Staking out a path to trustworthy AI

What exactly is “trustworthy artificial intelligence”? For now, there’s no clear, universally accepted definition, although the European Union’s AI Act goes one step in this direction with its seven principles of trustworthy AI: human agency and oversight; technical robustness and safety; privacy and data governance; transparency; diversity, non-discrimination and fairness; societal and environmental well-being; and accountability.

Applying these principles is much easier said than done, however, especially when biases come into play – whether ideological, political, religious, race- or gender-based, or more generally cognitive. “Biases are nothing new – as humans, we’re full of them,” says Olivier Crochat, the head of EPFL’s Center for Digital Trust (C4DT). “Generative AI should enable to bring this issue up, both for people who develop the programs and those who use them.”

“Biased output from a model could end up being used to train other models, thus compounding the bias,” he says. In other words, algorithms trained on biased data will merely reproduce the distortion. And because these algorithms are increasingly used for such things as making hiring decisions, reviewing mortgage applications and performing facial recognition, they could have a direct impact on peoples’ lives. The EU AI Act specifies that, to be truly trustworthy, a GenAI program must not only be transparent and secure, but also designed to detect and correct biases.

One concern is that algorithms trained on fake news, conspiracy theories, biases, propaganda and censored information will simply amplify that noise and become weapons of mass disinformation. “In theory, algorithms won’t make content any worse than it already is,” says Crochat, “however, if we don’t rectify this issue, it will carry on, and once amplified at large scale, it could have disastrous consequences.” In this regard, a recent report from the EPFL-based Initiative for Media Innovation (IMI) is encouraging. The IMI study examined various elections held in 2024 – a year when nearly half of the world’s citizens went to the polls. It found that AI-driven programs had only a marginal impact and didn’t swing the elections one way or the other. However, the study did find that the spread of manipulated content, boosted by algorithms, divided political opinion further and created a widespread climate of mistrust.

In theory, algorithms won’t make content any worse than it already is, however, if we don’t rectify this issue, it will carry on, and once amplified at large scale, it could have disastrous consequences.

Shared responsibility

The authors stress that the use of digitally manipulated content for propaganda purposes is nothing new – GenAI has only magnified this practice. “It’s a game of cat and mouse between the creators of AI technology to generate deepfakes and the developers of software to detect them,” says Prof. Touradj Ebrahimi, an EPFL expert in multimedia signal processing. His research group is working to develop systems for identifying and limiting the spread of manipulated content. C4DT director adds: “It still boils down to computer programming – if we know exactly what the problem is, we can find ways to solve it. And the upshot is that if we correct a bias in an AI model, it’ll instantly disappear for millions of users.”

However, that doesn’t make the issue of trustworthy AI any less important. Most GenAI programs are developed and hosted by profit-seeking businesses that shift some of the responsibility to users. This raises ethical and legal questions about the role that social values should play in AI governance.

“While ultimately the use of AI is the responsibility of users, regulatory mechanisms can – and should – be implemented to curb the risk of premeditated or accidental malicious use, much like the safety lock on guns or the safety cap on medicine bottles,” says Sabine Süsstrunk, head of EPFL’s Visual Representation Lab and president of the Swiss Science Council. “Proven strategies such as certification, regulation and education are needed to guarantee minimal acceptable performance, clarify responsibility and heighten public awareness.”

Johan Rochel, an EPFL lecturer on the ethics and law of AI and co-head of the Lab for Innovation Ethics (ethix), explains that responsibility isn’t a binary concept – it’s something that should be distributed along the entire value chain. Every decision that programmers make as they develop an AI system involves ethical choices and trade-offs that should be assessed well before their consequences become visible.

More isn’t always better

Another important issue relates to the huge datasets that are used to train GenAI algorithms. Many of these datasets come from businesses, with little transparency on where the data came from. Large tech companies often set aside ethical considerations when asked about intellectual property or how representative their datasets are, arguing that the more data they can feed into their algorithms, the better their AI programs will perform.

“But more isn’t always better – what matters is having quality data,” says Rochel. “Volume shouldn’t take priority over ethics.” Süsstrunk agrees: “If the data is copyrighted, it must either be licensed, purchased or not used at all. Unfortunately, given the inconsistent enforcement of copyright laws across jurisdictions, the current situation is far from satisfactory.”