Do Generative AIs Really Understand Safety?

Generative AIs like ChatGPT are moving into construction sites, factories, and offices. They are fast, answer questions, and help plan tasks. Yet their digital brain differs from ours. Researchers examined four core mechanisms of Transformer models: turning data into numbers (vectorization), marking word positions (positional encoding), calculating attention (attention scoring), and optimizing outputs (optimization functions). They compared these mechanisms with human reasoning. To test the gap, they asked an AI for fall‑prevention advice on a construction site, then measured how close the AI’s answers were to those of a human expert using cosine similarity and attention maps. Six “architectural distances” emerged: differences in focus, word meaning, and response stability. Such gaps can cause misunderstandings in hazardous situations where trust matters. The authors propose designing new architectures that better mimic human cognition, lowering risks and making AIs more reliable for safety applications.

The Transformer model processes text by converting it into high‑dimensional vectors, a process known as vectorization. These vectors represent words or tokens in a numerical space where Euclidean distance reflects semantic similarity. Positional encoding adds information about the sequential position of tokens, using sinusoidal functions or learned parameters, allowing the model to distinguish "before" from "after." Attention scoring computes, for each token, a relative weight toward all other tokens, creating an attention matrix that indicates which parts of the context the model focuses on. Optimization functions employ gradients and loss functions, typically cross‑entropy, to update weights during training. To assess cognitive divergence, the authors used cosine similarity between AI response vectors and those of human experts, obtaining average values below 0.7, indicating semantic gaps. They also examined attention maps: in safety scenarios, the AI tends to focus on technical terms while ignoring critical environmental context such as obstacles or adverse weather. The six identified architectural distances include: (1) semantic focus mismatch, (2) loss of spatial context, (3) response instability to minor input changes, (4) bias toward more frequent training data, (5) limited causal reasoning capacity, (6) difficulty handling uncertainty. These shortcomings suggest the need to hybridize Transformers with symbolic reasoning modules or cognition‑inspired neural networks to improve alignment between AI and safety‑critical decisions.