We’re releasing Conscience, a Rust CLI that evaluates the ethical impact of AI-assisted software development. It examines your GitHub activity and AI tool logs, maps what it finds to a framework of ethical principles, and generates questions for your team to discuss. It’s a discernment tool — it surfaces patterns, but humans decide what they mean.

The gap it fills

AI coding tools track tokens, latency, and acceptance rates. Project managers track velocity and throughput. Nobody tracks whether AI is making your team’s work more human — whether developers are growing or being de-skilled, whether benefits are shared or concentrated, whether the environmental cost is proportionate to the value delivered.

These aren’t hypothetical questions. Two recent documents make the case that they’re urgent ones.

Pope Leo XIV’s encyclical Magnifica Humanitas (May 2025) frames AI through the lens of human dignity. It warns that AI can “de-skill workers and subject them to surveillance” (MH 150), that “no one is saved alone” (MH 73), and that AI is not morally neutral — its effects depend on how it’s deployed (MH 104). Its central question — “Does AI make human life on earth ‘more human’ in every aspect of that life?” (MH 129) — is the question Conscience tries to operationalize.

The Leiden Declaration on Artificial Intelligence and Mathematics (June 2026), written by 60 researchers across 10 countries, provides concrete recommendations: disclose AI tool use transparently, retain responsibility for correctness, affirm the humanity of authorship, and evaluate the ethical consequences of your work. Conscience maps its analysis directly to these recommendations.

What it does

Conscience runs a three-layer pipeline via conscience examine:

Layer 1: Signal detection. Automated pattern recognition across GitHub data and Claude Code session logs. It detects contribution concentration, low review engagement, fast merges without discussion, AI-to-human turn ratios, token consumption patterns, sensitive file access (credentials, SSH keys, .env files), suspicious shell commands, and sessions that run for 12+ hours.

Layer 2: Ethical scorecard. Signals are mapped to seven principles, each grounded in the reference documents:

  • Human Agency (MH 150) — are humans directing, or becoming dependent?
  • Equity of Benefit (MH 73, 77) — are AI tools benefiting all team members?
  • Transparency (Leiden O1) — is AI involvement disclosed?
  • Developer Growth (MH 52, 129) — are team members learning, or being de-skilled?
  • Environmental Cost (MH 101) — is AI usage proportionate to value?
  • Code Provenance (Leiden O4-O6) — do humans understand and own the code?
  • Security (MH 104, Leiden O4) — are tools used safely?

Where automated judgment would be irresponsible — Code Provenance and Developer Growth when data is sparse — the scorecard explicitly says “Needs human input” instead of guessing.

Layer 3: Reflection questions. Data-informed prompts for team retrospectives, populated with real numbers from your project. The closing question borrows the encyclical’s central metaphor: “Is this work building Jerusalem — shared responsibility, piece by piece? Or is it building Babel — impressive but concentrated?”

Teams maintain a conscience.yaml manifest in their project root — documenting mission, beneficiaries, cost bearers, and monthly self-assessments including an AI sentiment rating. This is the human half of the analysis that no automated tool can provide.

From Conscience to Vigil

If Conscience evaluates your development process, Vigil watches the AI tools while they run. Vigil’s AI Security engine was ported directly from Conscience’s signal detection system — the same patterns that Conscience surfaces in retrospective analysis, Vigil detects in real time on your Mac. Sensitive file access, suspicious shell commands, excessive autonomous operation, tokenmaxxing — Vigil flags these as they happen rather than after the fact.

Together, they represent two halves of the same idea: you should know what your AI tools are doing (Vigil), and you should regularly ask whether what they’re doing serves the people involved (Conscience).

Why this matters

The easiest response to AI ethics is to ignore it — ship fast, measure output, optimize for tokens per dollar. The second easiest is to outsource it to a compliance checklist. Conscience is built for teams that want a third option: structured, evidence-based reflection that takes seriously both the productivity gains and the human costs.

It practices what it preaches. The project was built with AI assistance, tracks its own metrics in its own conscience.yaml, and notes the irony openly. It doesn’t pretend to have the answers — it insists on asking the questions.