Measuring the ethics of an artificial intelligence system requires defining what is being evaluated, with what criteria, and according to what thresholds. Since the gradual entry into force of the European AI Act, the question is no longer an abstract philosophical debate. The transparency obligations of Article 50, applicable since August 2, 2026, impose machine-detectable labeling for synthetic content and deepfakes. The ethics of AI is now reflected in documentary evidence, audits, and operational controls.
Transparency obligations of the AI Act: what changes concretely since August 2026
The implementation timeline of the AI Act distinguishes several phases. The most recent one transforms ethical principles into verifiable constraints for regulators.
| Obligation | Scope | Implementation Date |
|---|---|---|
| Reporting interactions with AI (chatbots) | Any conversational system accessible to the public | August 2, 2026 |
| Labeling of synthetic content and deepfakes | Content generated or modified by AI | August 2, 2026 |
| Powers of control of the AI Office | Supervision of general-purpose models | August 2, 2026 |
| Codes of practice for GPAI models | Providers of general-purpose AI models | Finalization in progress |
This table highlights a shift: the ethics of AI is moving from a principle-based logic to a proof-based logic. Companies can no longer rely solely on internal charters. They must document their processes, make their systems auditable, and ensure the traceability of algorithmic decisions.
The resources available on botagora.fr allow us to observe how certain technological sectors are integrating these regulatory constraints into their development practices.
Algorithmic bias and training data: the deficit lies in implementation
Recent research identifies a paradox. Ethical principles applicable to artificial intelligence are abundant (fairness, transparency, explainability, respect for privacy). In contrast, the main deficit concerns the operational implementation of these principles, not their formulation.
Algorithmic biases illustrate this gap. A model trained on historical data reproduces the imbalances present in that data. The technical challenge lies in detecting and correcting these biases before deployment, not after.
- Auditing training datasets requires an analysis of the representativeness of the covered populations, which necessitates skills in statistics and social sciences, not just computer science
- Robustness testing must include stress scenarios on minority subgroups, where classification errors are often more frequent
- Documenting design choices (which algorithms, which decision thresholds, which data excluded) is a requirement of the AI Act for systems classified as high-risk
Producing a “fair” model is not just about applying a post-hoc filter. Fairness is built from the data collection phase, upstream of any learning.
Harvesting personal data: a regulatory issue in its own right
Training generative AI models relies on massive volumes of data, often collected through automated web scraping. This practice raises legal questions distinct from those posed by biases.
Scraping personal data is becoming a specific regulatory issue within the European framework. Responsibilities differ depending on whether the company collects data online itself or relies on pre-existing datasets provided by third parties.
The intersection of the GDPR and the AI Act creates a tension zone. The right to erasure guaranteed by the GDPR clashes with the technical difficulty of removing specific data from an already trained model. No technical solution fully resolves this issue to date.
Transparency of GPAI models
The codes of practice currently being finalized for general-purpose AI (GPAI) models aim to frame this gray area. Providers will need to document the data sources used, the filtering methods applied, and the measures taken to respect copyright.
Voluntary compliance with the codes of practice conditions regulatory compliance. A provider adhering to the codes benefits from a presumption of compliance with the obligations of the AI Act. Those who do not adhere will need to demonstrate by other means that they meet the same requirements.
Impact on employment and transformation of ethical skills in companies
The transformation related to artificial intelligence is not limited to directly automated positions. It also affects the skills expected in development teams. Designing an AI system compliant with the AI Act requires hybrid profiles capable of navigating between technical, legal, and human sciences.
- Diverse development teams (in terms of education, gender, background) produce systems whose biases are detected earlier in the design cycle
- Ethical governance of AI in companies cannot rely on an isolated advisory committee: it must be integrated into technical validation processes, just like security testing
- Ethics by design incorporates ethical constraints from the outset, not as an added layer afterward
The main risk for companies is not regulatory sanction. It is the loss of user trust in systems perceived as opaque or discriminatory.
The AI Act now structures a framework where the ethics of artificial intelligence can be measured, documented, and controlled. The transparency obligations applicable since August 2026 mark the transition from declarative ethics to auditable ethics. Companies that have not yet aligned their practices with these requirements have a limited timeframe before the next phases of implementation cover high-risk systems.
