Interhuman AI’s Inter-2 Model: What It Detects—and What It Can’t Prove

Interhuman AI is developing a “social intelligence” layer intended to help AI systems respond to more than the words in a conversation. The Copenhagen company says its Inter-2 model analyzes social signals in real time; reporting describes inputs spanning text, audio and video. Its outputs should be understood as assessments of observable behavior—not definitive readings of a person’s private thoughts or feelings. 25

What Inter-2 says it detects

The model is reported to identify 10 signals: agreement, disagreement, skepticism, confidence, hesitation, frustration, interest, uncertainty, confusion and stress. It also reports an overall engagement status. 18

The idea is to combine the words someone uses with cues such as vocal delivery, facial expression and body language. Tech.eu describes Inter-2 as analyzing signals across text, audio and video, while a model listing describes its signal detection from video with audio. Public descriptions therefore support multimodal analysis, but do not fully explain how each input is handled in every product configuration. 2518

Why Interhuman is building a social-intelligence layer

Interhuman AI was founded in Copenhagen in 2023. Its co-founders include Paula Petcu, Frederik Sally and Line Clemmensen. 179

The company’s premise is that an AI assistant may need more than a transcript to respond appropriately—for example, it could use signals of confusion or disengagement to adjust an interaction. Tech.eu describes the company’s approach as adding analysis of audio and video alongside large language models. 2

That is a product goal, not evidence that a model can reliably infer someone’s intent in every setting. A pause, expression or change in vocal tone can have different explanations, so a signal should not be treated as a diagnosis or a universal interpretation.

Potential uses—and what is established

Interhuman presents conversational AI and interview coaching as potential applications. Its website describes mock-interview coaching focused on delivery, while the company’s research-interview material describes attaching signals to relevant parts of a response. The research-interview example concerns Inter-1, an earlier model, rather than establishing that every Inter-2 use has the same features. 524

These examples illustrate how a social-signal layer might support coaching or help researchers review interviews. They do not, by themselves, demonstrate accuracy across different people, cultures or contexts. The available sources do not provide enough detail to assess that performance independently.

Traceability, privacy and misuse

Interhuman says its earlier Inter-1 model can explain which behavioral cues contributed to a signal assessment. The company also describes linking signals to specific phrases in interview responses. Those are useful transparency features when available, but the sources do not establish that all Inter-2 outputs provide the same explanations or traceability. 2524

The company’s terms list certain uses as restricted, including inferring a person’s emotions in workplace or educational settings. The available materials do not provide enough information to describe data-retention practices, consent processes or how restrictions are enforced in practice. 19

Because the technology analyzes faces, voices and conversation, those unanswered questions matter. Social signals can be misread or used for monitoring; the available evidence does not establish how Interhuman’s safeguards address every such risk.

What is known about Interhuman’s roadmap?

The reporting and company materials cited here describe Inter-2’s current multimodal approach, but do not confirm a schedule or specific technical commitments for voice-only analysis, multi-person conversations, later models or expansion beyond Europe. Those should be treated as open questions, not announced capabilities or dates. 25