🔶 OpenAI 的超级 PAC 正资助一家用 AI 生成新闻、攻击行业批评者的网站

上周，倡议组织 Encode 的副总裁兼总法律顾问 Nathan Calvin 向我转发了一张他收到的电子邮件截图。发件人是一名叫 Michael Chen 的记者，正为一家名为 The Wire by Acutus 的刊物，就田纳西州一项 AI 法案征求评论。

Encode 收到的来自“Michael Chen”的电子邮件。

这封邮件有些不对劲。文章的完整标题被提前告知，叙事框架的倾向性很强，且唯一提供的采访形式是“书面问答”。在网络和社交媒体上搜索，没有发现任何与 Acutus 公开关联、名为“Michael Chen”的人。Chen 使用的也是一个通用邮箱地址（reporter@acutuswire.com）；对于一家宣称拥有许多供稿人、且近乎每日发布内容的刊物而言，这并不是记者常见的邮箱惯例。当我用 Pangram 检测这封邮件时——Pangram 是一款可信地声称其误报率近乎为零的 AI 内容检测器——结果显示该邮件“完全由 AI 生成”。

事实证明，Michael Chen 几乎肯定不是真实记者。而 Acutus 几乎肯定根本没有任何真实记者。

Acutus 如何描述自己

Acutus 是一个匿名运营的数字新闻网站，于 2025 年 12 月 29 日上线。在不到四个月内，它已发布 94 篇完整文章，涵盖 AI 政策、参议院选举、药房改革、核能、加密货币监管、特许经营、基于技能的招聘等话题。网站没有员工名单、没有署名、没有具名编辑，也没有解释究竟由谁运营，或它是如何诞生的。

AcutusWire.com 首页截图。

该网站将其内容描述为提供“独立报道”的“专家来源新闻”。其相对简略的“关于”页面将这家刊物定位为一个汇集“各行业专家声音”的协作式新闻平台。实际上，任何人都可以填写 Apply to be a Contributor 表格，申请为该网站撰稿；首页侧栏还邀请访问者就一系列轮换话题投稿。“关于”页面还提到了一个“编辑团队”，称其会主动识别并邀请供稿人为网站写作。

Acutus 实际上是什么

但似乎并没有任何供稿人，至少没有人类供稿人。我将网站上的每篇文章都通过 Pangram 检测（这款 AI 检测器可信地声称其误报率近乎为零）。94 篇文章中，69% 被标记为完全由 AI 生成，另有 28% 被标记为部分由 AI 生成。只有三篇文章被归类为人类撰写。

不过，证据不止于 AI 内容检测器。只要查看网站源代码，任何访问者都能在其中发现 Acutus 运营者留下的痕迹。以下内容是我基于这些代码，对该网站运作方式作出的最佳推测。

该网站是一个 React 应用。通过仔细查看发送到每位用户浏览器中的 JavaScript 文件，我们可以发现该网站幕后编辑界面的元素就这样公开放在那里，任何人都可以检查。这个界面中有一份用于创建新文章的表单，包含一个名为“AI Background Context”的字段，说明是“供 AI 在生成问题和撰写报道时使用的背景信息”。另一个名为“Question Prompts”的字段则被描述为“建议 AI 采访者提出的问题”。

完全依据 AcutusWire.com 公开 JavaScript 代码复原的供稿人后台界面。

该表单用于创建一个“主题”。主题创建后，可通过一个标有“Generate Story Draft”的大按钮将其转化为报道。报道生成后，另一个标有“Regenerate”的按钮可让你重复该流程。同一界面中的独立工具还会利用 AI 从“研究笔记”中提取引语、进行语法检查，并实施多轮由 AI 主导的编辑审查，针对多项编辑基准返回带评分的输出结果。

完全依据 AcutusWire.com 公开 JavaScript 代码复原的供稿人后台界面。

作为通讯社式内容分发安排的一部分，Acutus 将其发布的所有内容都以 Creative Commons 许可授权——这对新闻媒体而言并不寻常。“我们的报道以通讯社服务形式提供，供所有出版商使用，”其“关于”页面写道，指的是一条列出网站每篇已完成报道的 RSS feed。

但该网站暴露的不只是完成后的报道。任何浏览器都可访问的 URL acutuswire.com/api/wire——它负责将文章填充至网站首页——会返回该网站完整的报道数据库，以及 每篇内容的内部制作记录，包括 AI 编辑审查流程的详细结果。看起来 AI 会开展五类审查，其中四类按 100 分制评分（AP 文体规范合规度、引语准确性、来源核验，以及一项仅标为 acutus 的指标），另一类则返回状态（事实核查）。除 acutus 以外，AI 会在每一类中标记问题、提出修正建议，并以时间戳记录解决结果（推测由用户裁定）。

AcutusWire.com/api/wire 公开 API 响应中的一段代码。

时间戳表明，这一流程中的人工环节并未投入多少精力。在单篇报道中，从第一个问题被解决到最后一个问题被解决，整个多轮审查完成时间的中位数为 44 秒；在最后一个问题解决后，中位数仅 10 秒便点击发布。在 feed 中的 94 篇报道里，有 42 篇的自动审查器自身“overall”状态为“needs_revision”，表明 AI 系统自己也认为文章尚未准备好发布。但它们全都照样发布了。

一些被标记的问题包含一个名为“aiOriginalText”的字段，它会在建议修改内容旁保留 AI 模型的原始措辞——只有原始草稿由 AI 生成时，这样的命名才说得通。

AcutusWire.com/api/wire 公开 API 响应中的一段代码。

该网站有一个特点：有迹象表明，其开发目的之一是供 AI 模型自身消费，包括在 robots.txt 文件中对特定 AI 爬虫赋予宽泛权限、存在一个已废弃的供 ChatGPT 使用的 ai-plugin.json 文件，以及包含一个实验性的 llms.txt 文件。后者一方面误导性地将网站描述为“独立新闻”，并称其“立足于一手研究、经验证的来源和直接报道”，另一方面也泄露了有关内容来源的一条线索：“报道遵循 AP Style，并执行严格的零幻觉编辑标准。”

Acutus 的采访基础设施

如果整个流程都是自动化的，那么真实人物的引语是如何进入文章的？

仔细检查后，Acutus 文章中的许多引语似乎取自此前已发布的网络内容，这与它们由 LLM 抓取的情况一致。

Acutus 从此前发布的网络内容中挪用引语的一个例子。

但并非全部如此。至少有一条引语来自一名真实人物，他表示自己曾被征求评论。哈佛商学院教授 Joseph Fuller 在 LinkedIn 上写道，他“很高兴与 Acutuswire 分享我的看法；Acutuswire 是多位独立记者的作品”，并链接到一篇有关基于技能招聘的文章，文中引用了他的观点（该文章发布八天后，众议员 Nancy Mace 提出的《基于技能的联邦合同法案》H.R. 5235 即将提交众议院全体审议）。当然，Nathan Calvin 也收到了一封来自“Michael Chen”、请求采访的电子邮件。

客户端代码表明，为何 Calvin 得到的只是“书面问答”而非电话采访。JavaScript bundle 包含多个提及“AI interviewer”或“reporter agent”的字段。一个用于显示已上传采访内容的字段默认文字为：“尚未通过记者 API 上传采访内容。由 Agent 提交的采访内容将显示在这里供审核。”另一个字段则允许人类输入“建议 AI 采访者提出的问题”。

换句话说：Michael Chen 看起来就是那个 AI Agent。当 Acutus 需要一名在世专家的引语时，它会派出一个机器人去获取。Fuller——以及推测中任何在回复 Acutus 电子邮件后、姓名出现在这些报道中的人——都是在与伪装成人类的软件分享自己的想法。

一些文章还引用了未具名来源，但这些来源看起来并非出自 AI Agent。通常，这类来源是未具名的共和党参议院工作人员、政治操盘手或策略师，多次被描述为“接近参议院领导层的人士”。这些极为便利的匿名引语（与文章所偏向的利益相关方立场高度一致）表明，至少有一些引语是在报道生成过程中被直接输入 AI 的。

谁在 Acutus 背后？

Acutus 并未引起太多公众关注。Google 搜索几乎找不到网站本身以外的内容；在 Twitter 上，它也仅被相对不知名的账号链接过四次。但这四条帖子中有两条是由同一人撰写或转发：Patrick Hynes。他是一家名为 Novus Public Affairs 的共和党公关公司的总裁。Hynes 曾发推宣传一篇关于 Scott Brown 2026 年参议员竞选的 Acutus 文章，并单独转发了一篇由反 AI 安全账号 @DrTechlash 分享的 Acutus 文章。

Hynes 位于新罕布什尔州，这一点很重要，因为 Acutus 内容中有相当惊人的比例涉及美国东北部政治，尤其是新罕布什尔州和缅因州。该网站刊登了关于新罕布什尔州经济、前参议员 Scott Brown 的政治履历、缅因州教育体系、缅因州毒品政策，以及一名缅因州参议院候选人的文章。

Acutus 的报道还与 Hynes 代表 Novus 所开展的公关工作重叠。举一个例子：1 月 24 日——在特朗普总统签署大范围改革、将药品福利管理机构改革写入法律的 10 天前——Acutus 发布了一篇攻击药品福利管理机构的文章。美国药品研究与制造商协会 PhRMA 曾在 2025 年为推动这项确切改革而游说，并在游说方面创纪录地花费了 3819 万美元。尽管已发布文章中没有引述任何 PhRMA 人士，其内部来源日志中却将“PhRMA statements”列为来源。与此同时，Novus 在其网站上将 PhRMA 列为客户。

或许更能说明问题的是：Hynes 本人作为被引述来源，出现在 Acutus 自己的一篇文章中，以 Novus 的身份发言，却没有披露其公司似乎正在运营引用他的刊物。将自己公司的总裁作为所谓独立专家，却不承认这种关系，是对基本新闻规范的明显背离。被引述的言论也并非无关紧要：Hynes 借此称赞新罕布什尔州州长 Kelly Ayotte 推动劳动力住房的举措是在“削减繁文缛节”。这正是新罕布什尔州房屋建筑商协会所寻求的去监管叙事框架；该协会的名字出现在 Novus 公开的客户名单上。

Hynes 也有将政治运作包装成新闻的经历。2010 年，他与一群共和党政治操盘手共同创办了总部位于新罕布什尔州的新闻媒体 NH Journal——这种安排被 The

Last week, Nathan Calvin, vice president and general counsel of the advocacy group Encode, forwarded me a screenshot of an email he had received. The sender, a reporter named Michael Chen, was seeking comment for a story about an AI bill in Tennessee for a publication called The Wire by Acutus.

The email Encode received from “Michael Chen.”

Something was off about the email. The full title of the article was shared in advance, the framing was highly loaded, and the only format offered was “written Q&A.” Web and social media searches turned up no one by the name “Michael Chen” publicly associated with Acutus. Chen was also writing from a generic address (reporter@acutuswire.com), which would be an odd convention for a reporter at a publication that touts many contributors and has a near-daily publication rate. When I ran the message through Pangram, an AI content detector that credibly claims a near-zero false-positive rate, the email came back as “fully AI-generated.”

Michael Chen, it turns out, almost certainly isn’t a real reporter. And Acutus, it turns out, almost certainly doesn’t have any real reporters at all.

How Acutus describes itself

Acutus is an anonymously operated digital news site that launched on Dec. 29, 2025. In less than four months, it has published 94 full-length articles on AI policy, Senate races, pharmacy reform, nuclear energy, crypto regulation, franchising, skills-based hiring, and more. It has no masthead, no bylines, no named editors, and no explanation for exactly who runs it or how it came into existence.

A screenshot of the homepage of AcutusWire.com .

The site describes its content as “expert-sourced journalism” offering “independent reporting.” Its relatively sparse About page frames the publication as a collaborative journalism platform gathering “expert voices from across industries.” Indeed, anyone can fill out the Apply to be a Contributor form to apply to write for the site, and a sidebar on the homepage invites visitors to pitch on a rotating list of topics. The about page also makes reference to an “editorial team,” which proactively identifies and invites contributors to write for the site.

What Acutus actually is

But there don’t seem to be any contributors, or at least not human ones. I ran every article on the site through Pangram (the AI detector that credibly claims a near-zero false-positive rate). Of the 94 articles, 69% came back flagged as fully AI-generated, with another 28% flagged as partially AI-generated. Only three articles were classified as human-authored.

The evidence goes further than AI content detectors, though. Acutus’ operators left their fingerprints in places that any visitor can see by looking at the website’s source code. The following section is my best guess as to how the site operates based on this code.

The website is a React app, and by looking closely at the JavaScript file sent to every user’s browser, we can identify elements from the site’s behind-the-scenes editorial interface sitting there in the open for anyone to inspect. Inside that interface is a form for creating new articles, with a field labeled “AI Background Context,” described as “Background information for the AI to use when generating questions and writing the story.” Another field, labeled “Question Prompts,” is described as “Suggested questions for the AI interviewer to ask.”

A recreation of the user interface for the contributor dashboard, based entirely on the public JavaScript code available at AcutusWire.com .

That form is used to create a “topic.” Once a topic has been created, it is turned into a story using a large button labeled “Generate Story Draft.” After the story has been generated, another button labeled “Regenerate” allows you to repeat the process. Separate tools in the same interface use AI to extract quotes from “research notes,” run a grammar check, and perform a multi-pass AI-led editorial review that returns scored output across a variety of editorial benchmarks.

A recreation of the user interface for the contributor dashboard, based entirely on the public JavaScript code available at AcutusWire.com.

Acutus licenses everything it publishes under Creative Commons, unusual for a news outlet, as part of a wire-service syndication setup. “Our stories are made available as a wire service for all publishers to use,” the about page reads, in reference to an RSS feed listing each finished story on the site.

But the site exposes more than just its finished stories. The URL acutuswire.com/api/wire, accessible in any browser and responsible for populating the homepage of the website with articles, returns the site’s full story database alongside the internal record of how each piece was made, including detailed results from the AI editorial review process. There are five categories of review that appear to be conducted by the AI, four of which are then scored out of 100 (AP style compliance, quote accuracy, source verification, and one simply labeled acutus), and one which returns a status (fact-checking). For each of these categories aside from acutus , the AI flags issues, proposes corrections, and logs resolutions (presumably adjudicated by the user) with timestamps.

A snippet of code from the public API response available at AcutusWire.com/api/wire.

The timestamps suggest that not much effort goes into the human side of this process. Within a single piece, the entire multi-pass review completes in a median of 44 seconds from the first issue being resolved to the last, and the publication click lands a median of 10 seconds after the last issue is resolved. On 42 of the 94 stories in the feed, the automated reviewer’s own “overall” status reads “needs_revision,” a sign that the AI system itself flagged the piece as not yet ready to publish. Each of them was published anyway.

Some flagged issues include a field called “aiOriginalText” that preserves the AI model’s original wording next to the suggested edit — a naming choice that only makes sense if the original drafts were AI-generated.

A snippet of code from the public API response available at AcutusWire.com/api/wire.

One quirk of the site is that there are signs it was developed with the intention of being consumed by AI models themselves, including liberal permissions for specific AI crawlers in the robots.txt file, the presence of a deprecated ai-plugin.json file for ChatGPT, and the inclusion of an experimental llms.txt file. The latter one, despite misleadingly describing the site as “independent journalism” and “grounded in primary research, verified sources, and direct reporting,” also lets slip one clue about its content provenance: “Reporting follows AP Style and a strict zero-hallucination editorial standard.”

Acutus’ interview infrastructure

If the whole pipeline is automated, how do quotes from real people end up in the articles?

On closer inspection, many of the quotes in Acutus pieces appear to be lifted from previously published web content, consistent with them being scraped by an LLM.

An example of a quote on Acutus lifted from previously published web content.

But not all of them. At least one comes from an actual person who says he was approached for comment. Harvard Business School professor Joseph Fuller wrote on LinkedIn that he was “pleased to share my thoughts with Acutuswire, a production of a number of independent journalists,” and linked to an article on skills-based hiring in which he was quoted (an article released eight days before Rep. Nancy Mace’s Skills-Based Federal Contracting Act, H.R. 5235, headed to the House floor). And of course, Nathan Calvin received an email from “Michael Chen” asking for an interview.

The client-side code suggests why Calvin was only offered a “written Q&A” rather than a call. The JavaScript bundle contains multiple fields referring to an “AI interviewer” or a “reporter agent.” The default text for a field showing uploaded interviews reads: “No interviews uploaded via the reporter API yet. Interviews submitted by the agent will appear here for review.” Another field lets a human enter “Suggested questions for the AI interviewer to ask.”

In other words: Michael Chen appears to be the AI agent. When Acutus needs a quote from a living expert, it sends a bot to get one. Fuller — and, presumably, anyone else whose name appears in these stories after responding to an Acutus email — shared their thoughts with software pretending to be a human.

Some articles also cite unnamed sources that don’t look like they came from the AI agent. Usually this is an unnamed Republican Senate staffer, operative, or strategist, in multiple cases described as “close to Senate leadership.” These extremely convenient, anonymous quotes (which align neatly with whichever stakeholder the piece favors) suggest that at least some quotes are being fed directly into the AI during the story-generation process.

Who’s behind Acutus?

Acutus hasn’t attracted much public attention. Google searches return almost nothing beyond the site itself, and on Twitter it has been linked only four times by relatively obscure accounts. But two of those four posts were either written or retweeted by one person: Patrick Hynes, president of a Republican PR firm called Novus Public Affairs. Hynes tweeted an Acutus article about Scott Brown’s 2026 Senate candidacy and separately retweeted an Acutus piece shared by @DrTechlash, an anti-AI-safety account.

Hynes is based in New Hampshire, which matters because a surprising share of Acutus’ content concerns politics in the Northeast, especially New Hampshire and Maine. The site has pieces on New Hampshire’s economy, former Sen. Scott Brown’s record, Maine’s education system, Maine’s drug policy, and a Maine Senate candidate.

Acutus’ reporting also overlaps with Hynes’ PR work on behalf of Novus. To take one example: on Jan. 24 — 10 days before President Trump signed sweeping reform of pharmacy benefit managers into law — Acutus ran a piece attacking them. PhRMA, the pharmaceutical manufacturers’ trade group that spent a record $38.19 million on lobbying in 2025 while pushing for that exact reform, appears in the article’s internal source log as “PhRMA statements,” despite no PhRMA figure being quoted in the published piece. Novus, meanwhile, lists PhRMA as a client on its website.

Perhaps more telling: Hynes himself appears as a quoted source in one of Acutus’ own articles, speaking on behalf of Novus with no disclosure that his firm appears to be operating the publication quoting him. Using your own firm’s president as a supposedly independent expert, without acknowledging the relationship, is a striking departure from basic journalistic norms. The quoted remarks are not incidental, either: Hynes uses them to praise New Hampshire Gov. Kelly Ayotte’s workforce-housing push for “cutting red tape.” This is the precise deregulatory framing sought by the New Hampshire Home Builders Association, a name on Novus’ public client list.

Hynes also has a track record of dressing political operations in the clothes of journalism. In 2010, he co-founded a New Hampshire-based news outlet, NH Journal, with a group of fellow Republican operatives — a setup that The Hill reported led to “skepticism over its independence within New Hampshire circles.” When Hynes departed in 2014 to join a PAC supporting Scott Brown’s campaign for the U.S. Senate in New Hampshire, election experts told The Hill that his PAC work would have “opened up the publication to, at best, questions about its legitimacy as an independent news organization and at worst would have muddied the waters in terms of campaign finance law.” (Interestingly, Hynes has since soured on Brown and is now actively supporting his opponent in the upcoming 2026 election, a notable alignment given that Acutus published a piece critical of Brown).

Acutus’ reporting also blurs the lines of journalistic ethics. One piece criticizes AI-safety advocate and longtime broadcast journalist John Sherman for comments on his podcast suggesting that people would burn down data centers if they understood the risks AI posed. Rather than simply cover the quote, Acutus names each of the clients listed by a video production and consulting firm Sherman runs (unrelated to his podcast) and reports that it “contacted each of the organizations” to ask “whether they were aware of Sherman’s statements and whether they intended to continue working with his firm.” The article said none of the organizations agreed to give comment, but also noted that one “privately indicated” it no longer worked with him (a detail that itself raises questions about whether the AI reporter respected on/off-the-record norms). It also named a former lieutenant governor who serves as an advisor to Sherman’s nonprofit, reporting that “she did not respond to repeated requests for comment and did not denounce the calls for violence.”

From my review of the site, more than a third of Acutus’ published pieces read less like journalism and more like paid advocacy for a specific interest group: stories favorable to the pharmaceutical industry, the cryptocurrency lobby, real estate trade groups, theatrical-film exhibitors, the natural-gas and data-center lobbies, and multiple 2026 Republican Senate campaigns (Dan Sullivan’s in Alaska, John Sununu’s in New Hampshire, Susan Collins’ in Maine, and attacks on multiple Democratic candidates in Michigan). No editorial identity unites these topics, but the client list of a PR firm might.

The obfuscatory funding channel which seems to connect OpenAI to Acutus.

OpenAI’s growing astroturf campaign

There’s still the question of why “Michael Chen” was interested in Encode in particular. Looking through Acutus’ AI coverage is where the picture sharpens. Around 15% of the site’s articles cover AI, and they tend to hit a few talking points shared by the anti-regulation tech lobby. The site casts aspersions on Anthropic, criticizes blue states regulating AI and red states regulating AI, and blames rhetoric from journalists and grassroots organizers for recent violence against OpenAI.

A selection of the AI-focused articles on Acutus.

These are the talking points of Leading The Future, the $125 million super PAC funded primarily by OpenAI president Greg Brockman alongside OpenAI investor a16z, and assembled under the guidance of OpenAI’s chief political operative, Chris Lehane. Similar arguments are being pushed by the operatives linked to Leading The Future, including Chris Lehane, Nathan Leamer, Josh Vlasto, Zac Moffatt, and the anonymous Twitter account Doomer Daylight.

There’s a line connecting Acutus to that orbit. The third name on Novus’ aforementioned client list is Targeted Victory, the GOP consulting firm at the heart of OpenAI’s political apparatus: its CEO co-founded Leading The Future, and as I wrote a few weeks ago, Leading The Future’s websites share many legal and technical characteristics with the sites of other Targeted Victory clients. Hynes himself routinely retweets messages aligned with Leading The Future, including from Leamer directly.

A screenshot of Novus’ client list . Emphasis added.

If I’m right, OpenAI’s super PAC may be using Acutus to push its political agenda under the guise of independent journalism. That would directly contradict OpenAI’s own stated positions. The company prohibits the use of its products for political campaigning or lobbying, and its safety framework previously warned about the risk of AI-generated political influence campaigns — language that has since been removed. Through a network of super PACs and PR firms, the company now appears to be funding exactly that: an AI-powered Potemkin news organization.

It would not be OpenAI’s first foray into astroturfing. The company was in the news this month over accusations that it quietly astroturfed a children’s safety coalition to lend credibility to its policy agenda. Build American AI, the 501(c)(4) advocacy arm of Leading The Future, itself has touted nearly a million “grassroots” supporters — a list generated through extensive paid advertising.

These cases alone, while disingenuous, didn’t surprise me. I was more surprised when the same political machine began targeting me personally with an anonymous Twitter account, using AI-generated tweets to manufacture the appearance of grassroots opposition to the company’s critics.

Now OpenAI’s super PAC appears to be using AI, perhaps OpenAI’s very own models, to churn out content that advances its political interests and discredits its opponents. Worse yet, Acutus is blatantly lying about that fact, calling the articles it publishes “expert-sourced” and its reporting “independent journalism.” This seems like exactly the sort of AI-driven political influence campaign that OpenAI once considered a major risk category, and still prohibits in its usage policies to this day.