
Generative artificial intelligence is beginning to find its way into everyday applications ranging from virtual agents (or teams of virtual agents) that resolve queries when we call a customer service centre to intelligent assistants that automatically draft meeting summaries or report proposals in office environments.
These applications, often governed by foundational language models (LLMs), promise to revolutionise entire industries on the basis of huge productivity gains. However, their adoption brings new challenges because, unlike traditional software, a generative AI model does not follow fixed rules written by humans, but its responses are based on statistical patterns learned from processing large volumes of data. This makes its behaviour less predictable and more difficult to explain, and sometimes leads to unexpected results, errors that are difficult to foresee, or responses that do not always align with the original intentions of the system's creator.
Therefore, the validation of these applications from multiple perspectives such as ethics, security or consistency is essential to ensure confidence in the results of the systems we are creating in this new stage of digital transformation.
What needs to be validated in generative AI-based systems?
Validating generative AI-based systems means rigorously checking that they meet certain quality and accountability criteria before relying on them to solve sensitive tasks.
It is not only about verifying that they ‘work’, but also about making sure that they behave as expected, avoiding biases, protecting users, maintaining their stability over time, and complying with applicable ethical and legal standards. The need for comprehensive validation is a growing consensus among experts, researchers, regulators and industry: deploying AI reliably requires explicit standards, assessments and controls.
We summarize four key dimensions that need to be checked in generative AI-based systems to align their results with human expectations:
- Ethics and fairness: a model must respect basic ethical principles and avoid harming individuals or groups. This involves detecting and mitigating biases in their responses so as not to perpetuate stereotypes or discrimination. It also requires filtering toxic or offensive content that could harm users. Equity is assessed by ensuring that the system offers consistent treatment to different demographics, without unduly favouring or excluding anyone.
- Security and robustness: here we refer to both user safety (that the system does not generate dangerous recommendations or facilitate illicit activities) and technical robustness against errors and manipulations. A safe model must avoid instructions that lead, for example, to illegal behavior, reliably rejecting those requests. In addition, robustness means that the system can withstand adversarial attacks (such as requests designed to deceive you) and that it operates stably under different conditions.
- Consistency and reliability: Generative AI results must be consistent, consistent, and correct. In applications such as medical diagnosis or legal assistance, it is not enough for the answer to sound convincing; it must be true and accurate. For this reason, aspects such as the logical coherence of the answers, their relevance with respect to the question asked and the factual accuracy of the information are validated. Its stability over time is also checked (that in the face of two similar requests equivalent results are offered under the same conditions) and its resilience (that small changes in the input do not cause substantially different outputs).
- Transparency and explainability: To trust the decisions of an AI-based system, it is desirable to understand how and why it produces them. Transparency includes providing information about training data, known limitations, and model performance across different tests. Many companies are adopting the practice of publishing "model cards," which summarize how a system was designed and evaluated, including bias metrics, common errors, and recommended use cases. Explainability goes a step further and seeks to ensure that the model offers, when possible, understandable explanations of its results (for example, highlighting which data influenced a certain recommendation). Greater transparency and explainability increase accountability, allowing developers and third parties to audit the behavior of the system.
Open data: transparency and more diverse evidence
Properly validating AI models and systems, particularly in terms of fairness and robustness, requires representative and diverse datasets that reflect the reality of different populations and scenarios.
On the other hand, if only the companies that own a system have data to test it, we have to rely on their own internal evaluations. However, when open datasets and public testing standards exist, the community (universities, regulators, independent developers, etc.) can test the systems autonomously, thus functioning as an independent counterweight that serves the interests of society.
A concrete example was given by Meta (Facebook) when it released its Casual Conversations v2 dataset in 2023. It is an open dataset, obtained with informed consent, that collects videos from people from 7 countries (Brazil, India, Indonesia, Mexico, Vietnam, the Philippines and the USA), with 5,567 participants who provided attributes such as age, gender, language and skin tone.
Meta's objective with the publication was precisely to make it easier for researchers to evaluate the impartiality and robustness of AI systems in vision and voice recognition. By expanding the geographic provenance of the data beyond the U.S., this resource allows you to check if, for example, a facial recognition model works equally well with faces of different ethnicities, or if a voice assistant understands accents from different regions.
The diversity that open data brings also helps to uncover neglected areas in AI assessment. Researchers from Stanford's Human-Centered Artificial Intelligence (HAI) showed in the HELM (Holistic Evaluation of Language Models) project that many language models are not evaluated in minority dialects of English or in underrepresented languages, simply because there are no quality data in the most well-known benchmarks.
The community can identify these gaps and create new test sets to fill them (e.g., an open dataset of FAQs in Swahili to validate the behavior of a multilingual chatbot). In this sense, HELM has incorporated broader evaluations precisely thanks to the availability of open data, making it possible to measure not only the performance of the models in common tasks, but also their behavior in other linguistic, cultural and social contexts. This has contributed to making visible the current limitations of the models and to promoting the development of more inclusive and representative systems of the real world or models more adapted to the specific needs of local contexts, as is the case of the ALIA foundational model, developed in Spain.
In short, open data contributes to democratizing the ability to audit AI systems, preventing the power of validation from residing only in a few. They allow you to reduce costs and barriers as a small development team can test your model with open sets without having to invest great efforts in collecting their own data. This not only fosters innovation, but also ensures that local AI solutions from small businesses are also subject to rigorous validation standards.
The validation of applications based on generative AI is today an unquestionable necessity to ensure that these tools operate in tune with our values and expectations. It is not a trivial process, it requires new methodologies, innovative metrics and, above all, a culture of responsibility around AI. But the benefits are clear, a rigorously validated AI system will be more trustworthy, both for the individual user who, for example, interacts with a chatbot without fear of receiving a toxic response, and for society as a whole who can accept decisions based on these technologies knowing that they have been properly audited. And open data helps to cement this trust by fostering transparency, enriching evidence with diversity, and involving the entire community in the validation of AI systems.
Content prepared by Jose Luis Marín, Senior Consultant in Data, Strategy, Innovation & Digitalization. The contents and views reflected in this publication are the sole responsibility of the author.