With the EU AI Act entering into force, two of the most 𝐜𝐫𝐢𝐭𝐢𝐜𝐚𝐥 𝐫𝐞𝐪𝐮𝐢𝐫𝐞𝐦𝐞𝐧𝐭𝐬 for high-risk and general-purpose AI systems (GPAI) are 𝐄𝐱𝐩𝐥𝐚𝐢𝐧𝐚𝐛𝐢𝐥𝐢𝐭𝐲 and 𝐅𝐚𝐢𝐫𝐧𝐞𝐬𝐬. But current GenAI and LLM architectures are fundamentally at odds with these goals.
𝐀.- 𝐄𝐱𝐩𝐥𝐚𝐢𝐧𝐚𝐛𝐢𝐥𝐢𝐭𝐲 𝐛𝐚𝐫𝐫𝐢𝐞𝐫𝐬:
* 𝐎𝐩𝐚𝐪𝐮𝐞 𝐀𝐫𝐜𝐡𝐢𝐭𝐞𝐜𝐭𝐮𝐫𝐞𝐬: LLMs like GPT or LLaMA operate as high-dimensional black boxes—tracing a specific output to an input is non-trivial.
* 𝐏𝐨𝐬𝐭-𝐡𝐨𝐜 𝐈𝐧𝐭𝐞𝐫𝐩𝐫𝐞𝐭𝐚𝐛𝐢𝐥𝐢𝐭𝐲 𝐋𝐢𝐦𝐢𝐭𝐬: Tools like SHAP or LIME offer correlation, not causality—often falling short of legal standards.
* 𝐏𝐫𝐨𝐦𝐩𝐭 𝐒𝐞𝐧𝐬𝐢𝐭𝐢𝐯𝐢𝐭𝐲: Minor prompt tweaks yield different outputs, destabilizing reproducibility.
* 𝐄𝐦𝐞𝐫𝐠𝐞𝐧𝐭 𝐁𝐞𝐡𝐚𝐯𝐢𝐨𝐫𝐬: Unintended behaviors appear as models scale, making explanation and debugging unpredictable.
𝐁.- 𝐅𝐚𝐢𝐫𝐧𝐞𝐬𝐬 𝐁𝐚𝐫𝐫𝐢𝐞𝐫𝐬:
* 𝐓𝐫𝐚𝐢𝐧𝐢𝐧𝐠 𝐁𝐢𝐚𝐬: Models absorb societal bias from uncurated internet-scale data, amplifying discrimination risks.
* 𝐋𝐚𝐜𝐤 𝐨𝐟 𝐒𝐞𝐧𝐬𝐢𝐭𝐢𝐯𝐞 𝐀𝐭𝐭𝐫𝐢𝐛𝐮𝐭𝐞 𝐃𝐚𝐭𝐚: Limits proper disparate impact analysis and subgroup auditing.
* 𝐍𝐨 𝐆𝐫𝐨𝐮𝐧𝐝 𝐓𝐫𝐮𝐭𝐡 𝐟𝐨𝐫 𝐅𝐚𝐢𝐫𝐧𝐞𝐬𝐬: Open-ended outputs make “fairness” hard to define, let alone measure.
* 𝐁𝐢𝐚𝐬 𝐄𝐯𝐨𝐥𝐯𝐞𝐬: AI agents adapt post-deployment—biases can emerge over time, challenging longitudinal accountability.
𝐂.- 𝐂𝐫𝐨𝐬𝐬-𝐂𝐮𝐭𝐭𝐢𝐧𝐠 𝐃𝐢𝐥𝐞𝐦𝐦𝐚𝐬:
* Trade-offs exist between 𝐞𝐱𝐩𝐥𝐚𝐢𝐧𝐚𝐛𝐢𝐥𝐢𝐭𝐲 𝐚𝐧𝐝 𝐟𝐚𝐢𝐫𝐧𝐞𝐬𝐬—enhancing one can reduce the other.
* No standard benchmarks = fragmented compliance pathways.
* Stochastic outputs break reproducibility and traceability.
𝐖𝐢𝐭𝐡 𝐤𝐞𝐲 𝐭𝐫𝐚𝐧𝐬𝐩𝐚𝐫𝐞𝐧𝐜𝐲 𝐫𝐞𝐪𝐮𝐢𝐫𝐞𝐦𝐞𝐧𝐭𝐬 𝐛𝐞𝐜𝐨𝐦𝐢𝐧𝐠 𝐦𝐚𝐧𝐝𝐚𝐭𝐨𝐫𝐲 𝐬𝐭𝐚𝐫𝐭𝐢𝐧𝐠 𝐢𝐧 𝐀𝐮𝐠𝐮𝐬𝐭 𝟐𝟎𝟐𝟓, we urgently need:
• New model designs with interpretability-by-default,
• Scalable bias mitigation techniques,
• Robust, standardized toolkits and benchmarks.
As we shift from research to regulation, engineering 𝐭𝐫𝐮𝐬𝐭𝐰𝐨𝐫𝐭𝐡𝐲 𝐀𝐈 isn’t just ethical—it’s mandatory.