AI Reassesses Need for a Sense of Smell

Philip Maughan, writing in Noema Magazine, pushes back against breathless headlines that claim AI can now 'smell.' He notes that many stories conflate pattern recognition in language models with genuine olfactory sensing. The most dramatic claims come from media summaries, while careful reading of outlets like BBC Future shows the underlying behavior is an LLM repeating human associations about colors, shapes, and tastes captured in text corpora, not detecting volatile compounds.

The gap between perceived and actual progress is structural. Current systems that appear to 'smell' fall into two categories: multimodal models trained on correlational text-image or text-audio data, and prototype sensor stacks called 'electronic noses' that read chemical signatures. Neither is yet close to human olfaction in reliability or generality. Key technical pain points include:

• •sparse, noisy ground truth: labeled smell datasets are tiny and subjective, complicating supervised learning
• •sensor variability: chemical sensor arrays exhibit drift, cross-sensitivity, and require calibration
• •representation gap: mapping molecular structure and concentration to perceptual labels needs physics-aware features and chemoinformatics

Practitioners should note that LLM behavior alone is not evidence of olfactory capability. Building practical olfaction requires integrating hardware (sensor arrays, gas chromatography), domain models from chemistry, and multimodal training pipelines that handle temporal drift and environmental confounders.

This matters because olfaction unlocks high-value applications: early disease biomarkers in breath analysis, spoilage and contamination detection in food supply chains, and industrial safety monitoring for toxic leaks. The field sits at the intersection of AI, analytical chemistry, and instrumentation, so progress depends less on scaling transformer parameter counts and more on cross-disciplinary datasets, robust sensor platforms, and realistic field trials.

Track work that combines calibrated sensor hardware with principled molecular representations, reproducible field datasets, and benchmarking protocols for drift and cross-environment generalization. Funding and collaborations that bridge chemistry and ML will determine whether artificial olfaction moves beyond appealing headlines into reliable products.