---
title: "How Ethical Stack scores AI: criteria and evidence"
url: https://ethicalstack.ai/methodology/
content_signal: search=yes, ai-input=yes, ai-train=no
---

# How the ratings work

Ratings are editorial judgments about documented practices, safeguards, and concerns. They are not measurements or certifications.

## How scoring works

Each finding names a product, model, service, company, or setup; ratings do not automatically apply to related layers.

We equally average available numerical scores across energy practices, training data, privacy, openness, worker protections, and military ties, then round to one decimal. Unrated dimensions are excluded. BDS status is categorical and never enters this six-dimension average.

A profile needs at least four of six ratings to receive an overall score. For eligible profiles, a boycott target caps the final overall at 2/10; a pressure target subtracts one point, with a floor of 1. “Not listed” and “Not checked” do not change the score. These adjustments do not change the underlying ratings or raw average. With fewer than four ratings, the profile remains Not ranked. Coverage shows how many dimensions have a number; adding evidence can change the average. Within each category, higher final scores rank first; ties go to the profile with more rated dimensions, then alphabetically by name.

Our criteria prioritize permission and control for creators, privacy for users, inspectable systems, better worker conditions, lower environmental impact, and limits on harmful military or surveillance uses. These are editorial priorities, not a universal definition of ethics.

8 to 10: stronger safeguards4 to 7.9: mixed1 to 3.9: limited safeguards or concerns

## Energy practices

[Sections](https://ethicalstack.ai/methodology/#method-sections)

This score reflects documented environmental practices and transparency. It does not measure the electricity or emissions of a request.

1 base point plus up to 3 points in each area, for a maximum of 10.

| Criterion | How it is assessed |
| --- | --- |
| Reporting
0–3 points | 0: no useful reporting. 1: a stated runtime or operating boundary. 2: dated quantitative reporting with scope. 3: an independently reviewed product footprint. |
| Electricity sourcing
0–3 points | 0: no relevant sourcing evidence. 1: partial or annual certificate claims. 2: quantified regional or site sourcing. 3: verified hourly and local carbon-free supply. |
| Resource controls
0–3 points | 0: no documented controls. 1: one practical control. 2: several documented controls or efficiency measures. 3: measured physical resource reductions alongside controls. |

Water is considered separately from this score. A renewable-electricity claim doesn’t show low water use or a measured footprint for your request.

## Training data

[Sections](https://ethicalstack.ai/methodology/#method-sections)

We assess evidence about permission for training material, including inherited base models and later training.

1 base point plus up to 3 points for source transparency, 4 for documented permission, and 2 for creator controls. The total is capped at 7 when documented permission scores below 3.

| Criterion | How it is assessed |
| --- | --- |
| Source transparency
0–3 points | 0: no useful source detail. 1: broad categories. 2: named datasets or partners, with gaps. 3: an inspectable manifest or recipe covering the main training stages and inherited base. |
| Documented permission
0–4 points | 0: no documented basis. 1: general policies or limited licences. 2: substantial licensed, public-domain, or consented material with an unresolved remainder. 3: predominantly inspectable permitted material with bounded exceptions. 4: documented rights for all training stages, independently validated for the assessed release. |
| Creator controls
0–2 points | 0: no documented training-source remedy. 1: a working opt-out or rights request, with limits. 2: documented permission before use or demonstrated removal or retraining through an inspectable process. |

Open weights, public web access, and a no-training-on-prompts policy are not proof of permission for the original training data. Synthetic material still needs a permitted source. Prompt retention and reuse belong under privacy.

For speech and music, we check speaker consent, voice-clone verification, recording and composition rights, performer compensation, and whether permission can be withdrawn. Permission to clone one voice does not clear the base model’s training data. These checks inform the existing ratings; they do not add another score. [Read the voice and music guide](https://ethicalstack.ai/guides/ai-voice-consent/).

For model-agnostic clients, training rights are **N/A** at the client layer; assess the connected model. Model-dependent services need a selected model. The [Data Provenance Initiative](https://www.dataprovenance.org/data-provenance-explorer) and [Fairly Trained requirements](https://www.fairlytrained.org/certifications) inform these criteria but do not endorse our scores.

## Your data & privacy

[Sections](https://ethicalstack.ai/methodology/#method-sections)

This rating describes who can access your prompts, files and recordings, how long they may be retained, and whether they may be reused for training.

Editorial rating within the bands below; there is no component formula.

| Criterion | How it is assessed |
| --- | --- |
| 1–3 | Weak control or broad default reuse. |
| 4–7 | Published controls with retention, routing, or plan exceptions. |
| 8–10 | Strong local control or narrowly scoped retention and training protections. |

A local model can keep inputs on-device when the full setup is local. Cloud fallback, telemetry, extensions and search can send data elsewhere. For recordings and voice agents, we also check participant consent, transcript access, recording retention and deletion controls.

## Openness

[Sections](https://ethicalstack.ai/methodology/#method-sections)

This rating considers whether the system can be inspected, modified, reproduced, and moved, including the restrictions on code or weights.

Editorial rating within the bands below; there is no component formula.

| Criterion | How it is assessed |
| --- | --- |
| 1–3 | Closed service or model with limited inspection. |
| 4–7 | Important public components or downloadable weights with restrictions. |
| 8–10 | Permissive code or weights and substantial reproducibility. |

A client SDK does not make a hosted model open. A model or code licence also says nothing by itself about permission for the training data.

## Worker protections

[Sections](https://ethicalstack.ai/methodology/#method-sections)

We grade documented protections for employees and contractors, including data workers, moderators and performers whose work supports AI.

1 base point plus up to 3 in each area (maximum 10). Policies earn less than demonstrated implementation; top points require independent worker-level evidence.

| Criterion | How it is assessed |
| --- | --- |
| Pay & conditions
0–3 points | 0: no relevant detail. 1: published commitments or limited employment protections. 2: concrete pay, hours, safety, or contractor protections with implementation evidence. 3: independently corroborated outcomes across the relevant workforce, including data workers where used. |
| Worker voice & remedy
0–3 points | 0: no relevant mechanism. 1: a conduct policy or reporting channel. 2: accessible grievance processes, protection from retaliation, and evidence of representation or follow-up. 3: independently corroborated worker voice and effective remedy across the relevant workforce. |
| Evidence & follow-through
0–3 points | 0: no useful checks or outcomes. 1: a monitoring commitment or limited reporting. 2: substantive audits, coverage, or corrective-action results with gaps. 3: independent worker-level checks and published, effective remediation. |

The scope includes employees, contractors, annotators, moderators, and relevant suppliers. Staff benefits don’t confirm supplier conditions; contributor codes don’t confirm wages. Serious reported concerns and company responses remain part of the finding.

The criteria draw on [ILO labour rights](https://www.ilo.org/topics-and-sectors/fundamental-principles-and-rights-work) and [Fairwork’s AI-work principles](https://fairwork.oii.ox.ac.uk/en/fw/principles/ai-principles/). Neither organization endorses or certifies this rubric.

## Military ties

[Sections](https://ethicalstack.ai/methodology/#method-sections)

Known military ties rank below entries where no military relationship is established, even when those entries disclose no safeguards. A company relationship doesn’t show use of every product or release.

With no established military ties, start at 7 and add 1 for every 3 safeguard points, up to 10. Civilian government work uses the same rule. With known defense ties, start at 1, add the safeguard points, then apply the relationship’s cap.

| Criterion | How it is assessed |
| --- | --- |
| Relationship caps | Operational weapons, targeting, or combat-intelligence support: at most 2/10. Broad defense or intelligence enablement: at most 3/10. Bounded noncombat defense work: at most 6/10. |
| Use restrictions
0–4 points | 0: no relevant binding restriction, or lawful-use rules alone. 1: weapons or violence restrictions with broad exceptions or coverage gaps. 2: weapons and harmful-surveillance restrictions with narrower gaps. 3: an explicit military or intelligence ban, or tightly bounded noncombat use. 4: comprehensive binding restrictions on weapons, harmful surveillance, and military intelligence without broad exceptions. |
| Enforcement & oversight
0–3 points | 0: no relevant enforcement evidence. 1: published suspension, monitoring, or abuse-review mechanisms. 2: military-specific review, oversight, and remedies. 3: independently checked enforcement outcomes. A general right to terminate service alone does not show effective safeguards. |
| Relationship disclosure
0–2 points | 0: no relevant disclosure. 1: an applicable policy or named government relationship. 2: substantive disclosure of relationship purposes and limits. |

A 7 reflects uncertainty without established defense ties, not a verified clean record. We distinguish contracts, procurement access, partnerships, suppliers, and actual use; availability or an exercise alone doesn’t confirm weapons or targeting use. Hosted policies may not bind copies of open code or weights. The criteria draw on the [ICRC’s explanation of autonomous-weapons risks](https://www.icrc.org/en/document/what-you-need-know-about-autonomous-weapons); the rubric is our own.

## BDS targets

[Sections](https://ethicalstack.ai/methodology/#method-sections)

This is a categorical check of named companies in the Palestinian BDS National Committee’s official campaigns. Nationality and religion are never proxies for company conduct.

Campaign category, not a numerical dimension; its eligible-overall adjustment is described above.

| Criterion | How it is assessed |
| --- | --- |
| Boycott target | Named as a consumer boycott priority in the [BNC consumer boycott and pressure guide](https://bdsmovement.net/Guide-to-BDS-Boycott). |
| Pressure target | Named for pressure action in the [BNC campaign guide](https://bdsmovement.net/Guide-to-BDS-Boycott) or [No Tech campaign](https://bdsmovement.net/no-tech-oppression-apartheid-or-genocide). |
| Not listed | No target match in the two official sources. |
| Not checked | No completed list check is available for the named company. |

The check covers named targets in those two campaigns, not every BDS campaign, local group, investment, or supplier. “Not listed” is neutral, not BDS approval. Parent-company matches and known commercial or cloud relationships are described separately.

## Evidence & coverage

Profiles link findings to sources and state the argument, confidence, and known unknowns. Evidence-based findings follow relevant documents; setup-dependent findings change with configuration; reasoned views are inferences. Confidence reflects how clearly the evidence supports the claim, not whether a company is good. Primary material can still have limits; profiles describe scope and independent verification where relevant.

**Depends on the model** means a model must be selected; **Not enough evidence** means no defensible grade; **N/A** means the field does not apply at that layer. These fields are excluded from averages and sort after numerical grades. Missing disclosure can lower a practices or safeguards score, but does not prove high consumption, infringement, worker abuse, or a hidden contract. The BDS check is categorical; its specified adjustments affect only the final eligible overall. The catalog is a sample of products and model families, not every release or AI category.