AI Guides › Playbooks
By Nigel Guy · 7 min read
Most people form a view of a company's policy stance from a headline, a screenshot of a reply, or the fact that it did not sign something. That is how a large share of the internet came to believe Anthropic wants open-weight models banned. It feels like research because it involves reading, and it fails because nobody went to the company's own words.
The rule: before you repeat, cite or argue with anyone's stated position, read the primary statement, write down what it commits to and what it does not, and only then compare it with the claim you heard.
On 27 July 2026 Anthropic published "Our position on open-weights models", a post by CEO Dario Amodei (edited on 28 July to credit a research collaboration with AE Studio). It was a reply to a specific moment: reports that some US officials were weighing a ban on US companies using Chinese open-weight models, an industry letter backing open weights, and accusations that Anthropic wanted a ban to protect its business. CNBC reports that signatories to the letter included Nvidia, Microsoft, Meta and Palantir, that OpenAI signed later, and that Anthropic did not.
The post's headline sentence is in bold in the original: "Anthropic has never advocated for a ban on open-weights models." It also says: "we have not and are not advocating for a ban on open-weights models as a category."
It goes further than a denial. Amodei writes that open-weights models without dangerous capabilities are "a public good", and agrees with the letter that open weights expand access to the AI economy, strengthen competition at least for some use cases and give customers more control.
The post names two concerns, in order of priority. Keep the order; it matters.
He also disagrees with two assertions he attributes to the open letter: that open weights necessarily make safeguards easier to develop, and that broad access necessarily helps defenders more than attackers. His stated example is biology, where he suspects attackers hold a structural advantage. His position is that questions like this should be settled by pre-release testing, not assumed.
The post's remedy is three measures, which is the mechanism worth remembering. Anthropic says it has consistently advocated for all three.
| Lever | What it targets | What it is not |
|---|---|---|
| Keep powerful chips and chipmaking equipment out of China; crack down on smuggling | Concern 1 directly, concern 2 indirectly | A ban on any model |
| Crack down on industrial-scale distillation (building a model from a larger model's outputs, at scale) | Evasion of chip limits by a state-backed operation | A ban on open weights; the post says that is "neither the correct remedy nor something we have called for" |
| Mandatory safety testing for all sufficiently capable models, open and closed | Cyber, biological and alignment risk | A rule that open models fail; the post says whether open models carry extra risk "should emerge from testing" |
The post notes that effective testing would need to be global and says it thinks limited cooperation may be possible. Treat that as a hope, not a plan.
Say a colleague tells you, "Anthropic wants open models banned, so don't take their safety talk seriously." Run the check.
Use this with any assistant that can read a pasted document. Paste the primary text yourself so it is not working from memory.
Role: you are a careful policy reader. You summarise what a document says without adding to it.
Context: I will paste a primary statement from [ORGANISATION] dated [DATE], then a claim I have heard: [CLAIM_I_HEARD].
Goal: tell me whether the claim is supported, contradicted or not addressed by the statement, part by part.
Steps:
1. List what the statement explicitly supports and explicitly rejects, quoting each in no more than 15 words.
2. List important topics it does not address.
3. Split my claim into separate parts and score each as supported, contradicted or not addressed, citing the quote.
4. Write a two-sentence fair restatement of the organisation's position.
Rules: use only the pasted text. Do not use outside knowledge about the organisation. If the text is missing, truncated or looks like a summary rather than the original, say so and stop. If I have not given you a date or the claim, ask me for it.
Self-check before answering: is every score backed by a quote, and have you avoided guessing at motives?
Fill in the organisation, date, claim, then paste the statement under it.