How the ratings work.
These are editorial judgments about specific safeguards and concerns. They are provisional, open to challenge, and supported by the evidence shown on each profile.
Read the score and the confidence together
BDS status (Boycott, Divestment and Sanctions): Boycott target · Pressure target · No named target found. Based on the BNC guide and No Tech campaign. Absence from these lists is not approval. What we check ↗
Scores run from 1 to 10, with higher numbers reflecting stronger safeguards or fewer documented concerns under the criteria below. Green is 8–10, yellow is 4 up to 8, and red is below 4. The decimal boundaries apply to overall averages; individual dimension ratings remain whole numbers. They are coarse judgments, not measured percentages or scientific precision.
We only assign a number when the evidence supports a judgment. Unrated means the reviewed material does not support one. It is shown with a dash and a short assessment label, excluded from sorting as a number, and never silently replaced with 5 or 10. Every field still has a specific argument grounded in the reviewed material. Evidence reviewed means a documented finding; Setup dependent means the answer changes with your configuration; Reasoned view means an explicit inference whose limits are stated.
High confidence means the specific finding is stated clearly in a relevant primary document. Medium confidence means the evidence has a meaningful scope or verification limit. Low confidence means substantial uncertainty remains. Confidence describes the finding, not whether a company is good.
The overall score is a quick starting point. Individual dimensions, the amount of evidence and BDS status can change which option fits your values.
How the overall score works
We take the equal-weight mean of the available numerical ratings across energy, training rights, privacy, openness, worker treatment and government/military work, then round to one decimal place. BDS status is categorical and stays outside this average. Unrated dimensions are excluded; we do not insert a guessed 5 or treat an unknown as a clean record.
For example, ratings of 8 and 6 produce 7.0/10 with coverage of 2/6. The other four dimensions remain unknown. Every overall score shows its coverage, and the profile lists the components used. All current overall scores are partial, so a high average from one or two ratings should be read cautiously.
Each category starts with the highest overall score. Equal displayed scores put the entry with more rated dimensions first, then use name A–Z. Rank numbers stay attached to this category ranking when you filter or change the sort. These are provisional editorial averages, not certifications or a verdict about unknown conduct.
BDS status is a separate list check
BDS stands for Boycott, Divestment and Sanctions. The Palestinian BDS National Committee (BNC) defines its own campaign targets. Our column records the named companies in its live consumer boycott/pressure guide and No Tech campaign, checked on 2026-10-23. Microsoft is a consumer boycott priority target; Google/Alphabet and Amazon are pressure targets. Meta is named in the No Tech campaign. Labels apply to the developer, service company or explicitly identified parent, with that scope shown on the profile.
No named target found does not mean BDS approved. BDS uses selective campaigns; these two lists cannot certify every company or supply chain. We show documented partnerships and cloud choices separately, including OpenAI’s Microsoft partnership, Anthropic’s Amazon relationship and routing options. They do not turn a partner into a separately named BNC target. This column uses no numerical score or green approval badge.
What we value
The criteria favor permission and control for creators, privacy for users, inspectable systems, better worker conditions, lower environmental impact, and restrictions on harmful military or surveillance uses. These are stated editorial priorities, not a universal definition of ethics.
Energy
1–3: documented high-impact growth with few controls. 4–7: bounded efficiency or electricity claims with material gaps. 8–10: strong, service-specific clean-power and impact evidence. A corporate renewable claim alone cannot earn 8.
Training rights
1–3: broad undisclosed or disputed corpora with limited creator control. 4–7: meaningful provenance, licensing or opt-out steps with unresolved rights. 8–10: inspectable, appropriately licensed or public-domain corpora with strong rights controls. Open weights alone earn no rights credit.
Privacy
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. Every rating names the configuration it assesses.
Openness
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.
Worker treatment
1–3: documented serious worker concerns. 4–7: supplier standards or bounded worker disclosures with unresolved outcomes. 8–10: strong worker-level, independently checked protections. A general supplier code does not establish annotation wages.
Government & military
1–3: documented defense deployment with material weapons or surveillance exposure. 4–7: a narrower relationship or explicit limits with gaps. 8–10: enforceable, independently checked restrictions against harmful uses. Ordinary civilian government work is not automatically a negative.
BDS status
BDS stands for Boycott, Divestment and Sanctions. We check the named targets in the Palestinian BDS National Committee’s consumer boycott/pressure guide and official No Tech campaign. Boycott target and Pressure target follow those campaign categories. No named target found is limited to those sources and the review date: it is not BDS approval. Parent-company matches and known commercial or cloud relationships are explained separately. This column uses categorical labels, not 1–10 scores. Nationality and religion are never proxies for company conduct.
Scope changes the answer
A model’s weights, its developer’s business, a local client and a hosted API are different things. Worker and military fields include the parent company’s documented record where relevant, explicitly marked Company. BDS status follows the named company or identified parent; other business context is separate. This helps readers consider who their purchase supports without claiming that a specific coding tool or model was used in a contract. We identify the assessed setup on each profile. A self-hostable model’s privacy rating describes local use; it does not apply to every API serving that model.
A renewable-electricity claim can be annual, company-wide or based on certificates. It is not automatically hourly clean power, low water use or a measured footprint for your request. A permissive model license also does not prove the training corpus was obtained with permission.
Government contracts are assessed by their stated purpose. We distinguish a commercial presence, a documented contract, and evidence of harmful use. A founder’s nationality or religion is never a proxy for any of them.
Research coverage
This edition covers 58 products or model families and 338 distinct source URLs. All 406 dimension fields include sources and a reasoned assessment. Of 348 fields in the six scored dimensions, 170 support numerical ratings and 178 do not. The remaining 58 fields show categorical BDS status (3 boycott targets, 7 pressure targets and 48 with no named target found). A populated research note is not proof that every underlying fact is public.
Most sources are first-party policies, model cards, licenses, reports or official announcements. BNC sources establish that movement’s campaign position, rather than independently verifying every allegation in its campaign copy. A statement that no relationship was established means only that this review did not establish it. It never proves the relationship does not exist.
Model names follow the latest release a researcher could establish from the sources examined on 2026-10-23. This is a maintained sample of major families and alternatives, not a claim to list every release or every AI category.
Download the full catalog and per-field sources. The page text, ratings and confidence labels are generated from that record.
Corrections and accountability
This is an evolving research project by Vijay Michalik. Send a correction through the links on the about page, including the profile, the disputed claim and the supporting source. Ratings are not certifications, and the research has not undergone an independent audit.