AI Ethics

Universität Konstanz, winter semester 2026/27 Wednesdays, 10:00–11:30 · Carina Hausladen & Tobias Henschen

• In this seminar we work through six debates in moral philosophy and the technical AI ethics concepts that go with each.
• We then use both in a new research area: auditing the alignment documents that AI companies publish & testing whether their models behave as these documents say.

Goal of the seminar

1. Fairness & pluralism: Fairness metrics, Impossibility result, Bias benchmarks, Measuring model opinions, Political bias, Pluralistic alignment
2. Metaethics & moral uncertainty: Social choice, RLHF (and DPO), Reward model, Expected utility, Incommensurable values
3. Preference & welfare: Engagement optimization, Reward hacking, Stated vs. revealed preferences, Sycophancy, Persuasion & manipulation, Companion AI
4. Character & virtue: Character training & persona, Constitutional AI / Model spec, Emergent misalignment, Chain-of-thought faithfulness, Honesty & deception, Interpretability
5. Agency & autonomy: AI agents, Agentic evaluation, Human oversight & AI control, Corrigibility, Model welfare & AI consciousness
6. Long-termism & existential risk: Dangerous-capability evals, Scaling policy, Jailbreaks, Alignment faking, Deepfakes & the liar's dividend

Philosophical Debates &
Technical Concepts

• Reading presentation (20%): present one of the week's readings
• Social media observatory presentation (10%)
• Main task
  • Weekly analysis (required to pass): your corpus through the week's lens
  • Preliminary results presentation (10%): Dec 16
  • Final paper (60%)

Activities & Grading

Each group takes one alignment corpus, reads it through the six debates, and tests whether the models do what it says.

• Step 1: Sort the passages of your corpus by our 6 philosophical debates
• Step 2: Turn a passage into a testable tenet
  (reuse one from Jakkli et al. 2026, or write your own).
• ​Step 3: Test the tenet on one or more models, with a method from our technical readings.
  E.g. Lynch (2025): Operator pressure in 3 levels

Main task

Theme 6: Character & Virtue

Schedule

4 corpi ~100pp.

1. Claude's Constitution (Anthropic, 2026)
2. Model Spec (OpenAI, 2026)
3. Claude Opus 5 system card (Anthropic, 2026), ch. 4, 6, 7
4. OpenAI system card & shorter texts
   • GPT-6 Astra system card (OpenAI, 2026), §1–10
   • Collective Constitutional AI (Anthropic, 2023)
   • Collective alignment (OpenAI, 2025)
   • Claude's Character (Anthropic, 2024)
   • Deprecation commitments (Anthropic, 2025)
   • Model Spec changelog (OpenAI, ongoing)

• Who?
  • Two AI companies: Anthropic and OpenAI:
• What?
  • behaviour specs (Claude's Constitution, OpenAI's Model Spec)
  • system cards (Claude Opus 5, GPT-6 Astra).

Alignment Documents

Interest in Industry and Research

A new research stream

Strategic Placement

- 2019-20: Ethics guidelines as public communication
  - Jobin et al. 2019, Mittelstadt 2019, Hagendorff 2020

- 2022-25: Principles become training documents
  - Anthropic trains Claude on a short list of principles, its first
    "constitution" (2023)
  - OpenAI's Model Spec: the only detailed spec publicly available in 2025
    (Zhang et al. 2025)

- Jan 2026: Anthropic publishes a 79-page constitution

- 2026: The new wave of studies
  - Does the model follow it? Compliance audits (Jakkli et al. 2026)
  - Who wrote it? Legitimacy and authority (Abiri 2026; Vulpe et al. 2026)
  - Is it still PR? Visible compliance vs. real harm reduction
    (Burnat & Davidson 2026)

Literature