Guide Labs

Guide Labs is a product-focused research company developing interpretable and auditable AI systems and foundation models that humans and domain experts can understand, steer, and debug. The company's stated motivation is that models optimized only for narrow performance can produce explanations unrelated to their actual decision-making, and that self-explanations such as chain-of-thought can be unreliable. Its approach integrates interpretability, safety, and reliability constraints into the model development pipeline, with the aim of helping users identify causes of erroneous outputs, detect spurious signals, and correct models.

The company's work includes concept bottleneck generative models and language models constrained to explain outputs in human-understandable terms, applied to image generation and protein design. Other research areas include faithfulness-measurable masked language models, interpretable mixture-of-experts architectures, saliency-guided training, error discovery through clustering influence embeddings, and methods for attributing model bias to training data. The company's Clarity platform is presented as an interpretable AI platform, with a prompt-concept feature for identifying which part of a prompt is responsible for an output. The company describes itself as a product-focused research company with a team experienced in AI interpretability and reliability, and publishes research at machine learning conferences.