Research ledger
Case by case, category by category.
This is the public workbench for keeping the index honest. A provider does not get a trust score because it feels friendly or scary. It gets ten category scores, with links.
Update rule
No hidden maths.
Every major lab gets the same ten categories. Each category is worth 10 points. The total must be the sum of those ten scores, and each changed score needs a dated source note on the lab page.
Disclaimer: the current model scoring was completed by GPT5.5 on Extra High on 3 June 2026. Treat it as an AI-assisted draft for public review, not procurement, legal, safety or investment advice.
The ten sections
What each category has to prove.
This is the repeatable checklist. Start with official sources, then use credible reporting, court documents or public statements where official pages leave out the hard bits.
| Category | First source to check | Hard question | Beginner warning sign |
|---|---|---|---|
| Beginner usefulness | Official product pages, onboarding and app availability | Can a normal Australian use it without technical setup? | Jargon, waitlists, region locks or unstable access |
| Cost clarity | Pricing, billing, refund and cancellation pages | Can a beginner predict the monthly bill? | Hidden credits, surprise overages or confusing plan names |
| Data and privacy | Privacy policy, training opt-out, retention and business-data docs | What happens to prompts, uploads, files, voice and chats? | Unclear human review, retention or training use |
| Terms and rights | Consumer terms, output rights, upload rights and media policies | Can the user safely publish or sell what they make? | Unclear rights for images, music, voice, likeness or uploads |
| Transparency | Model cards, system cards, safety notes, docs and incident pages | Can the claim be checked without marketing fog? | Thin docs, shifting claims or missing model boundaries |
| Safety record | Safety policy, abuse reporting, red-team notes and incident reporting | How well does it handle scams, children, consent and misuse? | Repeated misuse stories without clear controls |
| Leadership ethics | Governance pages, leadership changes, investment links and public controversy | Who benefits when this tool becomes powerful? | Weak governance, erratic leadership or conflicts of interest |
| Defence involvement | Government pages, defence contracts, partner channels and acceptable-use rules | Is it used for military, intelligence, policing, borders or surveillance? | Direct contracts, battlefield/intelligence deployment or weak public safeguards |
| Country and stability | Company origin, data jurisdiction, surveillance law and political context | Which country's laws and pressure points sit behind the tool? | High censorship, surveillance, instability or unclear data routing |
| Public benefit | Open releases, science work, accessibility, education and nonprofit records | What useful public good is visible and sourceable? | Little evidence beyond brand claims |
Defence first pass
The first case-by-case correction.
These four providers were updated first because the CDAO page names them together. That means they should not be treated as vague "maybe government" cases.
| Lab | Defence score | What changed | Sources |
|---|---|---|---|
| OpenAI | 2 / 10 | Direct CDAO/DoD pilot and OpenAI government page describing custom models for national-security customers. Usage-policy limits keep it above zero, but the contract heavily lowers the score. | CDAO | OpenAI Government | Usage policy |
| Anthropic | 3 / 10 | Direct CDAO award and declared defence/national-security deployments. Public refusal to remove mass-surveillance and autonomous-weapons safeguards earns a small distinction from OpenAI, not a clean pass. | CDAO | Anthropic statement | AUP |
| Google Gemini and DeepMind | 3 / 10 | Named in the CDAO frontier-AI awards, with existing government-cloud and Project Maven history already in the watch note. | CDAO | Google AI principles |
| xAI Grok | 2 / 10 | SpaceXAI government positioning plus CDAO award. The government page names defence, intelligence and operational-planning contexts. | SpaceXAI Government | CDAO |
Major lab queue
Who gets reviewed next.
The first published pass covers a broad map. The next research work should move across this queue and update one category at a time, so the site does not sneak in unsupported totals.
| Group | Labs or projects | Next category priority | Status |
|---|---|---|---|
| Frontier super apps | OpenAI, Anthropic, Google Gemini/DeepMind, Microsoft Copilot, xAI Grok | Defence, data/privacy, leadership | Defence first pass started |
| Open and enterprise LLMs | Meta Llama, Mistral, Cohere, IBM Granite, AI21/Reka, Hugging Face | Open licence, data, government use | Queued |
| Cloud and infrastructure | Amazon Nova/AWS Bedrock, NVIDIA Nemotron/Cosmos | Defence, robotics, cloud data controls | Queued |
| China model families | DeepSeek, Alibaba Qwen, Baidu/Tencent/ByteDance, Moonshot/MiniMax/Z.AI | Country, data, censorship, open weights | Queued |
| Creative media | Midjourney, Adobe Firefly, Runway, Stability AI, Black Forest Labs, Luma, Pika, Kling, Hailuo | Rights, training data, cost credits | Queued |
| Voice and music | ElevenLabs, Suno, Udio | Consent, copyright, voice cloning, safety | Queued |
| World builders | World Labs, Google Genie, NVIDIA Cosmos and related spatial systems | Rights, safety, physical-world deployment | Queued |
| Australian and local relevance | Canva, Leonardo.Ai, CSIRO Data61, Harrison.ai, Advanced Navigation, Emesent, DroneShield | Country, public benefit, defence where relevant | Queued |
| New frontier labs | Thinking Machines Lab, Perplexity and similar fast-moving entrants | Leadership, transparency, terms, data | Queued |