how to kris
lexical note hlt · law_ethics_AI

Part 1

Part 1

1. Make a list of all international guidelines

2. International treaty currently in force applicable to AI

The Council of Europe (CoE) Framework Convention on Artificial Intelligence and Human Rights, Democracy and the Rule of Law (CETS No. 225) is currently the primary legally binding international treaty applicable to AI.

Its main provisions

  • Risk-Based Approach: It mandates graduated measures based on the severity and probability of adverse impacts on human rights throughout the AI lifecycle.

  • Scope: It applies to AI systems used by public authorities and private sector actors, though with differentiated compliance methods for the latter.

  • Core Safeguards: It establishes requirements for transparency, oversight, accountability, and the protection of democratic processes.

  • Remedies: Signatories must provide effective legal remedies and procedural safeguards for individuals whose rights are violated by AI systems.

Advantages and disadvantages

Advantages

  • It is the first legally binding international treaty for AI, creating a global standard for governance.
  • The treaty is open to non-European states, increasing its potential for global harmonization.
  • It allows states to choose between direct obligations or equivalent protective measures for the private sector.

Disadvantages

  • As a framework treaty, it provides general principles rather than the detailed technical rules found in domestic acts like the EU AI Act.
  • Western-focused
  • The convention generally excludes AI systems related to national security, defense, and pure research and development.
  • Differentiated obligations for private actors may lead to inconsistent levels of protection across different jurisdictions.

3. Formulate an advice to the management

Hard law is binding; soft law is non-binding.

  • A treaty is a hard law.
  • Guidelines, principles, recommendations, or standards are soft law.

Treaties include a provision establishing a body to monitor the implementation of the treaty’s provisions. States report to the monitoring bodies on their compliance. In practice, these monitoring bodies often produce reports and responses that are helpful in implementing the hard law. This shows how soft (reports) and hard (treaties) can interact and complement each other.

Multilateral (environmental) agreements can take the form of hard law instruments as well as soft law instruments.

The analysis and delimitation of the two concepts (hard law and soft law) starts from the research of the sources of international law, in particular the content of Article 38 of the Statute of the International Court of Justice (ICJ).

Hard law

Hard laws have three necessary conditions:

  1. The existence of an obligation.
  2. Precision in presenting the content of the obligation.
  3. The existence of a body in order to hold the state which does not fulfill its obligation accountable.

Soft law

Those non-binding rules or instruments that

  1. provide an interpretation for the understanding of binding legal norms, or
  2. that represent behaviors recommended to states in their future conduct.

5 Arguments for hard law

The following arguments focus on how the EU AI Act can achieve a “Brussels Effect,” where EU standards become the de facto global standard:

  1. Market Size Influence: The EU’s large internal market acts as a “gravitational pull,” forcing global producers to adopt EU standards to retain market access.
  2. Soft law instruments are ineffective because they do not have binding legal force, an efficiency that is present in hard law instruments
  3. They allow states to engage more credibly in fulfilling the provisions of international agreements (which makes state commitments more credible).
  4. They produce legal effects in national jurisdictions, which provides a guarantee of the fulfillment of international commitments.
  5. They clarify the content of the rules by creating interpretation mechanisms.
  6. They give states the opportunity to monitor the fulfillment of their commitments, including through dispute resolution bodies.

Brussels Effect is two-fold:

  1. Products are likely going to comply with EU rules globally.
  2. Other jurisdictions will base their legislation on EU law.

5 Arguments for soft law

  1. Ambitious commitments are easier to achieve through soft law instruments than hard law instruments, but not because they eliminate the complex process of ratification, but due to the flexibility offered by soft law instruments.
  2. They create the framework of flexible negotiations.
  3. They impose lower “sovereignty costs”.
  4. They allow states to engage more easily in international cooperation than in the case of hard law instruments where they are bound by enforcement and the application of appropriate sanctions in the event of non-enforcement.
  5. They respond much better to the diversity of international society.
  6. States are more likely to accept/adopt a soft law because it’s not binding.

Part 2

  • do the existing legal provisions provisions applicable to LLMs (see the AI Act) provide sufficient legal protection considering the increased use of LLMs?

LLMs present unique challenges that the Act struggles to contain completely. First, the “black box” nature of generative AI makes it incredibly difficult to explain exactly how an LLM arrived at a specific output, complicating accountability.

Second, while the Act categorizes certain sectors as high-risk, it does not always explicitly mandate specialized demographic audits (such as strict gender-disaggregated testing) to catch subtle biases.

Finally, the sheer scale and accessibility of these models mean that enforcing these laws—especially against bad actors generating deepfakes or misinformation—remains a massive logistical hurdle.

  • can LLMs manipulate data? Explain

Generative AI can alter data to achieve desired statistical outcomes. For example, in an experimental scenario, an AI tightened the standard deviation of a fictional clinical dataset to artificially force a statistically significant result (a p-value of less than 0.05).

LLMs can use mathematical and macroeconomic reasoning to execute cascading manipulations. If instructed to falsely reduce job loss numbers in a dataset, the AI can independently recalculate and alter downstream metrics, such as unemployment rates and GDP projections, to match the fabricated baseline.

An LLM can manipulate human transcripts to entirely change the sentiment of the text—such as shifting a patient’s feedback from positive to negative—while perfectly preserving the speaker’s unique conversational style and tone.

Visual generative AI can seamlessly alter scientific imagery, such as fabricating a realistic-looking body of liquid water into a photograph of the Mars terrain.

  • can LLMs discriminate? Explain

Yes, LLMs can and do discriminate. This occurs because AI models are trained on massive datasets scraped from historical human text, which naturally contain centuries of societal prejudices, stereotypes, and systemic inequalities.

1. Representational Harm: The AI might default to gendered or racial stereotypes, such as consistently assuming that a doctor is male and a nurse is female, or translating neutral pronouns into biased professional role.

2. Allocative Harm: When integrated into real-world tools, this bias scales dangerously. If a company uses an LLM to screen resumes, the model might automatically downrank female candidates or minorities because the historical hiring data it was trained on favored white men. The AI identifies this past discrimination as a "successful pattern" and replicates it.
  • Give 3 examples of responsible use of LLMs. Give also 3 examples of irresponsible use of LLMs

Responsible Use of LLMs:

1. Data Processing & Structuring: Using an LLM to clean, format, and merge large datasets to make them more useful for analysis, provided the data is handled transparently and without altering factual outcomes.

2. Administrative Streamlining: Utilizing AI to draft routine emails, summarize lengthy documents, or organize meeting notes, with a human-in-the-loop actively reviewing the output before finalization.

3. Educational Accessibility: Deploying LLMs to translate complex materials into simpler language or multiple dialects, making learning and information more accessible to diverse populations.

Irresponsible Use of LLMs:

1. Fabricating Legal Evidence: Generating deepfakes or falsified documents to submit as authentic evidence in judicial proceedings, such as a recent California housing dispute where a judge dismissed a case after discovering the plaintiffs submitted an AI-generated deepfake video as witness testimony.

2. Scientific or Clinical Fraud: Intentionally utilizing AI to manipulate trial data so that an ineffective treatment appears statistically or clinically significant, which could severely jeopardize patient safety and scientific integrity.

3. Unchecked Automated Decision-Making: Deploying an LLM as a "black box" to make high-stakes decisions regarding human beings—such as hiring, loan approvals, or criminal sentencing—without conducting rigorous bias audits or requiring human verification.

4. Violating data privacy legislature

5. Unsupervised used of LLM output, for example using hallucinated cases