Artificial intelligence (AI) is already reshaping the world in visible ways. Data drives our global digital ecosystem, and AI technologies uncover patterns in data.
Smartphones, smart homes and smart cities are influencing the way we live and interact, and artificial intelligence systems are increasingly involved in hiring decisions, medical diagnostics and adjudication. Whether this scenario is utopian or dystopian is up to us.
The potential risks of AI are listed many times. Killer robots and massive unemployment are common problems, while some people even fear extinction. More optimistic projections claim that AI will add $ 15 trillion to the global economy by 2030 and eventually lead us to some kind of social nirvana.
We certainly need to consider the impact such technologies have on our societies. One major issue is that AI systems reinforce existing social biases - to a devastating effect.
Several notorious examples of this phenomenon have received widespread attention: modern automated machine translation systems and image recognition systems.
These problems arise because such systems use mathematical models (such as neural networks) to define patterns in large training datasets. If this data is severely distorted in various ways, then inherent errors will inevitably be studied and reproduced by trained systems.
Biased autonomous technologies are problematic because they can potentially isolate groups such as women, ethnic minorities or the elderly, thereby exacerbating existing social imbalances.
If AI systems are trained, for example, from police arrest data, then any conscious or unconscious biases manifested in existing arrest schemes will be duplicated by the “police foresight” AI system trained from this data.
Promotional video:
Recognizing the serious implications of this, various reputable organizations have recently recommended that all artificial intelligence systems be trained on objective data. Ethical guidelines published earlier in 2019 by the European Commission suggested the following recommendation:
When data is collected, it may contain socially constructed errors, inaccuracies. This must be addressed before training the AI on any dataset.
This all sounds reasonable enough. Unfortunately, sometimes it is simply not possible to ensure the impartiality of certain datasets prior to training. A concrete example should clarify this.
All modern machine translation systems (such as Google Translate) learn from sentence pairs.
The Anglo-French system uses data that links English sentences ("she is tall") with equivalent French sentences ("elle est grande").
There could be 500 million such pairs in a given training dataset, and therefore only one billion individual sentences. All gender bias must be removed from this kind of dataset if we are to prevent results such as the following from being generated in the system:
The French translation was created using Google Translate on October 11, 2019 and is incorrect: "Ils" is a masculine plural in French and appears here despite the context clearly indicating that it is being referenced on women.
This is a classic example of an automated system preferring the default male standard due to bias in the training data.
Overall, 70 percent of the generic pronouns in the translation datasets are masculine and 30 percent are feminine. This is due to the fact that texts used for such purposes more often refer to men than women.
In order to avoid repeating the existing errors of the translation system, it would be necessary to exclude specific pairs of sentences from the data so that the masculine and feminine pronouns meet in a 50/50 ratio on both the English and French sides. This will prevent the system of assigning higher probabilities to masculine pronouns.
And even if the resulting subset of the data is fully gender balanced, it will still be skewed in various ways (eg, ethnic or age). In truth, it would be difficult to completely eliminate all of these errors.
If one person devotes just five seconds to reading each of the one billion sentences in AI training data, it will take 159 years to test them all - and that assumes the willingness to work all day and night, with no lunch breaks.
Alternative?
Therefore, it is unrealistic to require all training datasets to be unbiased before AI systems are built. Such high-level requirements usually assume that “AI” denotes a homogeneous cluster of mathematical models and algorithmic approaches.
In fact, different AI tasks require completely different types of systems. And fully underestimating this diversity masks the real problems associated with, say, highly distorted data. This is unfortunate as it means that other solutions to the data bias problem are neglected.
For example, biases in a trained machine translation system can be significantly reduced if the system is adapted after it has been trained on a large, inevitably biased dataset.
This can be done using a much smaller, less garbled dataset. Therefore, most of the data can be highly biased, but a trained system is not necessary. Unfortunately, these methods are rarely discussed by those who develop guidelines and legal frameworks for AI research.
If AI systems simply exacerbate existing social imbalances, then they are more likely to discourage than promote positive social change. If the AI technologies we increasingly use on a daily basis were far less biased than we are, they could help us recognize and confront our own lurking prejudices.
Of course, this is what we should strive for. Therefore, AI designers need to think much more carefully about the social implications of the systems they create, while those who write about AI need to understand more closely how AI systems are actually designed and built.
Because if we are really approaching either a technological idyll or an apocalypse, the former would be preferable.
Victoria Vetrova