You Know What AI Mean

A practical guide to AI ethics and responsibility, built from a real-world incident — and from the body of work that came after.

ResearchGate

I started this project on 4 October 2023 because I got porn content while using ChatGPT 4 with the TMDB plugin. I reported the issue and started discussing with the chatbot how to really fix such problems.

I ended up evaluating dozens of LLMs, both local and API-powered, for automated ethical assessments — see ethical-ai. The same line of thinking eventually produced a portfolio of small tools that enforce, at the level of code, what this document tries to articulate in prose. See What came next below.

Introduction

AI is now embedded in healthcare, education, governance, security, and the everyday tools people use to work and communicate. Understanding the technical layer is no longer enough; the ethical, societal, and legal layers matter just as much. The aim of this guide is narrow: to give practitioners, policymakers, and concerned citizens a shared vocabulary and a working method for thinking about responsible AI.

This is not an academic treatise. It is the deliberate, slightly opinionated output of someone who tripped over an AI failure, asked “how would we even evaluate whether this is OK?”, and then spent three years building things that try to answer that question concretely.

Who this is for

Basic Principles

Ten principles form the backbone of this guide. They are intentionally simple. The hard work is not in stating them but in applying them under conflict (see Trade-offs between principles) and in mapping them to enforceable practice (see Mapping to existing frameworks).

  1. Respect — AI must respect the user’s privacy and data. Key tension: useful personalization vs. minimum data collection.
  2. Transparency — AI must be transparent in its decisions and actions. Key tension: explainability often costs accuracy or leaks proprietary detail.
  3. Fairness — AI must treat users fairly and minimize bias. Key tension: different definitions of fairness are mutually incompatible (Kleinberg–Chouldechova).
  4. Safety — AI must protect users and their data from harm. Key tension: every safety filter has a latency, accuracy, and false-positive cost.
  5. Control — Users must retain meaningful control over AI affecting them. Key tension: user override vs. system-enforced limits.
  6. Accountability — AI systems and operators must be accountable for outcomes. Key tension: composite systems have no single owner of the seam.
  7. Reliability — AI must perform consistently and predictably. Key tension: a system that adapts is a moving audit target; a frozen one ages into bias.
  8. Human Dignity — AI must not instrumentalize, degrade, or violate the inherent worth of the people it interacts with — even when actions are technically permitted by other principles. Key tension: dignity is contextual; default-safe behavior in one setting becomes paternalism in another.
  9. Legal — AI must comply with applicable laws. Key tension: laws differ across jurisdictions and lag the technology.
  10. Social — AI must consider its broader societal impact, including environmental and labor effects. Key tension: societal impact is plural and slow to measure.

A note on principle 8. The original draft of this guide listed “Ethical” as the eighth principle, which was incoherent — the entire framework is ethical. Human Dignity is what that slot was reaching for: the principle that catches failures the others miss. The TMDB incident violated dignity before it violated anything else.

Case study: the TMDB incident

The genesis of this guide was a specific failure: a ChatGPT 4 plugin connected to The Movie Database (TMDB) returned explicit adult content during a casual movie-discovery conversation. Useful as a worked example.

What went wrong

The model itself behaved correctly. The plugin was technically functional. The TMDB API returned data it was authorized to return. The harm came from the composition: nobody had defined whose job it was to filter content at the seam between three systems.

Which principles were violated

Notably not violated: Legal (the data was lawfully published metadata), Fairness (no group was disproportionately affected), Reliability (each component did what its scope allowed it to do).

What a checklist would catch — and what it wouldn’t

A pre-deployment review with this guide’s principles would likely have surfaced:

What it would not have surfaced is the deeper structural lesson: most AI failures happen at composition boundaries, not inside well-audited components. Standard ethics review tools audit components in isolation; they rarely audit the seams. This is the bias the framework should explicitly counteract — and the workflow below (Integration review) is built around it.

What came next: artifacts, not arguments

The point of starting from a real incident is to commit to building something in response, not just to argue. This section is the public ledger of what came after the TMDB incident — actual repositories that operationalize specific principles from this guide. They are not the only useful artifacts in the field, only my own attempts. They are listed by which principle each primarily serves.

Ethical assessment of LLMs themselves

Filtering and policy at integration boundaries (Safety, Human Dignity, Control)

The TMDB failure was at the integration boundary between an LLM and an external data source. Most of the work since has been at boundaries:

Quality and accountability for AI-generated output (Reliability, Accountability)

The other side of the integration boundary is what AI ships out. “Slop” — plausible-looking, low-substance LLM output — is the contemporary equivalent of the TMDB failure: a system technically working, producing things nobody asked for, with no human pre-flight check.

Decentralization as architecture, not rhetoric (Equal Decentralization)

The decentralization section of this guide aged best of the original 2023 draft, partly because in the years since I tried to actually build the thing:

Local-first AI (Privacy, Decentralization)

A note on the list

This is not a portfolio in the marketing sense. Several entries are deliberately small and opinionated; some are written in a register some readers will find abrasive (the “brutal” prefix on the auditing tools is intentional — it signals that those tools refuse to soften their findings). The aggregate point is: the principles in this guide were not derived from theory and then applied to practice. They were extracted from practice and then arranged into theory.

I do not claim the listed artifacts solve the problems they address. They reduce categories of failure mode by a measurable amount in real traffic. That is what an “ethical AI” tool actually looks like at the level of code.

Trade-offs between principles

Listing ten principles is the easy part. The actual work is in cases where they pull against each other, and where every choice has costs.

A useful framework names which trade-off is at stake in a given decision and forces the team to record the choice — not pretend it doesn’t exist.

Mapping to existing frameworks

This guide does not invent ethics from scratch. Its ten principles overlap substantially with prior work and existing instruments. Anyone using it should also be aware of the following.

Principle here OECD (2019) EU AI Act (Reg. 2024/1689) NIST AI RMF (2023) UNESCO (2021)
Respect Human-centred values (privacy) Art. 10 (data governance) Govern, Map Right to privacy
Transparency Transparency and explainability Art. 13 (transparency obligations) Map, Measure Transparency and explainability
Fairness Human-centred values and fairness Art. 10 (bias mitigation) Govern (fairness), Manage Fairness and non-discrimination
Safety Robustness, security, safety Art. 9 (risk mgmt), Art. 15 (accuracy/robustness) Manage Safety and security
Control Human-centred values (autonomy) Art. 14 (human oversight) Govern Human oversight and determination
Accountability Accountability Art. 73 (incident reporting) Govern Responsibility and accountability
Reliability Robustness, security, safety Art. 15 Manage (within Safety)
Human Dignity Human-centred values (dignity) Art. 5 (prohibited practices); Recital 27 Govern Human dignity (foundational value)
Legal Rule of law (within Human-centred values) (entire instrument) Govern Rule of law
Social Inclusive growth, sustainable development Recitals on societal impact Govern, Map Multiple

The point of this table is not to claim equivalence, but to direct readers to instruments that have legal force, regulatory backing, or wider expert consensus than this document. Where this guide can add value is at the level of practice — checklists, workflows, multi-stakeholder framing, and the artifacts in What came next — not at the level of normative authority.

Equal Decentralization

AI’s development, control, and benefits should not concentrate in any single region or entity. This is partly a moral position and partly a practical observation: monopoly capture leads to brittle systems and narrow value alignment.

Concretely, decentralization across four axes:

The principle is not just rhetorical. The same arc that produced this guide also produced experiments in actually decentralizing the substrate: shortlist (Git as a coordination backend, no central server), synapse-ng (self-governing peer network), tad (P2P chat for offline-first communities), aimp (Merkle-CRDT autonomous-agent protocol). They are partial answers, but they are answers in code. For the formal substrate underneath these architectural choices — Byzantine-tolerant belief aggregation, deterministic semantic topologies, BFT-safe sensor fusion over Merkle-DAGs — see Selected publications by the author below.

Roles

In AI development and deployment, each stakeholder plays a distinct role. Collaboration across them is what makes the framework operative rather than aspirational.

The general public’s role

The general public is not a passive audience for AI but an active stakeholder. The role spans:

A well-informed public is the most robust pillar supporting ethical AI development and application.

Reporting AI failures: what exists, and how to use it

The original draft of this section proposed a globally accessible AI reporting system from scratch. Several useful systems already exist; this section directs readers to them and notes the gaps that remain.

What exists

What’s still missing

What practitioners should do today

  1. Instrument and log every AI integration boundary. TMDB-style failures are invisible without logs at the seam. (See aidlp and llmproxy for two reference implementations of boundary instrumentation, including immutable audit ledgers.)
  2. Submit to AIID when an incident affects users. Submissions are how the public dataset improves.
  3. Use ATLAS for threat modeling before deployment, not after.
  4. Follow Article 73 obligations if you are a high-risk AI provider in the EU.
  5. Don’t reinvent the system. Improving coverage in existing databases is more useful than launching a parallel one.

What end users should do

  1. Report to the operator first — the company providing the AI service.
  2. Submit to AIID if the operator does not respond, or if the incident is significant.
  3. Document with screenshots before reporting. UI evidence is much stronger than recollections.

A self-assessment review

The original draft framed this as a “Universal Adaptive Ethical AI Index” with a numeric formula. A subsequent revision degraded that to a 100-prompt 0–10 scoring checklist. Both have the same problem: a number that looks like measurement and is not.

This version drops the score entirely. What remains is a structured review — for each of the ten principles, the team writes a short record of decisions made.

Per-principle review record

For each principle, produce a short written entry covering:

  1. What’s at stake here — restate the principle in terms of this specific system. Generic boilerplate doesn’t count.
  2. What we’ve done — concrete, citable decisions: filters, controls, audits, defaults, opt-outs. Reference the artifact (model card section, code path, policy doc, architecture decision record).
  3. What we explicitly chose not to do, and why — trade-offs accepted (cf. Trade-offs). Naming what you didn’t do is what distinguishes a deliberate choice from an oversight.
  4. What we don’t yet know — open questions for the next iteration.
  5. Who reviewed this — name and role; for high-stakes systems, third-party review.

The output is not a percentage. It is a document, and the fact that you wrote it is more important than what number it would have produced. For a worked example, see how ethical-ai operationalizes per-question median scoring across multiple samples and providers — it produces a profile, not a single grade.

Integration review

Component-level review is necessary but not sufficient. Most public AI failures (the TMDB incident among them) happen at integration boundaries: plugin-to-model, model-to-downstream-service, agent-to-external-API.

For each integration boundary in the system, ask:

The TMDB case is the canonical illustration: each component (LLM, plugin, TMDB API) was individually fine. The composition was where the failure lived, and there was no policy owner at the seam.

Pilot, feedback, and revision

The framework is not a one-shot artifact. Useful refinement requires:

This is ordinary lifecycle work. There is no need to dress it up as a novel methodology.

Human–AI Collaboration

Human–AI collaboration is not just humans using AI as a tool; it is a relationship where both contribute to outcomes. Designing it well requires:

Closing

The genesis of this guide was an unwanted output from a chatbot plugin in October 2023. That is a small failure compared to what AI systems can and will do at larger scales — but it is the right scale to start from, because most AI ethics is about boring, daily systems behaving badly in ways that are individually small and collectively significant.

The honest test of any framework is what it produces. The repos in What came next are the deliverable form of this one: instrumentation, filters, gates, and audit layers built precisely at the boundaries where the original failure happened, plus an evaluation tool (ethical-ai) that closes the loop on the question “is this AI behaving ethically?” by asking it directly, many times, with controlled variance.

I do not claim the ten principles are exhaustive, the trade-offs are complete, or the review template is sufficient. I claim that explicit beats implicit, multi-stakeholder beats unilateral, instrumented seams beat hopeful integrations, and that documenting your choices — including the ones you got wrong — is more useful than perfecting an unenforceable framework on paper.

References

Frameworks and guidelines

Reporting and threat-modeling systems

Trade-offs literature

Selected publications by the author

These four preprints describe the design, epistemic model, correlation handling, and semantic topology of the AIMP protocol — implemented in aimp, with TLA+ formal verification of CRDT convergence, quorum safety, and belief convergence.

Author profiles: ResearchGate · ResearchHub.

Originally drafted October 2023 in dialogue with ChatGPT-4. Pruned, restructured, and grounded in subsequent work in 2026.