Table 1.

List of common barriers to the successful implementation of artificial general intelligence

Barrier to artificial general intelligence	Definition	Impact
Algorithmic bias^20,21	Outcomes reflective of the quality of the input and the knowledge of the designer.	This has the potential to reduce applicability of output in diverse populations. The learning algorithms may also adopt poor practice from historical data that humans would inadvertently avoid.
Reward hacking²²	Focus in algorithms on outcomes that are seen as ‘successful’ with poor consideration of the means, process or long-term goals.	The automated system finds methods to achieve short-term goals eg dosing heparin prior to coagulation level measurements.
Insensitivity to impact²²	Inconsideration of the consequence of the outcome in the decision-making process.	Human doctors tend to err on the side of caution with diagnoses surrounding malignancy, leading to a high false positive but also a high sensitivity (few false negatives). If machine learning systems focus on accuracy, the implications for higher positive predictive value and specificity (higher true positive, lower false positive) may not be optimised for care of the holistic patient.
Automation bias²³	Reliance on the output of automated systems.	This occurs where clinicians rely on the results of their tools despite evidence to the contrary. Furthermore, eagerness and funding pressure may lead to the premature or inappropriate application of under-developed systems.