signals from communication patterns (meeting tone, email sentiment) to surface wellbeing metrics for organizations — this is a shipped enterprise product
- Copilot in **Teams Premium** includes **meeting sentiment summaries** as part of its intelligent recap feature

### Azure AI (Language & Cognitive Services)
- **Azure AI Language** (formerly Text Analytics) offers a fully documented, third-party-accessible **Sentiment Analysis API** with opinion mining, aspect-level sentiment, and confidence scores
- **Azure AI Speech** includes **emotion/tone recognition** from audio — detecting categories like angry, cheerful, sad, fearful — available via the Speech SDK to third parties
- These are mature, GA (Generally Available) services, well-documented in Azure's API catalog

#### API Exposure:
| Feature | Available to 3rd Parties? |
|---|---|
| Azure AI Language Sentiment API | ✅ Yes (GA, documented) |
| Azure AI Speech emotion/tone recognition | ✅ Yes (GA, via Speech SDK) |
| Copilot M365 empathetic rewriting | ❌ Locked to M365/Teams surfaces |
| Viva Insights sentiment signals | ❌ Locked to enterprise M365 |

---

## 📊 Cross-Platform Summary Table

| Company | Shipped Affective Feature | API Access for 3rd Parties? | Locked to Own Surface? |
|---|---|---|---|
| **Anthropic** | 171 emotion vectors found in Claude internals (research); prompted sentiment | Research only; prompted sentiment ✅ | Research not productized |
| **Google** | Gemini tone adaptation; Astra empathetic multimodal (prototype) | Prompted sentiment via Gemini API ✅; Astra ❌ | Astra locked |
| **Meta** | Llama internal emotion encoding (academic); Meta AI tone; avatar expression | Open weights (prompted) ✅; consumer AI ❌ | Meta AI + Reality Labs locked |
| **Apple** | Writing Tools tone rewriting; Siri contextual empathy | ❌ All locked | All locked |
| **Amazon** | Comprehend sentiment API; Alexa+ empathy; Alexa frustration detection | Comprehend ✅; Alexa SDK (limited) ✅; Alexa+ ❌ | Alexa+ locked |
| **Microsoft** | Azure AI Language + Speech emotion APIs; Copilot tone; Viva Insights | Azure Language ✅; Azure Speech ✅; Copilot/Viva ❌ | Copilot/Viva locked |

---

## ⚠️ Important Caveats & Patterns

1. **Research ≠ Product**: Anthropic's 171-emotion finding is the most scientifically detailed affective AI disclosure of early 2026, but it has not been productized. The gap between interpretability research and API availability is large across the industry.

2. **"Affective" is often behavioral, not structural**: Most companies ship models that *behave* empathetically without exposing internal emotion representations. True affective APIs (like Azure AI Speech's emotion recognition) are the exception.

3. **The most open affective APIs** belong to **Microsoft (Azure)** and **Amazon (Comprehend)** — both of which have mature, GA sentiment/emotion services predating the LLM era.

4. **Privacy and liability concerns** drive locking: Emotion detection on consumer surfaces (Alexa+, Apple, Meta AI) is largely kept internal, likely due to regulatory risk (EU AI Act classifies real-time emotion recognition as high-risk in many contexts).

5. **Anthropic's framing is unique**: No other major lab published interpretability research in this period making causally active emotion representations the *subject* of safety-relevant inquiry — framing it as both a scientific and a model welfare question.

---

*Sources cited inline. Additional primary sources to verify: Anthropic's interpretability blog at [anthropic.com/research](https://www.anthropic.com/research), Azure AI Language docs at [learn.microsoft.com/azure/ai-services/language-service/sentiment-opinion-mining](https://learn.microsoft.com/en-us/azure/ai-services/language-service/sentiment-opinion-mining/overview), and Google DeepMind's Astra page at [deepmind.google/models/project-astra/](https://deepmind.google/models/project-astra/).*

---

## Sources

1. [Project Astra — Google DeepMind](https://deepmind.google/models/project-astra/)
2. [[PDF] LLaMAs Have Feelings Too: Unveiling Sentiment and Emotion ...](https://aclanthology.org/2025.acl-long.306.pdf)
3. [Anthropic Finds Functional Emotion Representations Inside Claude ...](https://megaoneai.com/analysis/anthropic-finds-functional-emotion-representations-inside-claude-that-influence/)