Share This Article
The recent Google’s Gemini scandal and its potential biases has thrust the importance of a rigorous Algorithmic Impact Assessment (AIA) for artificial intelligence (AI) systems into the spotlight.
Google’s Gemini Scandal
As reported by the press, when Google’s brand new Large Language Model, Gemini, began generating images of racially diverse Nazis, it not only showcased tastelessness but also historical inaccuracy, sparking a widespread internet outrage and a PR crisis of monumental proportions. Google’s Senior VP Prabhakar Raghavan’s admission of the system’s failure to discriminate appropriately in content generation highlighted a significant oversight in AI development.
The controversy escalated with further revelations, such as the chatbot’s refusal to compare Adolf Hitler with contemporary figures, underscoring the system’s biased and offensive responses. Google CEO Sundar Pichai’s response to the crisis, emphasizing the unacceptable nature of the outcomes and the company’s commitment to rectifying these issues, reflects the broader challenges facing AI development in terms of ethics and accuracy.
What Companies Willing to Exploit AI Shall Learn From the Scandal?
In an age where AI systems shape critical facets of our society, ensuring these technologies are deployed ethically and responsibly is paramount.
The AI Act obliges providers to train, validate and test data sets with data governance and management practices appropriate for the intended purpose of the high risk AI system. Those practices shall include, among others,
- examination in view of possible biases that are likely to affect the health and safety of persons, negatively impact fundamental rights or lead to discrimination prohibited under Union law, especially where data outputs influence inputs for future operations; and
- appropriate measures to detect, prevent and mitigate possible biases identified according to the process outlined above.
In this context, the performance of a Algorithmic Impact Assessment (AIA) becomes paramount to limit the risk of potential challenges. There is no “official” methodology to run an AIA, but the following might be a valid method:
1. Initiation:
- Start the AIA at the design phase of a project to understand the scope and nature of the automated decision system;
- Gather comprehensive information about the project, including the decision to be automated, the system’s design, and the data to be used.
2. Completion of the AIA Questionnaire:
- Answer a series of questions that assess various risk and mitigation factors related to the project. These questions cover areas such as project details, system capabilities, algorithm transparency, decision-making impact, data sourcing, and mitigation measures.
3. Scoring:
- Each response contributes to a risk or mitigation score, weighted according to the level of impact or risk mitigation it represents.
- The final score is calculated based on the responses, providing a measure of the project’s potential impact.
4. Impact Level Determination:
- Based on the scoring, determine the project’s impact level, ranging from Level I (little to no impact) to Level IV (very high impact).
- This step helps in understanding the severity and scope of the impact that the automated decision-making system might have.
Mitigation and Consultation:
5. Identify and implement mitigation measures to address the risks identified through the AIA:
- Consult with internal and external stakeholders, including privacy and legal advisors, digital policy teams, and subject matter experts.
6. Documentation and Transparency:
- Document the AIA process, findings, and decisions made based on the assessment.
7. Review and Update:
- Regularly review and update the AIA following any significant changes to the system functionality or scope of use.
Does this process sound complex to follow? That’s why we enriched our PRISCA AI Compliance tool with a module on the performance of an Algorithmic Impact Assessment to help organizations in performing such an intricated but essential analysis. You can watch a video of presentation on PRISCA AI Compliance HERE and reach an article on the AI Act and the changes introduced by it HERE.
