When Artificial Intelligence Fails
John Spacey, updated on August 08, 2023
Technology such as algorithms or artificial intelligence are often portrayed as logical entities that completely eliminate the errors in thinking that are common to humans such as biases. This is magical thinking that doesn't reflect the reality of computing. The following are a few examples of how computing produces illogical and irrational results.
Machine BiasesArtificial intelligence learns and self-improves. For example, a machine may build statistical models that can be used to make predictions. If you show an artificial intelligence a movie in which all the white people are bankers, the machine may develop a model that predicts most white people are bankers. As with human learning, artificial intelligence can develop biases based on factors such as data that is used in training.
Machine LearningMachine learning is by no means an exact science nor is it likely to become an exact science because learning is based on broad philosophical questions. Artificial intelligence produces vastly different results depending on the learning algorithm that is used. In other words, machines will disagree just as humans do.
Committee MachinesIn order to compensate for differences in machine learning algorithms, designers of artificial intelligence may use techniques such as committee machines that aggregate results from multiple learners. This is similar to social processes in humans and may open up the potential for the machine equivalent of social biases such as the abilene paradox.
HeuristicsBoth algorithms and artificial intelligence are commonly based on heuristics that come up with a good guess as opposed to an answer that is definitely correct. This is much how humans think and isn't anywhere close to purely logical.
Fuzzy LogicIf you did try to build an artificial intelligence based on traditional true/false logic that can be shown to be correct, such a machine would be incredibly unintelligent. Real life has grey areas between true and false that humans commonly understand. Machine techniques such as fuzzy logic can also handle grey areas. It order to account for grey areas you have to give up certainty whereby logic is "correct."
OutliersArtificial intelligence is often based on statistical models that aren't very good at explaining outliers in data. In other words, machines may stumble on obscure exceptions. Human abilities such as creativity and synthesis currently outdo machines in handling special cases and extreme situations such as black swan events.
Biases as ComplexityBacteria don't have as many cognitive biases as humans. Although biases are commonly viewed as unintelligent they occur because humans are extremely complex thinkers. It is possible that an advanced artificial intelligence would have far greater problems with biases than humans. Such biases would be very difficult to understand and might require a new field of practice such as machine psychiatry.
Artificial IntelligenceThis is the complete list of articles we have written about artificial intelligence.
If you enjoyed this page, please consider bookmarking Simplicable.
A few logic terms explained. A classical law of logic first established by Aristotle.
Logic that allows for partial truths.
The difference between logic and intelligence.
The definition of causality with examples.
The definition of magical thinking with examples.
The definition of scientism with examples.A list of logical fallacies.
The definition of mutually exclusive with examples.
The definition of false balance with examples.A few common types of artificial intelligence.
Technological singularity explained.Artificial intelligence and emotion.
An overview of artificial life.
A definition of deep learning with examples.
The difference between supervised and unsupervised learning with an example.The common types of natural language processing.
Common types of autonomous systems.
Common examples of artificial intelligence.
TrendingThe most popular articles on Simplicable in the past day. Recent posts or updates on Simplicable.