The Last Friday marks a pivotal moment in Brussels, a beautiful city. On this day, the European Union's groundbreaking AI regulatory proposals will advance into their crucial final stage. This landmark legislation, a first of its kind globally, seeks to rigorously regulate artificial intelligence based on its potential for harm.

The process now moves into the decisive 'trilogies' phase, where the EU Parliament, Commission, and Council will collaboratively finalize the bill's contents, paving the way for its integration into EU law. This represents a significant and high-stakes step in the legislative journey.

The intention of the bill is excellent but whether the execution of the bill is good or bad for the community - let’s look into it deeply.

🦚 Snapshot

The European Union's Artificial Intelligence Act of 2023 is a pioneering law aimed at regulating AI technologies. Key points include:

• The Basics: 

  • Purpose: To ensure AI safety, transparency, and respect for fundamental rights.

  • Risk Classification: Prohibited AI » High-Risk AI »Limited Risk AI » Minimal risk AI

  • Governance: Establishment of national and EU-level bodies for oversight and enforcement.Compliance Grace Periods between 6-24 months.

  • Prohibitions: Ban high-risk AI practices, like manipulative behavior and mass surveillance.

  • Transparency: Obligation for certain AI systems to disclose their AI-generated nature.

  • Global Impact: Expected to influence worldwide AI regulation standards.

  • Legislative Process: Involves trilogue negotiations among key EU institutions.

  • Implementation: Likely adoption in early 2024, with full effect expected between late 2025 and early 2026.

  • Exemptions: National Security, Military and Defence, R&D, Open Source(Partial)

• Prohibited AI:

  • Social credit scoring systems.

  • Emotion recognition systems at work and in education.

  • Al used to exploit people's vulnerabilities (e.g., age, disability).

  • Behavioral manipulation and circumvention of free will.

  • Untargeted scraping of facial images for facial recognition.

  • Biometric categorization systems using sensitive characteristics.

  • Specific predictive policing applications.

  • Law enforcement use of real-time biometric identification in public (apart from in limited, pre-authorized situations).

• High-Risk AI: 

  • Medical devices, Vehicles.

  • Recruitment, HR, and worker management.

  • Education and vocational training.

  • Influencing elections and voters.

  • Access to services (e.g., insurance, banking, credit, benefits, etc.).

  • Critical infrastructure management (e.g., water, gas, electricity, etc.

  • Emotion recognition systems.

  • Biometric identification.

  • Law enforcement, border control, migration, and asylum.

  • Administration of justice.

  • Specific products and/or safety components of specific products

This Act marks a significant step in establishing legal frameworks for AI, balancing innovation with ethical and public welfare concerns.

Source:

• 1, 2, 3, 4

• EU Act Draft.

🐣 LLM(read it GPAI) and Flop Games

• The GPAI has implemented an automatic classification protocol for systems deemed to pose 'systemic risk', contingent upon their computational capacity exceeding 10^25 Floating Point Operations per Second (FLOPs). 

  • Presently, this classification primarily pertains to advanced models such as GPT-4 and Gemini. We are unsure about the GPT4 the parameters count and flop details - Assumptions are - They are using multiple smaller models. 

• However, within the forthcoming two-year period, it is anticipated that a diverse array of models, encompassing both proprietary and open-source platforms, will fall under this categorization.

• Commission can designate other GPAI as "systemic risk" based on other criteria such as several business users, end users, modalities, etc. But this designation is not automatic. GPAI with systemic risk have obligations for testing and evaluation requirements (self-assessment). External red-teaming and cyber security are required.

• The models are required to comply with copyright directives. They also need to fulfill obligations for non-systemic risk GPAI. What are those requirements? Not clear and transparent.

• Technical documentation (incl. model cards) made available if AI Office or national supervisory authorities come knocking. This includes information on the training data set, infrastructure, computational resources used, info on training, testing, etc. We already have the open-source data and open-source code details. The community is looking for a GPU cluster to play with the variations but not mentioned in the draft.

• They also need to make a similar document available to those downstream in the supply chain. In addition, a detailed summary of the content used to train GPAI (as generative AI) is required. This summary should not be too technical and should be understandable for people.

• The AI Office is expected to provide a template so that GPAI providers provide this detailed summary in an understandable format. This information can then be used by copyright holders to opt out as the EU copyright directive allows.

• The commission has to also review the energy efficiency of GPAI regularly and propose measures or actions. There is an open-source exemption. The Open-source GPAI models (not systems) would have to fulfill copyright-related requirements and if they are "systemic risk".

✨ Observations

• The EU needs to understand that people want an industry that works, and regulation can be required to make sure this exists. But nobody is clamoring for regulation in the absence of industry. The Chatgpt is a year old and the industry is trying to decide the best architecture. The community is not looking for the early RED TAPE.

• I live in the EU and have seen that a lot of paperwork and extra infrastructure is required to integrate GDPR. And it is still not mature enough.

• The act will be an extra responsibility to integrate GEN AI into the Fintech Industry. They need to take a couple of complicated regulations - It will create extra complications for the fintech industry.

• Other countries like UAE/ Japan have better corporate laws favorable to the LLM Industry (corporate laws, data transparency). Europe will use soon the Non-European LLM Model that will be trained on the European Data.

• AI regulation office in Europe - What is that?

• Ending the Blog with a post from the Guardian.

**

I will publish the next Edition on Thursday.

This is the 24th Edition, If you have any feedback please don’t hesitate to share it with me, And if you love my work, do share it with your colleagues.

It takes time to research and document it - Please be a paid subscriber and support my work.

Cheers!!

Raahul

**