Artificial intelligence is a transformative tool for businesses, a catalyst for innovation in healthcare, finance and education. AI fascinates as much as it raises questions. Behind the apparent neutrality of algorithms, a fundamental issue emerges: that of bias. These statistical, cognitive or cultural distortions generate unfair, sometimes discriminatory decisions. The new European framework, the AI Act, intends to lay the foundations for more ethical AI. But how can companies concretely achieve compliance? And above all, how can they transform this regulatory constraint into a lever for responsible innovation?
A bias is never neutral
A bias in AI is not a simple calculation error. It’s a systemic imbalance that, if not controlled, can affect millions of decisions: hiring, credit granting, medical treatments. The white paper Bias in AI (March 2025), produced by Impact AI and Cercle InterL, offers a striking overview. Some recruitment algorithms exclude women, medical tools biased against minorities, delivery driver rating systems penalize certain religious practices.
Biases are everywhere: in data (historical, representative or not), in models (training on unbalanced corpora), and in uses (biased interactions between humans and machines). These failures can be statistical, methodological, cognitive or socio-cultural in nature. Ultimately, they reflect our own blind spots.
A legal framework in transformation: the AI Act enters the scene
Faced with this reality, the European Union has taken an unprecedented initiative: the AI Act, which came into force in August 2024. This regulation, which will be fully applicable in 2026, is based on a risk management logic. AI systems are classified according to four levels: unacceptable, high-risk, moderate-risk, or low-risk. The most sensitive ones — recruitment, health, education, justice, biometrics — will have to meet strict requirements of transparency, traceability, auditability and fairness.
This law does not merely impose general principles. It concretely engages the responsibility of companies, with sanctions that can reach 35 million euros or 7% of global turnover. A normative revolution on a continental scale.
Operational ethics: the corporate response
How can these obligations be transformed into concrete actions? Several pioneers, members of the Impact AI think tank, are already leading the way. Cercle InterL, which has been bringing together major industrial companies committed to equity for over 20 years, proposes a structured approach based on seven levers.
First, governance: establishing AI ethics committees, reporting to general management, with a clear mandate to prevent biases, particularly gender-based ones.
Then, responsible design: integrating non-discrimination principles from the design stage. This involves critical selection of data, rigorous documentation of models and continuous auditing.
Third, diversity in AI teams: an algorithm designed by a homogeneous group will reproduce its blind spots. Recruiting, training and promoting diverse profiles, particularly women, becomes a strategic necessity.
Finally, awareness: training all employees, but also the ecosystem (schools, partners, suppliers), in the culture of ethical AI.
Detect, evaluate, correct: three actions for fair AI
Bias management relies on a chain of vigilance. It first involves detecting differential effects between groups (age, gender, origin, socio-economic status). Then, understanding their causes: is it a statistical bias or a design bias? Finally, correcting by acting at different levels.
Three technical strategies coexist:
- Preprocessing: adjusting data before training (for example, balancing genders in a CV corpus);
- In-processing: integrating fairness constraints into the model itself;
- Post-processing: adjusting model outputs to make them more equitable.
But be careful: bias reduction must not harm performance or transparency. Trade-offs must be made between fairness, efficiency and explainability. This involves ethical as well as technical reflection.
The new challenges of generative AI
The arrival of large language models (LLMs), such as GPT-4, further complicates matters. These models, trained on massive corpora, are sensitive to the nature of the prompt, fine-tuning data, and alignment biases. They can amplify stereotypes, produce discriminatory content, or create behavioral influence effects.
Companies must therefore go beyond mere compliance. They must equip themselves with a governance framework specific to these models: enhanced documentation, traceability of adjustments, intrinsic and extrinsic bias evaluations.
Recommendations: an ethical roadmap for action
To support companies, the white paper proposes a five-step roadmap:
- Recognize biases and define the organization’s values;
- Map AI systems and their risk level;
- Prioritize legal and reputational risks;
- Implement documentation and audit procedures;
- Integrate bias management into CSR policy.
Public authorities also have a role to play. Teaching algorithm ethics from school, funding multidisciplinary research on biases. They must support SMEs through regulatory sandboxes, promote access to open source evaluation tools — these are the conditions for fair and effective regulation.
Bias in AI is not an anomaly: it’s a mirror of our collective choices, our historical data, our social structures. That’s why it’s not enough to correct it on a case-by-case basis: it must be framed, governed, thought through. The AI Act gives us a direction. It’s up to companies to seize it not as a constraint, but as a lever for transformation. And for civil society to remain vigilant. Managing biases means making AI a social intelligence.
References:
- White paper Bias in AI: What regulation requires and how enterprises can put it in action, Impact AI & Cercle InterL, March 2025.
- AI Act (European Regulation on Artificial Intelligence, 2024).
- OECD AI Tools Catalogue: https://oecd.ai/en/catalogue/tools
