Editorial standards

AI News is independent and non-partisan. Our news, analysis and explainers put evidence first, and they are written for readers, not for the companies, investors, campaigners or governments we cover. In practice:

• Claims are attributed. Claims about what a system can do, how it scores on a benchmark, how safe it is, how much money it makes or how many people use it are attributed to whoever makes them, and we say whether anyone independent has verified them.
• Announcements are not achievements. We distinguish announced from released from generally available, a preprint from a peer-reviewed paper, and a demo from a product.
• No anthropomorphizing. We don’t write that a model thinks, wants or feels, except inside an attributed quotation.
• No hype or doom vocabulary. Words such as “revolutionary,” “game-changing” or “terrifying,” and claims such as “AGI is here,” appear only in quotations.
• Predictions are labeled. A forecast about where AI is heading is presented as a prediction, with who made it.
• Every side, in its own words. In debates over policy, safety and the future of work, each side is represented by what it actually said, accurately and in context.
• Opinion argues in the open. Our opinion columns take clear positions, grounded in evidence, never along party lines, and are always labeled as opinion.
• Creators’ reviews are the exception. These rules don’t apply to our articles on YouTube creators’ reviews of AI models. Those articles relay a named creator’s view in the creator’s own spirit, enthusiasm and vivid language included, and are labeled as the creator’s review (see Video reviews).

You can read more about who we are on our About page.

• Grounded in reporting. Our articles are based on reporting by established news organizations and specialist publications and, where available, on primary sources such as company announcements, research papers, court filings, regulatory documents and government data. Our video reviews are based on the creators’ own videos.
• Every claim traceable. Each factual claim must trace to that source material. We do not publish facts, figures, events or dates we cannot source, and an AI model’s own memory is never a source.
• Primary sources in their place. A company’s or lab’s own announcement is authoritative about what it announces, such as a release, a price or a policy, but not about how capable, safe or successful its own products are. We cover an announcement only once another outlet has reported it.
• Attribution. When information comes from a single outlet’s reporting, or is disputed, we say who reported it.
• Corroboration. We give more weight to facts reported by several independent outlets. When credible sources disagree, we say so rather than choosing the version we prefer.
• Anonymous sources. We have no confidential sources of our own. When a report we cite relies on anonymous sources, we make that clear.
• Numbers. We report benchmark scores, funding, valuations and other figures exactly as our sources give them, with who measured them and the context needed to understand them, and we never invent or estimate statistics of our own.
• Show our work. Every article ends with links to the sources it draws on, so you can check them yourself.

We follow the published feeds of news organizations, specialist trade publications, research institutions and AI companies. Only stories about artificial intelligence are considered, and a story reported by a single source needs an established newsroom or trade publication behind it. We also follow the reviews of a list of YouTube creators, under the rules in Video reviews below.

• We respect publishers’ choices about automated use. With one exception, YouTube, whose captions and video pages we read (see Video reviews), we fetch nothing from a publisher whose terms or robots.txt forbid it, and publishers we know forbid it are also excluded from the web searches our AI models run while researching.
• When a publisher opts out of AI use, through a robots.txt rule for Anthropic’s crawlers or a Content-Signal: ai-input=no directive, our crawler never passes its text to our AI models, and publishers we know have opted out are excluded from those web searches and never cited. Their headlines can tell us that a story is being widely reported, but our articles are based on other sources.
• When a publisher blocks our crawler from its article pages, we don’t fetch them; we use the summary in its public feed instead.
• Our crawler identifies itself as InferenceImperiumBot (formerly AINews24Bot) and follows robots.txt, except on YouTube (see Video reviews). Our About page explains how publishers can opt out.

We follow the reviews of a list of YouTube creators, chosen by the publisher, who test AI models and publish what they think of them, and editors can also add a review from a channel on that list that our checks missed, such as an older one. When a creator reviews a model, we may publish an article on their review. When several have tested the same model the same way, we may publish a roundup that sets their results side by side, with each channel credited (see Roundups compare like with like, below).

These articles report the creators’ views, not ours. Every fact in them must trace to the videos, and they meet our standards for quotations, headlines, original writing, sensitive subjects, corrections and conflicts of interest. They don’t follow Independence, not hype, and they don’t add the outside context our other articles do. They follow these rules:

• How we read the videos. Our software reads each video’s captions (the text YouTube shows as subtitles, often generated automatically) directly from YouTube, or an editor pastes in a transcript. It finds new videos and their details on the channels’ and videos’ public pages, or through YouTube’s official API. YouTube’s robots.txt disallows automated access to the addresses that serve captions, and its terms of service don’t allow automated access to YouTube without its permission, which we don’t have. We read them anyway, to report accurately on reviews their creators have published for anyone to watch. It is the only exception to our practice of following robots.txt.
• The creator’s view, in their spirit. The article mirrors the review: its overall verdict, emphasis and tone match the creator’s. An enthusiastic review makes an enthusiastic article, a disappointed one a disappointed article, and a mixed one a mixed article in the same proportion. The creator’s criticisms appear as, and as much as, the creator gave them.
• Roundups compare like with like. We publish a roundup only when several creators tested the same model on the same kind of task, against the same rival model when they compared it with one, and reported concrete results: which model won or did better and at what, scores, times, costs, what worked or broke. Before a roundup is written, our software reads each video for the tests it ran and checks that enough channels share one; reviews that each test something different don’t make a roundup. The roundup sets those results side by side, test by test and credited to each channel in their spirit, and draws no conclusion broader than the tests.
• No second opinion. We don’t balance a creator’s verdict, critique their methods, add caveats or notes that their results weren’t independently verified, or compare them with benchmarks, other coverage or the model maker’s announcements. The article ends with the creator’s verdict, not ours. A creator’s enthusiastic or vivid language may be quoted or conveyed in attributed paraphrase.
• Clearly labeled. The headline and the opening name the creator and say it is their video review, and the article carries a note saying whose review it reports and that the opinions are theirs. A creator’s opinion is never presented as our own finding, and we never claim to have watched or tested anything ourselves.
• Only what the videos say. Everything in the article comes from the videos, apart from the model’s name, maker and release when the creator didn’t give them. No prices, benchmark scores or maker’s claims appear unless the creator discussed them, and nothing is invented.
• Quotations. Quotes come word for word from the transcript, are kept short and are attributed to the channel. Automatic captions can mishear names and numbers, so anything that may have been transcribed wrongly is paraphrased instead.
• Creators are credited, not guessed at. Articles name the channel and link every video they draw on. We don’t guess at a creator’s real name, gender or credentials.
• Sponsorships are disclosed. When a video or its description says it is sponsored, a paid partnership or carries affiliate links, the article says so in one plain, factual sentence. We never name a sponsor our sources don’t name, or one whose name automatic captions may have misheard.
• Creators can opt out. A creator who would rather we didn’t report on their videos can email [email protected], and we will stop using them.

Readers should always know what kind of article they are reading. Every article is labeled with one of four formats:

News

A report of what happened, drawn from the source reporting. It is non-partisan, attributes every claim to whoever made it, and keeps loaded language out.

Analysis

Interpretation of the news: what it means, why it happened and what could come next, grounded in the reported facts. Analysis can reach conclusions, but every conclusion must rest on sourced facts, and every prediction is labeled as one.

Explainer

Background that helps you understand an issue — how a technology or a rule works, how we got here and what the key terms mean.

Opinion

Argument. Opinion pieces argue clear positions openly and forcefully, grounded in evidence and never along party lines. They run under named columns and are held to exactly the same factual standards as everything else we publish.

News, analysis and explainers carry desk bylines, such as the Policy Desk. Opinion pieces appear under named columns; pieces from The Editorial Board represent the considered view of the publication as a whole. No byline on AI News is the name of an invented person.

An article on a YouTube creator’s review is labeled analysis, but it reports the creator’s view rather than our interpretation, and a note on the article says so (see Video reviews).

• We represent the views of people we disagree with accurately and in context, in their own words when we quote them.
• We report what people and companies say and do. We don’t present guesses about their motives as fact; interpretation belongs in clearly labeled analysis or opinion.
• We criticize public figures, companies and institutions on the basis of their public record. We do not target private individuals, publish their personal information, or mock anyone for who they are.
• People charged with crimes are described as accused until a court finds otherwise.
• When a person or organization faces serious allegations, we include their response when our sources report one.
• We never use dehumanizing language about anyone, including the people we criticize most sharply.

Quotation marks mean exact words.

• We quote people only when their exact words appear in a source, and we attribute every quotation to the speaker and to the outlet that reported it.
• We don’t alter quotations, stitch separate remarks together, or trim them in ways that change their meaning.
• Paraphrase is never placed inside quotation marks.
• Every quotation must match its source word for word. If the wording cannot be verified, it must be removed or expressed as an attributed paraphrase.
• Text produced by an AI system is never presented as a quotation from a person.

• A headline must be supported by the article beneath it. It may frame a story; it may not overstate it.
• We don’t write headlines that withhold the key fact to bait a click, promise something the article doesn’t deliver, or use sensational language the facts don’t justify. A company’s claim is never turned into a headline fact.
• A question headline is used only when the article substantively answers the question.
• Opinion headlines may argue a position. They may not misstate facts.
• We test alternative headlines on some articles, so readers may see different versions while a test runs. Every version is checked for accuracy against the article first, and the one readers engage with most is kept. Our privacy policy explains what headline tests record.

• We use licensed photographs from stock libraries, credited to the photographer and the library. Stock photos illustrate a subject; we never present them as photographs of the specific events in a story.
• We also use openly licensed and public-domain photographs from archives such as Wikimedia Commons, NASA and public collections. Each is credited with its author, its license and where it comes from, and is used only when it shows the story’s own place or subject, never as a stand-in for a different event.
• We also use official photographs that government offices publish for public use, such as the White House’s and the State Department’s, when they show the story’s own occasion or the official the story is about, and only when the story names everyone their caption names. Others at an official occasion, such as officials standing beside the speaker, may appear in them. Each keeps its original caption, with its date, and is credited to the office and its photographer.
• A photograph of a public figure named in a story is labeled as a file photo, with the year it was taken when known. Apart from these official photographs and video thumbnails (below), we don’t otherwise show photographs of identifiable people.
• A story about a particular website may show a screenshot of its homepage, credited with the date it was captured, or the Internet Archive’s recent capture of it, credited as such.
• An article on a creator’s video review shows that video’s thumbnail, the picture the creator chose for it, which may show the creator. It is credited to the channel and links to the video on YouTube; a roundup shows the thumbnail of one of its videos.
• When no suitable photograph is available, especially for a story about a scene such as a fire, a protest or a strike, an article may show a photorealistic image generated with an OpenAI image model. It shows the story’s own subject, such as the place, the objects or the kind of work the article describes, using only facts in the article. It is not a photograph of the event and was not taken there. Generated images never show a real, identifiable person or stand in for one, an injured person, or a company logo, and each is checked before it is used. These images are not individually labeled.
• Other articles show a typographic cover built from the section’s colors. Generated images and covers are not photographs, and we never alter photographs in ways that change their meaning.
• Images carry descriptive alternative text for readers who use screen readers.

This product uses the Flickr API but is not endorsed or certified by SmugMug, Inc.

• Misuse. We report on how AI is misused, but we never publish operational instructions for causing harm, such as working malicious code, instructions for weapons, prompts designed to defeat a system’s safety measures, or ways to evade detection.
• Health. We report on AI in medicine and health care, but we don’t give medical advice. Please talk to a qualified professional about your own health.
• Money. We report on AI companies, their finances and the markets around them, but we don’t give investment advice.
• Voting. We cover AI in elections, but we don’t publish practical voting instructions such as registration deadlines, polling places or eligibility rules, where a single error could stop someone from voting. Your state or local election office and vote.gov are the authoritative sources.
• Children. We don’t publish content about identifiable minors.
• Violence. We report on violence without graphic detail and without sensationalizing it, and we never publish anything that incites violence or harassment.

• We correct factual errors promptly and in public. A dated correction note explaining what was wrong is added to the article, and every correction is listed on our corrections page.
• When new reporting suggests that something we published was wrong, the corrected article is fact-checked again in full against its sources. It is published with its correction note only if that check passes cleanly; otherwise an editor reviews it first.
• We fix spelling, grammar and formatting without a note when the change doesn’t affect meaning.
• When a story develops after we publish it, we may update the article with new, sourced information. An update keeps the article’s address and its original publication date, is fact-checked again as a whole, and adds a dated note at the top of the article saying what changed. The article shows when it was last updated.
• Updates are not corrections and are not listed on the corrections page. If new reporting shows that something we published was wrong, the change is published as a correction, with a dated note, and listed on the corrections page.
• When a development is a new event, or an article is no longer recent, we publish a new follow-up article that links the earlier one, and the earlier article points readers to it. Opinion pieces are never updated with new facts.
• An update changes a headline only when the news has overtaken it, and the new headline is checked like the rest of the article. Headline changes that result from a headline test are not corrections, because every tested headline must already be accurate.
• We keep an internal record of every version of each article.
• If an article is fundamentally flawed and cannot be fixed with a correction, we withdraw it.

To report an error, email [email protected].

Advertising pays for AI News. It has no influence over what we cover or what we say.

• Advertisers, including AI companies, cannot buy coverage, review articles before publication or ask for changes. Editorial decisions are independent of advertising arrangements.
• Ads are placed by automated systems, mostly Google’s, and are always clearly labeled as advertising.
• An ad appearing on our site is not an endorsement of the advertiser, its products or its views.
• We don’t accept ads designed to look like our journalism. Our advertising standards list what else we won’t accept.

AI News does not publish sponsored articles, advertorials or paid native content.

If that ever changes, sponsored content will be labeled “Sponsored” at the top, will name the sponsor, will look clearly different from our journalism, and will never appear under a newsroom desk byline or an opinion column. Sponsorships in our newsletter are labeled as sponsored and kept separate from our editorial summaries.

Our articles are produced with AI models made by companies we cover: Anthropic’s Claude models and OpenAI’s GPT and image models. We pay for them like any other customer, and neither company has any say over, or advance sight of, what we publish.

• Coverage of Anthropic and OpenAI follows exactly the same standards as coverage of anyone else: no softer treatment, no harsher treatment and no promotion.
• Facts are checked against outside sources, never against what a model says about itself or its maker.
• Articles substantially about Anthropic or OpenAI, or their products such as Claude, ChatGPT, GPT, Sora and Codex, carry a short disclosure note.
• AI News is not affiliated with any AI company, investor, political party, campaign or candidate, and does not accept payment from them, or from anyone else, in exchange for coverage.
• We don’t accept gifts, payments or favors in exchange for coverage.
• We will disclose in the article any financial interest the publisher has in a company or subject we cover.
• Our service providers have no say over what we publish and are covered by the same editorial standards as everyone else.

• Comments under our articles are written by readers and kept visibly separate from our journalism. They are not reporting or opinion from AI News.
• We never write, generate, edit or pay for reader comments. We publish a comment as written or not at all.
• Comments are moderated against published community guidelines. Disagreement, and criticism of us, of public figures and of their public record, are welcome; harassment, hate, threats and personal information about private people are not.
• Reader comments are never used as sources for our reporting.

These standards apply to everything we publish, and every draft is checked before publication: quotations must match their sources, writing must be original, factual claims must be grounded in sources and headlines must be supported by the article. Drafts that do not meet these requirements are revised or discarded. The publisher is responsible for upholding these standards.

Mistakes can still happen. Our contact page explains how to report an error, and our corrections page records published corrections.

Questions about these standards, or concerns about an article: [email protected].

Factual errors: [email protected]. Messages go to AI News Media, which is responsible for everything we publish.