AI Ethics Manifesto
Ten lines that govern how OWI Labs models an organization: what the twin represents, what it is never allowed to become, and who stays accountable for the decision.
We model organizations, not people.
The twin represents structure: how work moves, where decisions actually get made, where ownership sits. Individual employees are the source of that picture, and no profile, score, or dossier is ever built for a named person. If a finding can only be stated about one individual, it does not enter the model.
Workforce intelligence is not surveillance.
OWI does not read keystrokes, email, calendars, or real time activity. What feeds the model is one anonymous 30 minute interview per person, in English or French, taken on their own schedule and with the purpose stated up front, and the answers are aggregated before anyone at the client organization sees them. Continuous monitoring of individuals is a different category of tool, and it is not what we build.
Transparent logic, not black-box authority.
Every pattern and every recommendation in a report traces back to a rule or a piece of evidence we can state in plain language. If a finding cannot be walked through step by step with the client reading it, it does not ship. A model nobody can interrogate is not intelligence, it is an opinion in a confident font.
Computational models are representations, not reality.
We hold the twin to one standard: the most accurate account of how the organization actually works that the evidence will support, built from what the people inside it said rather than from an org chart or a process diagram. A representation earns its authority by showing its work, not by claiming to be the thing itself. Where a reading is uncertain or the evidence behind it is thin, the report says so instead of rounding up to look more finished.
Human decisions require human accountability.
OWI's outputs inform decisions, they do not make them. No workflow we build lets an automated output trigger a consequential move on its own. A reorg, a role change, a headcount decision: a person reads the evidence, decides, and owns the call. The model can put the evidence on the table. It cannot be the one held responsible for what is done with it.
Bias is structural before it is algorithmic.
Before any pattern detection runs, we look at who was and was not in the conversation, and whether the questions were weighted toward one part of the organization. A model built on a skewed sample reproduces that skew no matter how careful the analysis downstream is, so which parts of the organization took part, and which did not, is reported in the deliverable rather than left as a footnote to it.
Data governance is non-negotiable.
We collect what an engagement needs and nothing more, encrypted in transit and at rest, kept only as long as the engagement requires, and reachable only by the people delivering it. Built to Loi 25, PIPEDA, and GDPR standards from the first conversation onward. Data gathered for one client is never used to train a model for another.
Simulation does not imply endorsement.
Modeling a scenario shows what that scenario would produce. It is not a recommendation to pursue it. An AI adoption path, a restructuring option, a merger sequence: we simulate options precisely so a leader can see which ones are bad ideas before committing to one. Showing where a path leads and arguing for it are different acts, and the report is written to keep them apart.
The asymmetry of power demands active responsibility.
The people whose day to day work we study rarely hold the leverage the leaders commissioning the study do. That imbalance means the duty to protect them does not stop at consent. It runs through aggregation thresholds, anonymization by design, and turning down engagements built to identify individuals rather than understand structure.
We submit our systems to scrutiny.
We treat our own methodology as something to be checked, not trusted by default. A client can ask how any specific finding was derived and get the derivation. Outside review of our anonymization and bias controls is welcomed rather than resisted. A manifesto that cannot be questioned is not an ethics practice, it is a marketing page, and this is meant to be the former.