Open, accountable research on AI and the industries building it: built by people, powered by AI.

Dangerous Robot is a structured research project. It starts with a human question, is this claim true?, and routes it through a Python pipeline that uses LLM agents to find sources, fetch content, evaluate evidence, and draft a verdict. Every published claim aspires to be reviewed and approved by a human operator before it appears on the site.

The goal is to make it easier to verify what AI companies and products actually do, especially around environmental impact, safety practices, and transparency.

Research on this site focuses on:

• Claims made by or about AI companies and AI products regarding environmental impact, energy use, carbon footprint, sustainability practices, safety claims, transparency disclosures, and responsible AI practices.
• Any AI company or product where a verifiable claim can be evaluated, including new and smaller players. No minimum size or prominence required.
• Claims that can be evaluated against sources that are publicly accessible.

• Financial or investment advice about AI companies.
• Legal or regulatory interpretation.
• Non-AI technology companies. Coverage is limited to the AI sector.
• Claims about individuals rather than organizations.
• Claims that rely solely on non-public or paywalled sources with no accessible backup.
• General technology product reviews unrelated to environmental impact, energy use, sustainability, safety, or responsible AI practices.
• Medical or health claims unrelated to AI environmental or AI safety practices.

How a claim is published

Claims move through a pipeline before they appear on the site:

1. Research. An AI agent searches for sources relevant to the claim. A second agent fetches and stores those sources.
2. Analysis. An analyst agent evaluates the evidence and drafts a verdict with supporting reasoning. An auditor agent checks the draft for consistency with the sources.
3. Human review. An operator reviews the draft via the dr CLI and either approves it, rejects it, or requests a revision. Nothing is published without human sign-off.
4. Schema validation. All claims, sources, and entity data are validated by Zod schemas at build time. A claim that fails validation cannot be built into the site.

Verdicts

• True, accurate and well-supported by evidence.
• Mostly true, substantially accurate but with minor inaccuracies, missing context, or outdated details.
• Mixed, partially accurate; evidence supports some aspects but contradicts or fails to support others.
• Mostly false, contains a kernel of truth but is misleading in its overall characterization.
• False, contradicted by available evidence.
• Unverified, insufficient publicly available evidence to reach a verdict.
• N/A (not applicable), this criterion does not apply to this entity.

Confidence levels

• High, multiple independent primary sources confirm the verdict. Evidence is consistent with little meaningful uncertainty.
• Medium, evidence is present but incomplete. Sources may be partially secondary, limited in scope, or not fully consistent. Reasonable but not definitive.
• Low, sources are limited, conflicting, or carry significant uncertainty. Best available assessment; treat with caution.

Rechecks and "as of" dates

Each claim has a recheck cadence, typically 60 days (pricing: 14 to 30 days; policy: 90 to 180 days). When due, the pipeline re-researches and the operator reviews again before any change publishes. The "as of" date shows when evidence was last reviewed.

Limits

• Point-in-time. Verdicts do not update automatically. Check the "as of" date.
• AI-assisted, human-reviewed. Agents do the initial work. Operators approve. Human reviewers can be wrong.
• Coverage is not comprehensive. Gaps exist and are expected.
• No paid research. No commissions, no sponsored verdicts.

Six channels can put new research in the queue. The first four are operator-driven; the last two are public submission paths.

1. A new criterion. A reusable claim template is added to the template library. The system fans out a draft claim for every entity the criterion applies to.
2. A new company or product. An operator onboards an entity. The system fans out a draft claim across every active criterion.
3. A new source. An operator adds a citable reference. The system matches it to existing criteria (queuing new claim work) or to existing claims (queuing a reassessment).
4. A topic or URL drop. An operator adds free-form work to the queue file for triage and routing.
5. A public source submission. Anyone can propose a source via a GitHub issue. An operator reviews and ingests if accepted.
6. A public claim request. Planned, not yet live. Anyone will be able to propose a claim worth investigating; an operator triages.

None of these channels publish anything directly. Every claim goes through the methodology pipeline above before it appears on the site.

Sources are classified into three tiers:

• Primary, the organization's own output: company disclosures, official documentation, sustainability reports, regulatory filings.
• Secondary, independent reporting or research: journalism, academic papers, analyst reports, third-party certification records.
• Tertiary, advocacy, opinion, consumer guides, review aggregators.

Every source is linked and accessible from the claim page that cites it. Sources marked with an archived copy include a cached version in case the original is removed.

Research content is published under CC-BY-4.0. Site code is MIT licensed. Sources are cited inline on each claim page.

Dangerous Robot and TreadLightly AI share the same founder. Many claims here back assertions on the TreadLightly site, so it's worth reading with that in mind: what gets researched is shaped by that work.

No verdict is paid, sponsored, or commissioned. The operator holds no equity, debt, or paid advisory role with the AI companies covered here, and will note it on the relevant entity page if that changes.

If a verdict looks shaped by this, flag it on GitHub with the claim slug and your reasoning.