A company publishes a post. Within hours there are a dozen articles about it. You cannot use the thing — access is gated, there is no specification, nothing is downloadable — so every word you read is somebody’s account of somebody else’s account.
This happens constantly, and the usual advice about it is useless. “Check your sources” is not a method. What you need is a small set of questions that separate an article that did work from an article that restated a press release well, and those questions are answerable in about ninety seconds per piece.
The launch of Anthropic’s Model Hardware Standard on 27 August 2026 is a good worked example, because it has every feature of the genre: a gated preview, no public specification, an impressive partner list, and a wave of same-day coverage. Below is the method, and what happened when it was applied to nine outlets covering that launch. Evaluating tech news is a repeatable skill, and the example is only there to make it concrete.
The Four Questions
1. Does this article quote anyone the company’s own post did not already quote?
This is the single highest-signal check, and it takes seconds. Company announcements come with pre-approved quotes from named partners. If every quoted person in an article also appears in the original post, the reporter did not talk to anyone — they read the same page you can.
2. Did the writer load the primary source, or a summary of it?
Look for signs of first-hand contact with the actual material: a direct quote of an odd phrase, a note about what the page does not say, a correction of a detail circulating elsewhere. Articles built from a summary tend to reproduce the summary’s emphasis, including its errors.
3. Does it connect the announcement to something outside the announcement?
The most valuable reporting on a launch is usually not about the launch. It is a regulation the thing will run into, a prior standard that already tried this, a dated precedent from the same company. That connection cannot come from the press release, so its presence is proof of independent work.
4. What is it careful not to claim?
Good coverage of an unavailable product is conspicuously hedged in specific places — “named as a partner” rather than “is using,” “the company says” rather than a flat assertion. Vagueness everywhere is padding; precision about where the uncertainty sits is a sign someone thought about it.
What Those Questions Found
Applying them to coverage of the MHS launch: of the nine outlets identified, seven could actually be loaded and assessed. Of those seven, exactly two did something a press release could not have produced on its own.
The Quantum Insider obtained original interview quotes from QuEra’s own executives — people not quoted in Anthropic’s post. That is question 1, answered yes.
The Next Web connected the MHS safety-tag reference file to a specific, dated EU regulation that neither Anthropic’s post nor any other outlet in the sample mentioned: Regulation (EU) 2023/1230 on machinery, of 14 June 2023, which repeals Directive 2006/42/EC and whose recitals address software performing safety functions and machine-learning systems with self-evolving behaviour. That is question 3, answered yes, and it is the most useful single contribution in the whole sample — because it points at a constraint the technology will actually meet.
The rest, including CNBC, largely reproduced Anthropic’s framing, partner list and numbers. CNBC did carry one genuinely original quote, from Anthropic’s head of beneficial deployments, Elizabeth Kelly: “We built this for science to sort of show the promise of AI, but there’s also huge benefits here for enterprise and for industry.” (Worth flagging honestly: a direct fetch of that CNBC page returned HTTP 403 during this research, and the quote was confirmed from a corroborating search-indexed excerpt rather than by reading the full page. That is a weaker form of verification, and saying so is part of the method.)
This is not a scandal, and reading it as one misses the point. It is what fast coverage of an embargoed announcement looks like everywhere, for structural reasons — outlets receive material under embargo, publish at the same minute, and the ones that add reporting are the ones that had a relevant contact or a relevant obsession already. Naming which outlet did which is more useful than a generic warning, because it teaches you what the difference actually looks like on the page.
Two of the nine could not be located or loaded at all. The correct handling of those is to attribute nothing to them — not a claim, not a framing, not a quote. An outlet you could not read is not evidence, and treating “I saw it referenced somewhere” as sourcing is how errors propagate.
The Failure Mode This Method Catches
Here is the most instructive thing that happened, and it is instructive because the people doing the research got it wrong first.
A quote attributed to an Anthropic staff member — to the effect that what MCP did for software, MHS will do for the hardware world — circulated widely alongside the launch. It is a tidy, quotable line. It is exactly the sort of sentence that ends up in the second paragraph of a dozen articles.
It does not appear in Anthropic’s announcement. Checked directly, more than once: the person is named as a contributor on that page and is not quoted at all. One outlet attributes a different quotation to him — about USB — while also giving his title in a form that conflicts with the primary source.
Nobody fabricated anything dramatic here. A plausible line got attached to a plausible person, and each retelling made it more citable. The only thing that catches it is question 2: load the primary source and look for the sentence. It takes a minute, and it is the minute that separates a claim from a rumour with good manners.
The general shape is worth naming, because it recurs: a quote that is too well-suited to the headline is worth checking before it is worth repeating.
What the Primary Source Was Actually Good For
Applying the method to the announcement itself, rather than to the coverage, produces things most of the articles did not carry:
- The status, stated plainly. MHS is a gated research preview; access “is by application during the preview,” and as of 29 August 2026 there is no public specification, schema, SDK or repository. Every architectural claim in circulation, including in this article, rests on one vendor narrative nobody outside the preview can inspect.
- The safety mechanism is prose. The driver “contains tags that let the user write this information directly in natural language,” which a person writes or supplies “by chatting to an agent that interviews them about their hardware setup.”
- The strongest enforcement-sounding sentence in circulation is a researcher’s quotation, not a specification. Virginie Ruetten, a scientist in the Ahrens lab at HHMI Janelia, is quoted saying “MHS enforces device-level safety limits.” Anthropic’s own prose is narrower, describing what the generated reference file documents.
- Anthropic makes no claim in either direction about certification or safety-rating. The category is not addressed — which is a fact about the announcement worth reporting precisely, rather than filling in from either direction.
None of that required special access. It required reading the page and noticing who was speaking in each sentence.
The Habit, Compressed
For any announcement you cannot try yourself:
- Open the primary source first, before any coverage. Ten minutes on the original changes what you notice everywhere else.
- List who is quoted, and where else they appear. New voices mean new reporting.
- Search the primary source for the quote everyone is repeating. If it is not there, find out whose words they are before repeating them.
- Note what the source is silent about. Silence is information, and it is the thing summaries lose first.
- Separate “named as a partner” from “is using in production.” They are printed identically and mean very different things.
- Attribute nothing to a page you could not load. Including a paywalled one.
That is the whole method. It is not sophisticated, and it does not require expertise in the subject — every check above is about provenance, not about hardware standards. Which is why it transfers to the next launch, and the one after that.
Why This Matters More for a Preview Than a Product
With a shipped product, the coverage stops mattering fairly quickly, because people use the thing and reality asserts itself. Reviews accumulate. Bugs surface. The gap between the announcement and the artifact closes.
With a gated preview there is no such correction. The only available evidence is the company’s own account and whatever partners have agreed to say, and that state can persist for months. Everything written during that window is describing a description.
That is not a reason for cynicism about the technology — MHS may well be exactly as useful as its early users say. It is a reason to keep the epistemic status attached to the claim: who said this, were they in a position to know, and could I check it? When the answer to the last one is “not yet,” the honest move is to say so and keep reading, rather than to round the uncertainty off in either direction.
Frequently Asked Questions
What is the fastest way to tell if a tech article did original reporting?
Check whether it quotes anyone the company’s own announcement did not already quote. Press releases ship with pre-approved quotes from named partners; if every voice in the article also appears in the original post, no one was interviewed. It takes about thirty seconds and separates original work from restatement more reliably than any other single check.
Why does connecting an announcement to outside context signal real reporting?
Because that connection cannot come from the press release. When an outlet ties a product to a specific regulation, a prior standard, or a dated precedent, someone had to know that context existed and go check it. In the MHS coverage sample, that is what distinguished The Next Web’s piece — it linked the safety-tag mechanism to Regulation (EU) 2023/1230 on machinery, which no other outlet in the sample mentioned.
How should I treat an article I cannot load or that sits behind a paywall?
Attribute nothing to it. Not a claim, not a framing, not a quote sourced from someone else’s summary of it. An article you have not read is not evidence, and “I saw it referenced somewhere” is precisely the mechanism by which a plausible-sounding error acquires a citation trail.
What is a “too good” quote and why check it?
A quotation that fits the headline unusually neatly — memorable, symmetrical, quotable. Those spread fastest and get checked least. During the MHS launch, a widely repeated line about MHS doing for hardware what MCP did for software was attributed to a named Anthropic staff member; the sentence does not appear in Anthropic’s announcement, where he is listed as a contributor and not quoted at all. Searching the primary source for the exact sentence is a one-minute check.
Does this method require understanding the technology?
No, and that is the point. Every check is about provenance rather than subject matter: who is quoted, was the primary source read, what outside context was added, what the piece is careful not to claim. You can apply all of it to a field you know nothing about, which is exactly when you most need it.
📦 Recommended Resources
The method in this article costs nothing to practise — the primary sources are free, and the whole skill is the habit of opening them. If you want to go deeper on why confident claims survive without evidence, and how organisations end up believing their own summaries, these three are worth the time.
- Reproducibility and Replicability in Science — the National Academies consensus study, and the most rigorous available treatment of the gap between “a result was announced” and “a result holds.” Its central distinction — between reproducing an analysis and independently replicating a finding — is exactly the distinction between an outlet restating a press release and an outlet checking something. The full text is free to read online from the National Academies, so buy the paperback only if you want it on a shelf.
- Normal Accidents: Living with High-Risk Technologies by Charles Perrow — a masterclass in reading official accounts of complex events critically. Perrow’s reconstructions of how incidents were described versus what actually happened are the best available training in noticing whose account you are reading, and what that account is structurally unable to see. Directly transferable to reading vendor announcements.
- Human-in-the-Loop Machine Learning: Active Learning and Annotation for Human-Centered AI by Robert (Munro) Monarch — included for one specific reason: it is a serious book about how humans and automated systems produce claims together, and about where human judgement actually adds information rather than rubber-stamping. Useful if the announcements you are evaluating increasingly come from, or are about, systems that generate text. (Published under both Munro and Monarch — same author.)
✅ Pro tip: The cheapest habit in this whole article: before reading any coverage of an announcement, spend ten minutes on the original post. You will notice omissions and hedges in the coverage that are invisible if you arrive at the primary source second — and you will never again mistake a partner list for a customer list.