Dto the analysis of the curricula of candidates for a job to health diagnostics: it is now evident and documented that Artificial intelligence systems can amplify gender inequalities. And if they do it It is not because they are bad But because they were trained on data – and bias – distorted. That is, on our distorted society.
The artificial intelligence trained to gender prejudices
Learning from data full of stereotypes, the Ai reflect and strengthen, in particular, gender prejudices. Prejudices that can limit the opportunities and diversity of women, especially in areas such as the decision -making process, hires, approval of loans and legal sentences.
Let’s imagine a car trained to make hiring decisions. Examples of the past are shown, and in most of these examples there are conscious or unaware prejudices. For example, men or scientists prefer men for managers or scientists. For roles as a nurse or secretaries, women.
It is only obvious that artificial intelligence interprets that men and women are more suitable for certain roles and make distorted decisions When filtering the candidates. Gender prejudice in artificial intelligence already has profound consequences in real life today, and not only in the recruiting sector.
A frame of the film Coded Bias by Joy Buolamwini on the risks of the unregulated use of artificial intelligence which, in his case, did not “recognize” the face of a black woman. In this photo, the facial recognition enables the access of a tenant to its accommodation. However, the anthropometric data are associated with various sensitive data (such as – penultimate line – the number of rental installments paid late).
From hiring to health care, prejudices are confirmed
“Vocal assistants that use female voices of default strengthen the stereotypes that women are suitable for service roles, and linguistic models such as GPT and Bert often associate you work as a “nurse” to women and “scientist” to men»,, explains (here) Zinnya del Villar, DIrermer of data, technology and innovation at Data-Pop Alliance. Again, “in critical sectors such as health care, artificial intelligence could focus more on male symptoms, leading to incorrect diagnosis or inadequate treatments for women ». The examples of gender prejudices in artificial intelligence are widely documented.
Strategies to reduce gender prejudices in artificial intelligence systems
How can we reduce gender prejudices in artificial intelligence systems? The debate is open, and fought.
Training on correct data
To reduce gender prejudices in artificial intelligence, the most important thing It is to train it using diversified data that represent all genres, ethnic groups and communities. “This means actively selecting data that reflect different social backgrounds, cultures and roles, while eliminating historical prejudices, such as those that associate specific works or characteristics with a genre”, continues of the villar.
The idea of Enter the correctives in the datasets, Assuming that they are default discriminators, he is promoted with conviction by the Council of Europe (COE). In a study published in 2023 by the Gender Equality Commission of COE, the need to adopt positive Actions, or transient tools for “positive discrimination” useful to compensate structural discrimination is supported.
The first solution is therefore trying to Clean the data before they are swallowed by the machine: balance the representativeness of the classes that the system must learn to recognize in the data set. And improve the transparency of algorithms.
Ex post filters
Another way of intervening, How (here) explains Professor Daniela Giordano, professor of artificial intelligence of the University of Cataniais to put ex post filters. «That is, a certain Large Language Model is used and then The result is filtered in retrospect. These filtering techniques can also be learned based on the indications of human being who have the task of select the best answers provided by the model, And they are necessary to ensure the alignment of the model to the values of the company and a minimum of safety. If you ask the system how you can build a bomb to create an attack, the system, for example, must not answer ».
Heterogeneous team upstream
Not only that: artificial intelligence systems should be created by heterogeneous development team, Composed of gender people, ethnicity and different cultural backgrounds. Del Villar explains: “This helps to integrate different perspectives in the process and reduces the blind points that can lead to distorted artificial intelligence systems”.
Implementing this strategy is not that easy, considering that, if the women who work in the STEM field are a minority, the gender gap in the AI sector It is even greater. Women constitute 22% of the workforce in the field of the AI in the world and only 16% in Italy.
As IA systems do to incorporate a feminist perspective if a female point of view is scarcely present among those who program them? International organizations such as UNESCO and UN promote initiatives to fill the technological gender gap: the Women4ethical AI UNESCO program, for example, aims to train experts and support governments and companies in guaranteeing equal female representation in the development of AI.
The education of people to detect distortions
Another fundamental point is then the awareness and education of the public. Help people a comprehend the functioning of artificial intelligence e the distortion potential It can allow them to recognize and prevent distorted systems and maintain human supervision on decision -making processes.
An ethical picture
Another fundamental point is the adoption of solid ethical paintings, integrating policies attentive to gender issues in the development of AI systems.
An artificial feminist intelligence? Can be done
Since the work is based on the analysis of pre-existing data, it is an intrinsically conservative technology. But it is not said that it cannot be used for social transformation projects, also in a feminist sense. Of course, because this happens, a proactive and anti -system effort is needed. And this from design but then in all stages, up to monitoring.
From the identification of the gender gap to Credit Scoring: what can do to the for gender equality
That artificial intelligence is not “the enemy” demonstrates the positive initiatives that involve it. For example, it has helped to analyze large quantities of data to identify gender salaries in the workforcewith tools like Glassdoor.
In finance, he is helping to overcome Long -date gender prejudices in Credit Scoring (i.e. the statistical method that allows to evaluate the credit reliability and solvency of a person). And it is also improving access to microfinance services for entrepreneurs, allowing them to access loans and financial services, in particular in disadvantaged areas.
Artificial intelligence has contributed to detect the disparity in the registration rates between men and women on platforms as coursera and edx. And he discovered prejudices in textbooks, helping teachers to review the educational materials to make them more inclusive.
Continue Villar. “In the future, artificial intelligence could help governments to evaluate the potential gender impact of proposed laws and help prevent discrimination and gender inequalities”. Not only that. “Algorithms based on artificial intelligence can also be used to make the safe digital space for everyone, detecting and removing harmful content It is discriminatory and blocking the spread of non -consensual intimate images, “adds Villar.