A framework for scoring whether AI systems are trustworthy to the people they affect — before the failure shows up in your outcome metrics.
Working paper · under peer review at AI and Ethics (Springer Nature) · doi.org/10.5281/zenodo.20671478
If you have applied for a job in the last five years, an algorithm likely screened your résumé before a human saw it. If you have applied for a loan, a credit model scored you. Inside your organization, AI now drafts the analysis, ranks the candidates, and shapes the decisions. These are not distant technologies. They are the infrastructure of daily life — and what they share is opacity.
The evidence that opacity produces harm is no longer in dispute: résumé screeners that favoured white-associated names 85 percent of the time; facial-recognition error rates forty-three times higher for darker-skinned women than light-skinned men; a recruiting tool that taught itself to penalize the word “women’s.”
Inside organizations, the same gap has a different face: the pilot that stalled, the tool nobody uses, the shadow-AI workarounds your policy pretends don’t exist. When it happens, leaders reach for the same theory — we chose the wrong tool — and buy another one. In both worked cases in the research behind this guide, that theory was wrong. What was failing was not the tool. It was the space between the people who built the system and the people living with it.
Call that space the trust gap. It is measurable — and that changes what you can do about it.
NIST answers to the organization managing its risk. The EU AI Act answers to the regulator. ISO answers to the auditor. Each measures what its own audience needs — and none of them answers to the person living with the system. This framework is built to.
| Instrument | What it primarily asks | Who assesses | Primary focus |
|---|---|---|---|
| NIST AI RMF | Is the system’s risk managed across its lifecycle? | The organization managing risk | Risk process |
| EU AI Act + FRIA | Is a high-risk system lawful; does it respect rights? | The deployer, reporting to the regulator | Rights & contestability |
| ISO/IEC 42001 | Is there an auditable AI management system? | The auditor | Organizational process |
| Research instruments (participatory AI · CIRCLE · SCAF · IEEE 7000) | Involvement, deployment behaviour, societal resilience, design values | Design teams, stakeholders, society in aggregate | One facet each |
| This framework | Is it trustworthy to the people it affects? | Builder and adopter, scored independently | All five dimensions — capacity and power operationalized as scored criteria |
Few of these instruments operationalize, as scored criteria, whether a system builds or depletes the capacity of the people who use it, or which way economic power flows around it — who owns the data, who captures the value, who can leave. These were never oversights. They were simply never in scope.
The gap between what matters and what gets measured is not accidental. It is structural. This framework begins where that work leaves off.
Every criterion is scored twice — by the builder (what was actually designed into the system) and by the adopter (what happens when the system meets real people, processes, and context). The diagnostic power lives in the difference.
The Alignment Gap is the divergence between builder and adopter scores, read as the primary diagnostic output:
Because the gap is criterion-specific, it does more than register that something is wrong — it locates where, and implies the intervention. Builder scores epistemic integrity high, adopter scores it low? The system sends uncertainty signals the organization cannot read — a literacy failure, not a technology one. Capacity high/low? A tool designed to augment is being deployed to replace — a deployment failure. Power high/low? Portability exists but no one built an exit plan — a strategic failure. In every case, “we picked the wrong tool” is the wrong diagnosis, and a single averaged score would have hidden the real one.
The following cases are illustrative — constructed to show how the instrument reads a live deployment, not client engagements.
Warning zone. The model was signalling uncertainty; nobody could hear it. Fix: verification protocols + literacy, not a third vendor.
Clear governance failure. A co-pilot deployed as a replacement. Fix: restore the co-pilot role, redesign roles with the reps, fund capability investment.
A single score tells you a deployment is failing. Only the gap tells you why — and therefore what to change.
Pick one AI system you rely on — the tool at work, the model you prompt daily, the algorithm deciding what reaches you. Hold it against the five dimensions, twice.
Does it flag its own uncertainty, or state wrong answers with full confidence? Does it push back, or agree with you? Does it carry the biases of its training data into your context?
Do you catch it when it’s guessing, or take the output on faith? Are your people growing more capable — or more dependent? Who captures the value it produces, and who carries the risk when it fails?
The dimension where the two answers diverge most — that is where your trust is most at risk. That gap is the finding.
The framework supports three modes. A builder assessment surfaces the gap between intent and what was actually built. An adopter assessment scores organizational reality. An alignment assessment — both sides scoring independently, then comparing — is the most powerful mode, and the one the framework is built for. Scoring a single perspective takes under an hour with the rubric in hand.
Do this with one system this week — and do it with the people the system affects in the room. The conversation it forces is half the intervention.
→ Use the AI Assessment Canvas below to score a system by hand.
The hands-on worksheet. Print it, take one system, and score all fifteen criteria — Builder and Adopter, independently. Where the two diverge is the finding.
Where the two lines pull apart is the Alignment Gap — here widest on capacity orientation and accountability.
What you have here is enough to run the lens on one system and see where the gap opens. Where you go next depends on how deep you want to work.
Trust Infrastructure for AI — the full working paper: fifteen criteria, complete scoring rubric, evidence standard, two worked case studies. Under peer review at AI and Ethics (Springer Nature). → doi.org/10.5281/zenodo.20671478
From Hero to System — the collective-intelligence framework this instrument belongs to. → doi.org/10.5281/zenodo.21314259
Sonali Sharma writes from inside the 90 percent this work is for. Two decades building across continents and sectors — enterprise and leadership development across Asia, public-sector technology innovation strategy in Canada, a closed-loop impact business with Indigenous artisans in South India. She speaks on AI governance and mentors tech startups. Vancouver, BC.