BBVA evaluates AI bias

Strongly committed to the deployment of the latest artificial intelligence technologies in banking services and processes, the Spanish bank BBVA partnered with IBM Research to test artificial intelligence, in particular, to measure the impact of biases and distorted results on discrimination.

FACTS

• BBVA has therefore partnered with IBM Research to share the expertise of researchers from both entities. The latter must therefore work to test the resistance of artificial intelligence to discriminatory biases based, for example, on possible stereotypes related to origins, sexuality or disabilities.

• Models of generative artificial intelligence are particularly concerned, as are models in languages other than English (English already benefiting from extensive work to combat discriminatory biases).

• BBVA has therefore adapted a data set called SocialStigmaQA (SSQA), developed by IBM's research team to test the resistance of general AI responses to conotic questions and prompts.

• BBVA teams have thus identified a greater presence of biased responses in Spanish, compared to the English model, thus demonstrating the value of further optimising the models.

• In addition to the researchers at BBVA and IBM Research, the work was also opened in Open Source to enable a broad community to work on model optimization.

ISSUES

• Promoting positive democratization : The deployment of the latest artificial intelligence technologies within the Bank's processes must not take into account the potential risks posed by these new technologies. However, these risks may arise from the extrapolation of discriminatory training data and the polarization of responses.

• Use and optimize : BBVA is here showing vision by working itself to improve AI, thus ensuring responsible use and deployment of new technologies. In this way, the bank underlines its commitment to establish a safe generative AI.

• To be a reference : The work of BBVA and IBM Research was accepted at the
  Neural Information Processing Systems (NeurIPS), lecture series on artificial intelligence. In addition to the data sets adapted to the Spanish language, BBVA is also working to deploy a data set specifically adapted to the banking sector.

PERSPECTIVE

• For the past year, BBVA has sought to mobilize its teams on the current issue of deployment of AI technologies in banking services. The bank's positioning includes: Organization internal challenges aimed at developing the expertise of its own employees on the subject.

• And this mobilisation has more recently led to the formalisation by the Spanish bank of Employment artificial intelligence technologies within these processes and its large-scale teams. BBVA offers ChatGPT Enterprise to its teams to facilitate their work.

Traduit automatiquement via Libretranslate / Automatically translated via Libretranslate