70 Years of AI.1956–2026

Dartmouth gave a developing field a shared name and an ambitious agenda. It did not invent every idea behind AI: work on computation, logic and neural models came before it.

This was an important route toward machine learning, not a complete model of the brain. A single-layer perceptron has strict limits on the patterns it can separate.

ELIZA exposed a gap that still matters: convincing conversation can encourage people to attribute understanding to a system. Fluency alone is weak evidence of intelligence.

An “AI winter” describes a loss of confidence and support, not a worldwide halt in research. The lesson was to judge systems by what they could reliably do beyond a demo.

AI found commercial uses. But acquiring, maintaining and reconciling large rule sets was difficult, and systems were often brittle outside the situations their experts had anticipated.

AI’s history is not a clean succession of replacements. Learning systems and symbolic approaches coexisted, and important work survived the retreat of the AI label.

A machine could surpass an exceptional human in a prestigious intellectual task. Deep Blue’s victory did not demonstrate general intelligence or human-like understanding.

The result accelerated a shift toward learning visual features from data. Deep learning was older than 2012; ImageNet supplied a compelling demonstration at scale, not its invention.

Learning and search together could tackle a game long regarded as a difficult frontier. The achievement widened expectations for AI while remaining a specialised result.

The architecture became a foundation for many language and multimodal models. It was a major design change, not the invention of attention or a guarantee of truthful output.

Prompting became a practical alternative to training a separate model for every task. Performance was uneven, and convincing text could still be factually wrong.

2022 was a turning point in access and use, not the birth of diffusion research. Text-to-image tools expanded creative possibilities and intensified questions about training data, consent and authorship.

The turning point was also social: language models became easier to encounter and use. The launch explicitly acknowledged limitations, including plausible but incorrect answers.

The interface moved closer to the way people encounter the world. Recognising an image or responding to speech still does not ensure reliable reasoning about it.

Tool use makes AI useful in a different way, and makes mistakes consequential. The frontier is reliability: permissions, verifiable results and human oversight matter as much as an impressive answer. Benchmark success is not a promise of dependable autonomy.