HITL Tracker Codebook v1

Version: 0.1 Published: 2026-04-18 Maintainer: David Felsmann Framework source: Center for Humane Technology, The AI Roadmap: How We Ensure AI Serves Humanity (2026)

Purpose

This codebook specifies how signals — discrete events in AI governance — are tagged against the Center for Humane Technology's seven principles, across three domains of change. The codebook is applied by a Gemini-powered tagging agent; 10% of auto-approved tags are sampled and human-audited each cycle.

This is an independent editorial project of Humane in the Loop, applying CHT's framework as a lens. It is not a CHT publication.

Structure

Each principle has three domains (Norms, Laws, Design). Each cell (principle × domain) contains 3–4 indicators — operational definitions of what can move that cell.

Every signal in the tracker is tagged to exactly one indicator. The signal's direction (+1 / 0 / −1) and magnitude (Major / Minor) are evaluated against that indicator's rules, not against the principle abstractly.

Cells with fewer than 5 signals over the relevant time window display "Insufficient data" rather than a direction.

---

Three Domains of Change

Domain	Code	Scope
Norms	N	Public discourse, expert consensus, institutional expectations, civil society pressure
Laws	L	Binding rules, regulations, court decisions, international agreements, enforcement actions
Design	D	How AI products and systems are actually built, deployed, and operated

---

Signal fields

Every signal carries:

• indicator_id (e.g., 1.L.d) — the specific indicator it updates
• direction ∈ {−1, 0, +1} — regressing / neutral / advancing against the indicator's rules
• magnitude ∈ {Minor, Major} — systemic weight
• direction_of_power ∈ {regulators, industry, public, unclear} — who gains leverage
• rationale (≤140 chars) — human-readable justification including base-rate context
• confidence ∈ [0, 1] — tagging agent's self-reported confidence
• triangulation_count — number of independent source types corroborating

---

Magnitude rubric

A signal is Major if it meets at least one of:

• Affects ≥100M people (population-weighted for policy actions)
• Sets legal precedent in a Tier-1 jurisdiction (US, EU, UK, Japan, G7 body)
• Originates from a frontier AI lab or top-5 global firm by compute
• Is a first-of-its-kind (novel legal/normative/design category)
• Has credible enforcement or implementation path within 12 months

Otherwise: Minor.

---

Direction rubric (general)

• +1 Advancing — the action moves governance in the direction CHT's "Better Future" describes for this principle. Consult the indicator's positive_direction_rule.
• 0 Neutral / Mixed — no net directional effect, or effects cut both ways with roughly equal weight.
• −1 Regressing — the action moves away from the "Better Future" vision. Consult the indicator's negative_direction_rule.

When the direction is genuinely ambiguous (common for research or industry announcements), direction is 0 and confidence should be below 0.85.

---

Principle 1 — AI should be built safely and transparently

1.N — Norms

1.N.a — Public expectation of transparency on AI capabilities

• Positive (+1): Rising public / media / expert demand for visibility into AI training, evaluation, and deployment.
• Negative (−1): Normalization of opacity; "AI is too complex to explain" framings gaining traction.
• Major: Coverage in ≥2 tier-1 outlets in the same week; major policy coalition letter.

1.N.b — Open publication of safety evaluations as industry norm

• Positive (+1): Labs publish eval suites, red-team findings, limitations openly.
• Negative (−1): Withdrawal from voluntary transparency (e.g., removed system cards, deleted benchmarks).

1.N.c — Incident disclosure as expected behavior

• Positive (+1): Near-misses, failures, jailbreaks disclosed proactively.
• Negative (−1): Incidents concealed, NDAs on safety findings, whistleblower retaliation.

1.L — Laws

1.L.a — Pre-deployment evaluation mandates

• Positive (+1): Binding requirements to evaluate frontier models before release (capability, misuse, alignment).
• Negative (−1): Mandates repealed, narrowed, or exemptions expanded.
• Major: Federal/EU/UK statute or binding regulation.

1.L.b — Mandatory incident reporting

• Positive (+1): Regulated actors required to report AI incidents to authorities within defined windows.
• Negative (−1): Reporting rules relaxed or rescinded.

1.L.c — Third-party audit requirements

• Positive (+1): Independent audits required for high-risk AI systems; right-of-access for auditors.
• Negative (−1): Audits made voluntary, weakened, or captured by industry.

1.L.d — Red-team disclosure rules

• Positive (+1): Binding red-team findings disclosure (to regulators, to public, or both).
• Negative (−1): Red-team results kept confidential or never required.

1.D — Design

1.D.a — Model cards / system cards published

• Positive (+1): New model release ships with substantive model card (training data summary, evals, limitations).
• Negative (−1): Release without card, or cards that are marketing documents lacking technical substance.

1.D.b — Public evaluation results

• Positive (+1): Labs publish benchmarks, red-team findings, dangerous-capability evals.
• Negative (−1): Results withheld, cherry-picked, or replaced with self-attestation.

1.D.c — Capability disclosure at point of interaction

• Positive (+1): Users told what model they are talking to, its limitations, confidence bounds.
• Negative (−1): AI systems deployed without disclosure, or disguised as humans.

---

Principle 2 — AI companies owe a duty of care to the public

2.N — Norms

2.N.a — Professional standards and codes of conduct

• Positive (+1): ML/AI engineering codes of ethics, professional licensing discussions gaining traction.
• Negative (−1): Industry pushes "move fast and break things" framing; dismissal of responsibility.

2.N.b — Public expectation of company accountability for AI harms

• Positive (+1): Mainstream expectation that AI companies are responsible for foreseeable downstream harms.
• Negative (−1): Normalization of "user is responsible" / "just a tool" deflection.

2.N.c — Researcher and whistleblower protections as norm

• Positive (+1): Strong norm that safety researchers inside labs can publish critical findings.
• Negative (−1): Chilling effects — non-disparagement NDAs, retaliation, equity clawbacks.

2.L — Laws

2.L.a — Liability for foreseeable AI harms

• Positive (+1): Courts or legislatures assign liability to AI developers/deployers for predictable damage.
• Negative (−1): Section-230-style shields extended to AI output.
• Major: Supreme court ruling, federal statute, EU directive.

2.L.b — Duty-of-care statutes applied to AI

• Positive (+1): Explicit duty-of-care obligations on AI developers/deployers.
• Negative (−1): Carve-outs from existing duty-of-care frameworks.

2.L.c — Consumer protection enforcement against deceptive AI claims

• Positive (+1): FTC / regulator enforcement against false AI claims, dark patterns, deceptive AI marketing.
• Negative (−1): Enforcement declined, agencies defunded, no action despite complaints.

2.L.d — Private right of action / class action enablement

• Positive (+1): Individuals can sue AI firms for harms; class actions viable.
• Negative (−1): Mandatory arbitration, pre-emption, standing doctrine shutting out claims.

2.D — Design

2.D.a — Safety-by-design practices

• Positive (+1): Red teaming, threat modeling, safety evals integrated pre-release.
• Negative (−1): Safety teams cut, evals skipped for shipping speed.

2.D.b — User reporting / abuse channels

• Positive (+1): Functional reporting channels acted on; published response times and resolution rates.
• Negative (−1): Reports ignored, channels removed, trust and safety teams dismantled.

2.D.c — Post-deployment monitoring and rapid response

• Positive (+1): Labs monitor deployed systems, respond to emergent harms within defined windows.
• Negative (−1): Ship and forget; no feedback loop from deployment to model updates.

---

Principle 3 — AI design should center human well-being

3.N — Norms

3.N.a — Public discourse on well-being metrics over engagement

• Positive (+1): Rising critique of engagement-maximization; mainstream coverage of time-well-spent framings.
• Negative (−1): Engagement metrics defended as proxy for value; critiques marginalized.

3.N.b — Critique of dark-pattern and addictive design

• Positive (+1): Civil society, academics, regulators naming specific harmful patterns.
• Negative (−1): Dark patterns rebranded as "UX optimization."

3.N.c — Mental health implications in mainstream discourse

• Positive (+1): Mainstream reporting on AI effects on attention, loneliness, cognitive development.
• Negative (−1): Topic marginalized as moral panic.

3.L — Laws

3.L.a — Restrictions on dark patterns and manipulative UX

• Positive (+1): Binding restrictions on specific dark patterns (consent traps, dark nudges).
• Negative (−1): Proposed rules dropped, enforcement rare, industry self-regulation accepted as sufficient.

3.L.b — Protections for minors

• Positive (+1): Age-appropriate design codes, default protections for under-18 users, content restrictions.
• Negative (−1): Minor-protection rules weakened, pre-empted, or challenged successfully.
• Major: Federal/EU/UK law, binding code.

3.L.c — Ad / recommendation system transparency

• Positive (+1): Regulation requiring recommendation-system auditability and user control.
• Negative (−1): Algorithmic opacity protected as trade secret.

3.D — Design

3.D.a — Opt-out and de-personalization defaults

• Positive (+1): Products ship with de-personalized defaults or easy toggles.
• Negative (−1): Personalization forced, opt-out buried or impossible.

3.D.b — Well-being features shipped by default

• Positive (+1): Time limits, break prompts, default quiet hours enabled out of the box.
• Negative (−1): Features exist but hidden, disabled by default, or removed.

3.D.c — Attention respect in UX

• Positive (+1): Fewer interruptive notifications; batch delivery; user-controlled attention.
• Negative (−1): Notification aggressiveness increases; attention engineered for maximum capture.

---

Principle 4 — AI should not automate away meaningful work and human dignity

4.N — Norms

4.N.a — Public discourse on AI and labor

• Positive (+1): Nuanced public conversation on displacement, augmentation, and worker power.
• Negative (−1): Binary "AI will replace X" rhetoric dominates without worker-voice balance.

4.N.b — Worker voice in AI deployment decisions

• Positive (+1): Union position statements, works councils, collective bargaining over AI deployment.
• Negative (−1): Workers systematically excluded from deployment decisions.

4.N.c — Public skepticism toward AI replacement rhetoric

• Positive (+1): Mainstream pushback against inevitabilist framings.
• Negative (−1): Replacement treated as inevitable, debate shut down.

4.L — Laws

4.L.a — Worker displacement protections and transition funding

• Positive (+1): Binding transition support, retraining obligations tied to AI deployment.
• Negative (−1): Displacement externalized onto workers; no safety net expansion.

4.L.b — Algorithmic management regulations

• Positive (+1): Rules limiting algorithmic supervision, scheduling, and discipline of workers.
• Negative (−1): Algorithmic management expands unregulated.
• Major: Federal/EU statute, landmark case.

4.L.c — Notification requirements before AI deployment in workplace

• Positive (+1): Employers required to notify and consult workers before deploying consequential AI.
• Negative (−1): No notification duty; workers learn of AI via effects.

4.L.d — Automated decision-making rights

• Positive (+1): Rights to explanation, human review, and contestation of consequential automated decisions.
• Negative (−1): Automated decisions opaque and uncontestable.

4.D — Design

4.D.a — Human-in-the-loop for consequential decisions

• Positive (+1): Product design keeps humans in authoritative roles for high-stakes decisions.
• Negative (−1): Full automation in high-stakes contexts without review.

4.D.b — Augmentation-over-replacement framing

• Positive (+1): Products positioned as augmenting human work, with measurable augmentation outcomes.
• Negative (−1): Products explicitly marketed as replacing human workers.

4.D.c — Transparent attribution of AI-generated work

• Positive (+1): AI-generated outputs clearly labeled; provenance tools deployed.
• Negative (−1): AI work passed off as human; labeling avoided.

---

Principle 5 — AI innovation should not come at the expense of rights and freedom

5.N — Norms

5.N.a — Public debate on AI surveillance and civil liberties

• Positive (+1): Journalism, civil society sustained pressure on surveillance applications.
• Negative (−1): Surveillance normalized as security or convenience.

5.N.b — Attention to algorithmic harms

• Positive (+1): Investigations, exposés on discriminatory outcomes, algorithmic bias.
• Negative (−1): Harms under-reported; "bias is solved" narrative gains.

5.N.c — Recognition of algorithmic discrimination

• Positive (+1): Mainstream acknowledgment that AI systems can systematically disadvantage groups.
• Negative (−1): Framing shifts to "AI is neutral, humans are biased."

5.L — Laws

5.L.a — Biometric and facial recognition limits

• Positive (+1): Bans or binding limits on law-enforcement or commercial use.
• Negative (−1): Expansions of use, rollbacks of moratoria.
• Major: Federal, EU AI Act, state-level bans.

5.L.b — Algorithmic bias and discrimination protections

• Positive (+1): Anti-discrimination laws explicitly cover algorithmic decision-making.
• Negative (−1): Algorithmic exceptions to existing anti-discrimination law.

5.L.c — Data protection strengthening

• Positive (+1): New / expanded privacy rights (access, portability, deletion, purpose limitation).
• Negative (−1): Weakening or pre-emption of state privacy laws.

5.L.d — Restrictions on predictive policing and algorithmic sentencing

• Positive (+1): Bans or oversight regimes on these systems.
• Negative (−1): Expansion without oversight.

5.D — Design

5.D.a — Privacy-preserving design defaults

• Positive (+1): Minimization defaults, local/on-device processing, differential privacy shipped.
• Negative (−1): Maximum data capture, centralized processing by default.

5.D.b — Bias testing and fairness tooling in development

• Positive (+1): Fairness audits as standard practice; published mitigation results.
• Negative (−1): Fairness work deprioritized or eliminated.

5.D.c — User rights surfaced in UX

• Positive (+1): Access, correction, deletion surfaced as first-class UI.
• Negative (−1): Rights exist on paper but buried or dark-patterned.

---

Principle 6 — AI should have internationally agreed-upon limits

6.N — Norms

6.N.a — International scientific consensus on risks

• Positive (+1): IPCC-for-AI-style bodies, published risk assessments, expert consensus statements.
• Negative (−1): Consensus processes stall; risk denial gains.

6.N.b — Multilateral civil society coalitions

• Positive (+1): Cross-border coalitions coordinating on AI governance demands.
• Negative (−1): Fragmentation, national retreat.

6.N.c — Public expectation of cross-border governance

• Positive (+1): Mainstream recognition that AI risks require international coordination.
• Negative (−1): Techno-nationalism dominant.

6.L — Laws

6.L.a — Multilateral treaties and conventions

• Positive (+1): New binding multilateral instruments; ratifications.
• Negative (−1): Withdrawals, stalled negotiations, unsigned protocols.
• Major: Council of Europe AI Convention, G7/G20 commitments with binding elements.

6.L.b — Export controls on frontier compute and models

• Positive (+1): Export controls targeting dangerous-capability compute, coordinated across jurisdictions.
• Negative (−1): Controls relaxed, loopholes exploited, uncoordinated.

6.L.c — Compute governance and licensing

• Positive (+1): Frontier-compute licensing regimes, thresholds codified.
• Negative (−1): Proposals shelved; opposition from industry successful.

6.L.d — Safety thresholds codified internationally

• Positive (+1): Shared red-line definitions (bioweapon uplift, cyber offense, autonomy thresholds).
• Negative (−1): No shared definitions; racing to ship past unstated thresholds.

6.D — Design

6.D.a — Voluntary industry safety commitments

• Positive (+1): Frontier labs publish specific, testable safety commitments.
• Negative (−1): Commitments withdrawn, watered down, or shown to be unmet.

6.D.b — Information sharing on dangerous capabilities

• Positive (+1): Labs share capability evaluations with peers and regulators via defined channels.
• Negative (−1): Competitive secrecy prevails; no structured sharing.

6.D.c — Evaluation protocols aligned internationally

• Positive (+1): Cross-lab and cross-border eval methodology converging.
• Negative (−1): Proliferation of idiosyncratic self-evals that aren't comparable.

---

Principle 7 — AI power should be balanced in society

7.N — Norms

7.N.a — Antitrust and competition discourse applied to AI

• Positive (+1): Mainstream conversation about AI market concentration and vertical integration risks.
• Negative (−1): National-champion framing dominates; concentration celebrated.

7.N.b — Open-source vs closed-source debate

• Positive (+1): Substantive public debate; neither side monopolizes framing.
• Negative (−1): Debate captured — either open-sourcing dismissed as dangerous, or closed AI dismissed as illegitimate, without serious engagement.

7.N.c — Concerns about democratic capture

• Positive (+1): Civil society surfaces AI-firm influence on policy; lobbying scrutinized.
• Negative (−1): Regulatory capture normalized; AI firms set the agenda unchallenged.

7.L — Laws

7.L.a — Antitrust action against AI market concentration

• Positive (+1): Merger challenges, structural remedies, conduct rules applied to AI firms.
• Negative (−1): Mergers waved through; investigations dropped.
• Major: FTC/DOJ/EU Commission action with teeth.

7.L.b — Interoperability and data portability mandates

• Positive (+1): Requirements for data portability, model interoperability, API access on fair terms.
• Negative (−1): Walled gardens protected; DMA-style rules weakened.

7.L.c — Public-option AI and sovereign compute funding

• Positive (+1): Public investment in open infrastructure, academic compute, non-commercial alternatives.
• Negative (−1): Public AI starved; private incumbents dominant.

7.L.d — Restrictions on political uses of AI

• Positive (+1): Rules on deepfakes in campaigns, AI-generated political ads, disinformation.
• Negative (−1): No rules; AI-manipulated political speech unregulated.

7.D — Design

7.D.a — Open-source model releases

• Positive (+1): Open-weights releases of frontier-class models, with responsible-release practices.
• Negative (−1): Frontier models universally closed; open releases withdrawn or restricted.

7.D.b — Decentralized and federated architectures

• Positive (+1): Federated training, on-device inference, decentralized serving growing.
• Negative (−1): Centralization accelerates; only incumbents can serve frontier models.

7.D.c — Third-party access to closed models

• Positive (+1): API parity for third-party developers vs internal teams; fair access terms.
• Negative (−1): Internal-first APIs, discriminatory pricing, sudden access revocation.

---

Versioning

Codebook changes are versioned. When an indicator's rule changes, all historical signals tagged against that indicator are re-evaluated by the tagging agent to maintain comparability across time.

• v1.0.0 (2026-04-18): Initial publication. 69 indicators across 7 principles × 3 domains.

Known limitations

• Single-coder (with LLM agent): Inter-rater reliability is not measured. A future upgrade path is to recruit 2–3 volunteer coders and compute Cohen's κ.
• Source universe is finite: Signals from sources outside the registered universe (see source-universe-v1.md) are not ingested.
• Framework lock-in: If CHT updates its AI Roadmap framework, this codebook will be re-versioned; historical signals will be re-tagged against the new indicator set.
• No China-domestic indicators: CHT's framework is premised on civil society and rule-of-law institutions that do not translate cleanly to authoritarian regimes. See methodology.md for the rationale.

Attribution

This codebook operationalizes the Center for Humane Technology's AI Roadmap. It is not authored, reviewed, or endorsed by CHT. Errors of interpretation are solely the editor's.