The Context Problem in Artificial Intelligence (2022) (cacm.acm.org)

The 2022 column diagnoses a pervasive “context problem” in AI: many high-profile failures—from facial-recognition and road-sign errors to LLM hallucinations and deadly drone targeting mistakes—stem from systems being used outside the contexts they were trained for. The piece traces this to the longstanding AI “frame problem”: machines (including modern ANNs) encode fixed internal connections learned from training data and cannot reliably sense or adapt to subtle, real‑time changes in situ. That brittleness is amplified by biased or unrepresentative training sets and the impracticality of rapid retraining, making it unsafe to assume models will generalize across contexts—an urgent concern where lives and high-stakes decisions are involved. Practically, the author argues the solution is not purely technical but design‑centric: embrace human–machine teaming supported by rigorous Human–Computer Interaction. Machines should handle high-speed search and simulation; humans should retain tasks requiring contextual judgment, ethics, and relevance. Four guiding principles are offered for trustworthy couplings: design for human–machine differences, use AI to overcome human bias, exploit simulations to surface possible futures, and test relentlessly in the actual context of use. For practitioners, the implication is clear: prioritize context-specific testing, HCI-informed interfaces, and system designs that detect or defer when context falls outside model competency rather than overtrusting opaque, brittle autonomy.