YecoAI

YecoAI Complies with AI Act, But Challenges Its Technical Feasibility

YecoAI publishes its position on the EU AI Act: full compliance with transparency obligations, with technical concerns on high-risk evaluation requirements.

Marco Nasi

Founder & CEO

Formal Compliance, But Scientific Reality

The Scientific Evidence: Why Detection Fails

The False Positive Problem

The Fundamental Indistinguishability

The Watermarking Illusion

A Science-Based Proposal: Focus on Harm, Not Technology

1. Focus mandatory labeling where harm is concrete

2. Rely on voluntary transparency for text today

3. Establish technical standards with error thresholds

4. Create platform accountability for metadata preservation

A Constructive Contribution

YecoAI: Compliant and Constructive

References

Explore More

Today, YecoAI announces full compliance with the European AI Act. But we also believe it is fair to raise a technical point openly: for textual content, some obligations rest on capabilities that current science cannot reliably deliver.

We have implemented transparency systems in our platforms (DeskNexo and EnderDevelopment) that fully comply with the regulation. We provide clear disclosures to users and deactivation options.

However, I must draw a line between legal compliance and technical reality. The AI Act requires labeling AI-generated text, but current scientific literature demonstrates this is fundamentally unachievable with any legal certainty.

Peer-reviewed research from the University of Maryland (Sadasivan et al., 2023) found that AI-text detectors show markedly higher false-positive rates for non-native English writers. In a legal context, even a small false-positive rate risks accusing innocent authors.

The core finding is widely accepted: detector reliability drops in high-stakes settings, and error rates are unevenly distributed across user groups.

Multiple studies have shown that when an LLM is trained on human-written text, its output approaches human writing at the semantic level: for well-written texts, even expert human evaluators identify AI-generated content barely better than random guessing.

As linguistics researchers have long argued, a model trained to imitate human language produces text whose origin cannot be reliably determined after the fact.

A pivotal study by Sadasivan et al. (2024) from the University of Maryland showed that watermarking techniques can be removed through simple paraphrasing with 100% success rate while maintaining semantic meaning. The researchers concluded: "Watermarking provides a false sense of security. It can be defeated by any motivated adversary."

Furthermore, Zhao et al. (2023) demonstrated that watermarking can be removed automatically by prompting the same LLM to "rewrite more naturally," achieving watermark removal in over 95% of cases without human intervention.

For images and videos, the AI Act relies on metadata signatures. But this approach has a fatal flaw: social media platforms systematically strip metadata.

When an image is uploaded to major platforms, platform documentation and independent analyses consistently show:

Digital-forensics researchers, including Professor Hany Farid (UC Berkeley), have repeatedly observed that platform processing commonly strips the metadata needed to establish an image's origin.

When the first sanctions under the AI Act are challenged, judges will face a heavy burden of proof. EU administrative law expects evidence of a quality that detection technology, in its current state, cannot guarantee for text.

With current technology, this standard cannot be met:

"Sanctions based on unreliable technology risk being overturned on appeal, wasting judicial resources and eroding trust in EU institutions. We advocate evidence-based regulation, built with the technical community."

  • Marco Nasi, Founder YecoAI

Based on the scientific evidence, I propose a more targeted and enforceable approach:

The AI Act should require labeling only for images and videos that can damage reputation or defame individuals. This is where real harm occurs and where enforcement resources should focus.

For textual content, the scientific consensus is clear: reliable detection is fundamentally impossible. Mandating something that cannot be verified creates legal uncertainty and potential for abuse.

If detection is mandated, the EU must define acceptable error rates. A technology with a 26% false positive rate should not be admissible as legal evidence.

If metadata signatures are the enforcement mechanism, social platforms must be required to preserve them - not strip them automatically.

I urge the European Commission, Parliament, and Council to:

"We are a small company, but we work with this technology every day. We are sharing the scientific evidence in a constructive spirit, so that regulation and technical reality can converge before enforcement begins."

  • Marco Nasi, Founder YecoAI

YecoAI is an Italian startup developing AI solutions for the hosting and gaming sectors, with platforms like DeskNexo and EnderDevelopment already compliant with European regulations.

We will continue to comply with the law - while advocating for evidence-based reform. Responsible innovation means not just following rules, but improving them when they fall short of scientific reality.

The future of AI in Europe depends on regulation that is both protective and technically sound. We are committed to making that vision a reality.

Transparency note: this announcement was drafted with AI assistance and reviewed by our team, consistent with our AI-transparency commitments.

  • Instagram removes 98% of EXIF metadata for compression and privacy
  • X (Twitter) strips all metadata except basic color profile
  • LinkedIn removes geolocation and camera data by default
  • Facebook compresses images with up to 70% quality loss, destroying forensic markers
  • False positives: A 9-26% error rate means courts cannot convict without risking wrongful prosecution of innocent parties
  • Removed watermarks: If metadata is stripped by social platforms, defendants can argue the law punishes them for third-party actions beyond their control
  • Modified content: If an AI-generated image is edited in Photoshop or similar tools, forensic traces are lost, making provenance untraceable
  • No technical standard: The AI Act fails to define what detection method constitutes legal proof, leaving courts without a technical baseline
  • Consult domain experts: Include computer scientists and forensic specialists in the regulatory process, not just legal scholars
  • Wait for mature technology: Do not mandate what science cannot deliver
  • Focus on harm prevention: Target deepfake defamation, not AI assistance for writing
  • Avoid creating unenforceable laws: Laws without technical viability undermine the rule of law
  • Sadasivan, V. S., Kumar, A., Balasubramanian, S., & Wang, S. (2023). "Can AI-Generated Text be Reliably Detected?" University of Maryland.
  • Kirchenbauer, J. et al. (2023). "Reliability of New AI-Text Detectors: Report." University of Maryland.
  • Sadasivan et al. (2024). "Watermarking Paraphrase Attack." University of Maryland.
  • Zhao, X., Ananth, P., Li, L., & Wang, Y.-X. (2023). "Provable Robust Watermarking for AI-Generated Text."
  • Platform documentation on metadata handling for major social networks (Instagram, X, LinkedIn, Facebook).
  • Farid, H., public commentary on image provenance and metadata stripping, UC Berkeley.

Related content