The ARPP launches Invenio (from the Latin “I find”), an artificial intelligence capable of detecting suspicions of breaches of rules applicable to digital advertising. The prototype is currently focusing on the detection of misleading therapeutic claims and the advertising of banned financial products in display format. It may be extended to other formats and other breaches.
This is the second prototype of ARPP based on AI, after the analysis of gender and age representation in audiovisual advertising. It was developed after workshops conducted with the advertising profession and animated at the School of Mines (Workshop No. 1 with the audiovisual advertising industry and Workshop No. 2 the digital advertising industry).
To help it automate the detection of digital advertising failures through artificial intelligence, ARPP chose Sicara, a French company specializing in image recognition, after the publication of a call for a declaration of interest.
The ARPP thus earns just over a third of the time. In the past, it was necessary to visit a site and identify possible breaches of regulations and ethical rules. Today, suspicions of failures automatically go up in a dashboard to save on less value-added tasks. This allows ARPP’s legal advisors to validate or invalidate the breach, contributing to the annotation of the data and improving the accuracy rates of the algorithmic models implemented.
The tool is based on a customizable list of URL (universal addresses) of websites, in order to collect advertising on the web pages of these sites or to alert the spread of campaigns pointing to the specified domain names. Image and text-based detection algorithms, based on pre-trained models, analyze content and trace suspicions of failures.
This information is then validated or invalidated, then qualified by the lawyers and experts of the ARPP, which allows the actors concerned (publishers, their boards and SSPs …) to be alerted to any shortcomings present in programmatic campaigns and to act accordingly.
Now in use, the device has shown its relevance and performance, with accuracy rates of 98% on text and 95% on images.