Summary
- Claims should separate broad task coverage from robust reasoning, memory, agency and adaptation.
- Evaluation needs unfamiliar tasks, changing conditions, failure costs and independent replication.
The idea of general AI describes capability across domains rather than excellence on one narrow task. Modern models can appear broad because one interface reaches many learned patterns, yet breadth may collapse under new rules, missing context or prolonged action. Developers and buyers should publish capability boundaries, contamination controls, resource needs and the consequences of error. The next useful evidence is an independent evaluation on genuinely unseen work, including requests the system should refuse or hand back. A label should follow demonstrated transfer and control, not lead them.
