Artificial Intelligence today is very sophisticated, complex and can be very accurate. But can we trust it?
Artificial Intelligence is a term already widely accepted and used in our technological society: in large and small companies, in the world of research and applied to practically all fields and areas in which research and development is carried out.
Artificial Intelligence consists of copying and formalizing qualities and characteristics that we perform with human intelligence, to apply them in an automated way to solve and optimize problems.
One of the most interesting qualities that we have copied from the brain is the ability to learn, which is called automatic learning. The computer, already for decades, can process some data (which are called training, because they are used to learn from them) and then face other new data that, even being a little different, can be recognized by putting into practice what it has learned (not memorized).
For example, one can learn from many images of cute kittens (among other animals) and subsequently be able to recognize a previously unseen image of another cute kitten.
Companies and research centers have been using this technology for several decades in order to solve complex problems that require learning from existing data and then applying the algorithm that has learned from it to new complex and unknown cases and scenarios.
The accuracy with which they get their prediction right has been increasing and the types of algorithms used to learn have become increasingly sophisticated. As a result, sometimes pharaonic projects are carried out to create “predictors” to anticipate with very high reliability what may happen with a customer in a company, with a diagnosis in health, with an incident in industry or with an event in society. The great confidence in these algorithms is leading to decisions being left in their hands, even in the criminal or professional recruitment fields.
However, experts in artificial intelligence and machine learning have long been aware that if the data the algorithm uses to learn is incorrect, the prediction the algorithm will make will also be somewhat incorrect. This is what is informally known as “trash in, trash out”.
The big problem is that not in all cases the data is clearly incorrect or deficient. There are numerous techniques for dealing with such data, cleaning it, correcting it, and completing it, even with unknowns in it. On many occasions, the data are apparently correct, but include a disproportion or imbalance in their quantity or caseness that makes them not quite complete. This is known as bias and can be thought of as learning only from a specific and homogeneous part of the information.
These biases have occurred on numerous occasions and in important projects with great repercussions, such as the attribution of crime risk associated with specific races (in which the algorithm learned from data in which the majority of the data on criminals were of that race, but had practically no cases from which to learn about non-criminals), or in the hiring of profiles of a specific gender over another (and equally motivated by not having enough information in the training data of one of the genders to be able to make a fair decision).
Hence the later coining of the informal phrase “bias in, bias out”, i.e.: bias in, bias out. This means that if the data we use to learn are biased, our prediction or attribution will be biased as well.
Apart from this unintentional situation, what is more dangerous is that there is also the possibility of malicious modification of the input data (called data poisoning) or of the scope of application to modify the prediction made by the algorithm in favor of the person who has made it. Just like a hacker.
This risk triggered alarms among the scientific community and made them rethink the reliability of artificial intelligence, not with the aim of losing faith in it, but quite the opposite: to generate a formal framework in which we can make it reliable. This concern led to the term already coined in Europe and with a very focused current follow-up, called: Trustworthy Artificial Intelligence (Trustworthy AI).
This framework is based on three fundamental pillars, which are to ensure that AI is lawful (and complies with all existing laws and regulations), ethical (ensuring basic principles of fairness and justice) and robust (both from a technical and social point of view).
To achieve these objectives, a series of requirements have already been put in place that must be reviewed, evaluated and assured in our smart systems to meet this reliability. The following seven requirements have been established at European level:
- Human action and oversight: including fundamental rights, human action and oversight.
- Technical robustness and security: including attack resilience and security, a fallback plan and overall security, accuracy, reliability and reproducibility.
- Privacy and data management: including respect for privacy, data quality, data integrity and access to data.
- Transparency: including traceability, explainability and communication.
- Diversity, non-discrimination and equity: including freedom from unfair bias, accessibility and universal design, and stakeholder participation.
- Social and environmental well-being: including sustainability and respect for the environment, social impact, society and democracy.
- Accountability: including auditability, minimization and reporting of negative impacts, trade-offs and trade-offs.
It is no longer enough to treat and clean data to apply artificial intelligence, nor to choose the best possible algorithm to make the best prediction. It is now necessary, at last, for our intelligent systems to be reliable.
Author: Alejandro Echeverría Rey, Senior Researcher in Artificial Intelligence/Machine Learning and Cybersecurity, Funditec.
