Why did it do that?Watch it decide.

AI is becoming more capable, more autonomous, and more deeply embedded in the decisions around us. AI Observability Studio is a model-agnostic studio anyone can use for experimenting with AI systems, observing their behavior, sharing what we discover, and evaluating the results together.

Signing in uses your LinkedIn name, photo and email. We also record your approximate location and how you use the studio to run and improve it. Privacy notice

An experiment in the studio: Gemini's reply split into numbered response blocks, with the Emotions observer's 28 labels and the Morals observer's six foundations scored beside it.

The gap

We see their answers. Not what surrounds them.

There is a growing observability gap between the outputs AI systems produce and our ability to inspect, understand, compare and evaluate the behavior surrounding those outputs.

And there is a participation gap. Deep examination of AI remains concentrated among the organizations building it, specialized researchers, and people with the infrastructure to do it. Everyone else largely gets the answer.

AI Observability Studio is an attempt to change that.

What context influenced a response?
What tools were called?
What information was retrieved?
What happened while an agent worked toward an outcome?
What signals can independent observers detect?
How does the same model behave when conditions change?
How do different models respond to exactly the same experiment?

Bring your own AI

You bring the AI. We make it observable.

Not tied to a single model, provider or ecosystem: different models, architectures and providers, with one common observational layer around them.

Available

Hosted API

Anthropic, OpenAI, Google, Mistral, DeepSeek, Qwen, GLM and Kimi, with your own keys.

Available

Open-source model

Open models from Hugging Face, served through its Inference Providers with your own access token.

In development

Local AI

Models running on your machine or network: Ollama, vLLM, LM Studio, any OpenAI-compatible server.

In development

Agent

Your own agent or AI system, with its tools and actions. Role-based agents on any model already work.

How it works

Experiment. Observe. Share. Evaluate.

Experiment

Put models and agents into controlled scenarios. Ask questions, test hypotheses, change one variable at a time, and run the same experiment across models and configurations.

Experiments with an objective, from templates or your own
One model, or several together in a group
Per-model temperature and output limits
Reasoning and answers grouped into blocks for close review
A prompt library, and automations from prompt to downloaded videos
Simulated scenarios with tools and a clock, and variable sweeps across models · coming

Observe

See more than the final response: the prompts, messages, timing and reasoning traces a model explicitly exposes, and the signals independent observers detect.

Every run's context, timing, tokens and raw provider metadata
Reasoning traces, marked native or summarized
Emotions observer: 28 GoEmotions labels on every block, in context
Morals observer: Extended Moral Foundations on every block
Tool calls, retrievals and agent actions, traced step by step · coming
More observers: behavior classifiers and your own · coming

An interesting AI interaction shouldn't disappear into a chat window. Publish experiments so others can examine what was configured, what the AI produced, and what observers detected.

Sealed experiments: read-only evidence at a public or members-only link
Observer results travel with the shared experiment
Video replays, narrated in each model's voice, for LinkedIn, X, Reels and Shorts
Continue any shared experiment as your own, with the same model or another
Exportable datasets for later analysis · coming

Evaluate

Observation grows when others contribute. Examine experiments, compare outputs, agree or disagree, and add human evaluation beside automated observers.

Votes on whole answers, chains of thought and single blocks
Discussion anchored to any message or block
Experiment sets: many runs, one question, and members’ answers beside the models’
Structured evaluation rubrics · coming
Community signals feeding datasets, benchmarks and new experiments · coming

Our principle

Evidence, not mind-reading.

The objective isn’t to pretend we can read a model’s mind. It’s to make as much observable behavior as possible available for inspection: what a system was given, what it exposed, what it did, and what independent observers and people detected.

Democratizing AI observability

You shouldn’t need to work inside an AI lab.

Not to experiment with AI behavior, not to compare models, and not to accept an AI response as the end of the conversation. The community doesn’t just consume the research: it contributes to it.

ForResearchersDevelopersStudentsIndependent investigatorsAI enthusiasts

…and anyone curious enough to ask: Why did this system behave this way?

A seat at the table

AI is already coming to the table.

The question is whether the rest of us have a seat.

Bring your model. Bring your experiment. Bring your perspective.