Espresso Code: Italian Dev Insights

“How can I help you today?” – These are the words of one of the best-known Artificial Intelligences; in fact, few people have never read them and never tried asking it questions.

But what is an AI? What is it capable of doing? Is it impartial, being a machine?

From the first artificial neuron to Big Data

The first step in this field was taken more than 80 years ago, contrary to what one might imagine. It was a neurophysiologist and a logician, Warren McCulloch and Walter Pitts, who proposed the first model in 1943, built by trying to imitate a biological neuron. Since then, efforts have been made to refine the imitation, but the basic model has always remained the same.

But why is it only in recent years that artificial intelligence has captured so much media attention, on everyone’s lips, from governments trying to keep pace in order to regulate it, to companies looking for talent to be able to implement it, to ordinary people who use it widely instead of the more traditional search engines?

The reason why the time is ripe only now comes down to two factors: Big Data and the enormous computing power we now have thanks to technological progress.

Big Data is a vast set of data, which may or may not be structured, and it is only since the advent of the internet that people have begun to write enormous amounts of text and to generate an inconceivable mass of content such as images and videos: just consider that more than 500 hours of video are uploaded to YouTube every minute¹.

The human brain and artificial intelligence

Just as a child begins to speak by imitating their parents, who day after day, tirelessly, talk to them and care for them, so too is this tool the result of the “training data” it repeatedly comes into contact with.

But however useful this metaphor may be, we have not yet reached an artificial intelligence as sophisticated as the cognitive abilities of a human being. In fact, the activation functions, the neural structure, the efficiency of operations, the capacity to adapt and the plasticity of the whole are very far from being able to imitate a biological organism.

Biological synapses are much more than simple connections: they are complex structures involved in transmitting, processing and modifying signals. Their strength and behavior change dynamically, contributing to learning, memory and other complex brain functions².

The parameters of artificial intelligence are mathematical values within a model, adjusted during the learning process. Artificial intelligence models have difficulty reasoning and thinking critically. They are not able to form their own opinions, analyze situations from different points of view or learn from experience in the same way a human brain does. They can analyze enormous amounts of text and generate human-like text by identifying patterns and statistical relationships between words. However, they do not intrinsically understand the meaning of the words they use. In fact, an LLM (Large Language Model) analyzes a very large body of text to predict which word is most likely to follow a sequence of words already written. It considers several factors, such as co-occurrence, that is, how often two words (n-grams) appear together, their syntactic and grammatical role, etc. Although governed by sophisticated processes, it is a purely statistical choice: the model assigns a probability to each different word and usually chooses among the most probable ones according to the patterns it was trained on.

The quality of an AI is closely tied to that of its training data.

Applications and hallucinations

Neural networks can be applied in different sectors and can specialize in different functions:

• we have talked about text generation, but they can also generate pieces of music and images;
• they can predict an event and tell us how likely it is to occur, such as the efficacy of a drug or its side effects;
• they can analyze determining factors, extracting and generating value from data;
• they are capable of recognizing patterns, highlighting anomalies or understanding a consumer.

They have very broad, many-sided capabilities, but this does not make them immune to errors such as “hallucinations.”

We speak of hallucination when a model generates content that makes no sense or, worse, seems to make sense but is false and not supported by any logical or scientific basis. The more versatile an AI is and the more capable it is of handling different topics, the more insidious the hallucination³. Faithfulness hallucinations occur when the output generated is not consistent with the instructions or the context provided by the user, or contradicts itself. For example, a model asked to summarize a news article might generate a summary that seems realistic and well written, but that contains information absent from the article or that contradicts it.

Factuality hallucinations occur when the text generated is grammatically correct and consistent with the training data, but impossible or highly improbable in reality. For example, a model trained on a dataset of recipes might generate a new recipe that includes nonexistent ingredients or impossible cooking instructions.

One of the main reasons that can lead to hallucinations is the bias hidden in the training data, which favors certain behaviors and discriminates against others, producing a “biased output.”

An impartial algorithm?

There was a time, and perhaps for some it is still so, when, dreaming of an external and impartial system of judgment, people believed that an algorithm could make choices that were based on data and therefore objective⁴. But software cannot be immune to human influence and partiality, much less an Artificial Intelligence that learns from inequalities entrenched over the years. It is not able to tell good from evil: what is “right” and what is “wrong” is introduced through labels on the data, and the model manages to generalize this concept, becoming able to repeat it on unlabeled data too, without, however, understanding the reason. A simple “mathematical operation.”

Using automated systems to make decisions amplifies discrimination, instead of reducing or eliminating it, taking it to a larger scale.

Inequality: a limitation or an asset?

Although bias plays an important role in the process of discrimination, it does not necessarily cause it: bias is defined as a deviation from the standard, and it is sometimes necessary in order to identify the existence of statistical patterns in data or language. Classifying and finding differences would be impossible without bias⁵.

Indeed, the relationship between bias and discrimination is not always clear and understandable.

Many claim that by eliminating bias, discrimination can also be eliminated or reduced, but since implementations often ignore the context in which the algorithm will be used, the decision as to whether a biased output results in a case of discrimination is often left to the user to assess.

To explain this concept better, let’s think of a sophisticated algorithm that recommends which candidates are the best to hire after analyzing their résumés. We might notice that it favors candidates aged between 25 and 35, and this could be discrimination; but if instead the goal is precisely to hire younger candidates in order to have generational turnover within the company, then it would not be a case of discrimination by the algorithm, but an expected result.

Measures of bias and awareness

There are different ways of measuring bias; the statistical measure is the most intuitive and comprises three aspects. The first is that different categories should be able to appear in the results with the same probability: for example, if four out of five applicants from the advantaged group were granted a mortgage, the same proportion of applicants from the protected group should obtain the mortgage as well. The second aspect is satisfied if the unprotected and protected groups have an equal probability that a positive classification is correct, while the last aspect is that of well-calibration: the probability estimated by the algorithm must match the actual frequency of outcomes in each group⁶.

As already mentioned, this is the simplest measure, but other, more complex measures are used in order to have a model that limits inequalities and does not create discrimination, such as the measure of similarity between individuals called “fairness through awareness”⁷, according to which, for fairness to hold, the distance between the distributions of outcomes for individuals should be at most the distance between the two individuals as estimated by means of a similarity metric.

This shows us how much energy is spent in the field on analyzing a model with the aim of improving it as much as possible and minimizing discrimination. But without the awareness of users, every effort risks being in vain: we must use common sense and intelligence before applying an algorithm or a model to automate decisions on a large scale. Artificial intelligences are powerful tools and, as such, can lead to great benefits or great disparities. There is a growing need for honest, God-fearing people⁸, aware of the limits and risks of these tools and able to use them wisely.

Originally published in the dossier “Tempo del Creato 2024” of the GLAM commission of the FCEI (Federation of Protestant Churches in Italy)

1. How YouTube search works – YouTube Help (“over 500 hours are uploaded every minute”; figure reported by YouTube in 2019). See also Statista. ↩
2. Proverbs 20:5 “A person’s thoughts are like water in a deep well, but someone with insight can draw them out.” (GNT) ↩
3. L. Huang et al., A Survey on Hallucination in Large Language Models: Principles, Taxonomy, Challenges, and Open Questions, arXiv:2311.05232, 2023. ↩
4. Can an Algorithm Hire Better Than a Human? – The New York Times (nytimes.com) ↩
5. The definitions and examples in this section and the next are drawn from, and in places closely follow the wording of, X. Ferrer, T. van Nuenen, J. M. Such, M. Coté, N. Criado, Bias and Discrimination in AI: a cross-disciplinary perspective, arXiv:2008.07309, 2020. ↩
6. Criteria such as calibration within groups and balance for the positive and negative classes cannot, except in very special cases, all be satisfied at the same time: J. Kleinberg, S. Mullainathan, M. Raghavan, Inherent Trade-Offs in the Fair Determination of Risk Scores, 2016 (ITCS 2017). ↩
7. C. Dwork, M. Hardt, T. Pitassi, O. Reingold, R. Zemel, Fairness Through Awareness, ITCS 2012. ↩
8. Galatians 3:3 “How can you be so foolish! You began by God’s Spirit; do you now want to finish by your own power?” (GNT) ↩

Scripture quotations marked (GNT) are from the Good News Translation in Today’s English Version – Second Edition Copyright © 1992 by American Bible Society. Used by Permission.

GOOD NEWS TRANSLATION™ is a trademark of the American Bible Society and may be used only under license.