The alignment problem was popularised by author Brian Christian in his 2020 book The Alignment Problem: Machine Learning and Human Values. In the book, Christian outlines the challenges of ensuring AI models capture “our norms and values, understand what we mean or intend, and, above all, do what we want.” The alignment problem describes the problems associated with building powerful artificial intelligence systems that are aligned with their operators.
Understanding the alignment problem
Artificial intelligence has come a long way in recent years, with humankind now creating machines that can perform remarkable feats.
But after six decades of intensive research and development, aligning AI systems with human goals and values remains an elusive task.
With every major field of artificial intelligence trying to replicate human intelligence, problems invariably arise when developers expect AI to act with the rationality and logic of a person.
Growing interest in machine and deep learning has meant the algorithms underpinning everything from baseball games to oil supply chains are being digitized.
This process is helped by high-speed internet, cloud computing, the internet of things (IoT), mobile devices, and a plethora of emerging technologies that collect data on anything and everything.
While machine learning algorithms scale well with the availability of data and computing resources, they are nonetheless complex mathematical functions comparing observations to programmed outcomes.
In other words, artificial intelligence is only as robust as the data used to train it.
When training data is poor quality or simply insufficient, algorithmic output suffers. This scenario represents the essence of the alignment problem.
Real-world examples of the alignment problem
In his book, Christian explains several cases where machine learning algorithms have caused embarrassing and sometimes damaging failures.
An algorithm used by the search engine giant in facial recognition software tagged people with dark skin as gorillas.
Had Google trained the algorithm with more examples of people with dark skin, the failure could have been avoided.
Amazon’s recruitment tool once used artificial intelligence to give job candidates a score between one and five stars.
In theory, this would allow the company to identify promising candidates amongst hundreds of resumes.
However, the model was trained to vet applicants by observing patterns in resumes submitted over a decade-long period.
Since most applications came from men, the algorithm automatically disqualified female applicants as a result.
Real-world examples of the alignment problem were also mentioned by author Cathy O’Neil in her book Weapons of Math Destruction: How Big Data Increases Inequality and Threatens Democracy.
In the book, O’Neil explained how blind faith in algorithms caused pervasive damage to many aspects of consumer life. Some examples include:
- Credit scoring systems that wrongfully penalize people.
- Recidivism algorithms give defendants of a certain race or ethnicity a heavier prison sentence.
- Teacher-scoring algorithms reward teachers who game the system and terminate honest, high-performing teachers.
- Trade algorithms that make billions of dollars profit at the expense of low-income classes and so-called “mom and pop” investors.
- The alignment problem describes the problems associated with building powerful artificial intelligence systems that are aligned with their operators. The concept was popularised by Brian Christian in his book The Alignment Problem: Machine Learning and Human Values.
- At the core of the alignment problem is poor quality or insufficient algorithm training data. With data now being logged in almost every aspect of daily life, there is a higher likelihood of algorithms making poor decisions because of an overreliance on their mathematical functions.
- The alignment problem resulted in Google facial recognition models classifying people with darker skin as gorillas, while a similar mishap at Amazon caused its recruitment algorithm to ignore female applicants. Blind faith in algorithms has also resulted in arguably more sinister and pervasive consequences for the average consumer.
Main Free Guides: