Artificial intelligence is not one machine with one purpose. The term covers systems designed for different tasks, using different methods and operating with different degrees of autonomy. Useful discussion begins by identifying what a system does, how it was evaluated and who is responsible for its deployment.
AI has a history, not a single birthday
The Dartmouth research proposal was dated 1955 and proposed a summer project for 1956. It explicitly used the phrase “artificial intelligence” and explored learning, language and problem-solving. This primary document corrects the common shorthand that the term first appeared only at the meeting. It records an ambitious research agenda, not the arrival of complete human-like intelligence.
Jobs: neither total replacement nor guaranteed protection
The ILO’s 2025 occupational-exposure study distinguishes tasks that generative AI might affect from jobs that will actually disappear. Occupations contain mixtures of activities, and adoption, work organisation and training matter. It is therefore misleading either to promise that AI will replace everyone or to insist it can only complement workers. Exposure is not an unemployment forecast, and outcomes are not determined by technology alone.
Learning from data takes different forms
Supervised learning uses examples with target answers; unsupervised learning looks for structure without those labels; reinforcement learning uses feedback about actions. Curriculum learning concerns how training examples or tasks are ordered, rather than being an equivalent separate category. Training changes a model according to an objective. It does not mean every deployed system continually learns from every conversation.
Capability is not the same as consciousness
Producing fluent language or solving a benchmark does not, by itself, establish human-like understanding, feelings or subjective experience. Researchers studying possible AI consciousness examine theory-derived indicators, not just convincing conversation. Their assessments are provisional and theory-dependent. Neither an impressive demonstration nor a confident slogan settles all questions about minds.
Limitations extend beyond training data
Data quality matters, but so do model design, objectives, evaluation and deployment conditions. A system may work well in a test and fail when its environment changes. Claims about understanding sarcasm or using common sense should be tested in specified situations, rather than treated as universal abilities or permanent impossibilities.
“Neutral technology” does not remove responsibility
NIST describes AI as sociotechnical: technical choices interact with people and institutions. Bias, privacy risks and harmful outcomes can arise from design and use, not only a malicious user’s intentions. Developers and deploying organisations remain responsible for suitable evaluation, oversight and ways to address failures. Calling a system neutral is no substitute for examining consequences.
The cover is a fictional robot illustration, not a depiction of a particular AI system or proof of machine consciousness.
Sources: Bengio et al. (2009); Dartmouth proposal (1955); ILO & NASK (2025); Google Machine Learning; Butlin et al.; NIST AI RMF.




Discussion