The wonderful world of intelligent algorithm bias control

The biases that algorithms can develop are known and numerous:

• Many are optimized to maximize profits above all.
• Others are fed data that is unrepresentative of the real population or neglect populations such as the disabled, single-parent families, etc.
• Finally, some are technologically limited while others take advantage of the most sophisticated advances.

While all recognize the need for regulation, the challenge of writing enforceable and effective regulations is on the table.

For example, the Digital Markets Act (DMA) legislation specifies:

"A marketplace shall not favor its services and products over those of sellers using its platform or exploit seller data to compete with them."

The Digital Services Act (DSA) legislation on the other hand states that:

"A digital service will have to explain how its recommendation and advertising algorithms work."

Easy to say, still need to be able to demonstrate or simply be able to do so. If the algorithm is constantly learning and changing, we need to find a way to test it in a representative way on the one hand, and on the other hand not to induce bias by the test itself and to be able to do it continuously and regularly.

This is what the Regalia pilot project tries to do. Benoit Rottembourg, project manager at Inria, takes stock after two years of activity. His recommendations are awaited with interest, especially as the development of legislation advances.

For the full article: Regulating algorithms: where does the Regalia pilot project stand?

Illustration: AndrewLozovyi - DepositPhotos

Visit inria.fr