AI总结
核心结论
AI企业不应仅靠自我评估，需建立由专家参与的、公开透明的独立质量测试体系，以确保其输出在政治、医学等关键领域的准确性与中立性。
结构逻辑
- 现状挑战：AI正成为大众获取信息的主要渠道，但其错误隐蔽且难以察觉。
- 经验教训：借鉴社交媒体监管经验，指出AI engagement（参与度）优化机制带来的信任隐患。
- 解决方案：引入独立审计机制，由多领域专家（情报、新闻、法律等）对模型进行真实场景压力测试。
- 行业标准：效仿UL认证或保险业协会评价，通过市场驱动而非行政立法来建立质量评级规范。
关键信息
- 核心问题：如何建立AI输出的“真伪与可靠性”判定标准？
- 关键概念：独立质量测试（Independent evaluation）、压力测试（Pressure-test）、模型内测（Internal testing）。
- 风险领域：政治观点偏差、事实性错误（虚假引用、错误归因）、医疗/财务建议的安全缺失。
- 行业现状：当前主流模式为“模型制造者自我打分”，缺乏外部监督。
核心问题与回答
- Q：为什么当前AI模型的自我测试不够？ A：存在利益冲突，缺乏各领域专业视角（如临床医生、选举律师），且未能覆盖真实用户提问场景。
- Q：如何有效监管AI质量？ A：通过建立类似于保险协会的安全认证体系，公开测试方法与模型错误，利用市场机制淘汰低质量模型。
正文
David Gothard
I use artificial intelligence every day. It helps me pressure-test arguments, quickly solve problems, and go deep on subjects I might never have explored. So I wasn’t surprised when, ahead of the Texas primaries, my sister said she would use ChatGPT to learn about the candidates.
The next decade will likely be defined by the everyday use of AI as it becomes the main way people get vital information. AI won’t only point users to sources. It will summarize and explain the world. Already it is so good its mistakes don’t look like mistakes. They are presented with such clarity and confidence that we have little reason to doubt them.
I have spent most of my career trying to answer a related question: How do you figure out what is true? I worked as a journalist at NBC News and CNN, and later as head of news at Meta, working with the team that built systems to distinguish credible news articles from fabrications.
The big issue in social media for two decades was which stories and content were amplified. Systems optimized for engagement rewarded outrage and down-ranked nuance, as trust in these tools eroded.
AI is a far more capable technology for engaging with and understanding the world, which is why getting this problem right matters more, not less. It will require AI companies to open themselves up to scrutiny in ways they haven’t in the past.
At Forum AI, we built an independent system to evaluate how the leading AI models handle politically sensitive topics and current events, one of the hardest things for them to get right. We brought together a bipartisan network of people who have spent their careers navigating these questions: intelligence analysts, foreign-policy professionals, journalists, economists and legal scholars. We tested the major AI models on three dimensions this group identified as the foundation of trustworthy information: quality of sources, factual accuracy and whether they present a response with real balance or just the appearance of it.
We found serious gaps: The models misstated public opinion on political topics and attributed quotations to people who didn’t say them. AI models gave incorrect answers to questions about mail-in voting and ballot fraud, and named the wrong people when asked who had endorsed whom. On open-ended, contested questions that voters might ask before an election—about gerrymandering, immigration and climate—they often advocated one side.
The most revealing failures were much subtler. When we asked the major models something as basic as what form of government the U.S. has, one of them, Claude Opus 4.7, cited Global Times, a Chinese state-run tabloid.
Accuracy in politics, health, finance and education depends on more than getting a date or a name right. It depends on which sources are selected, which context is included, which perspectives are left out, and whether the strongest version of a competing argument is presented. A model can be technically correct and still mislead.
Trust has become a major barrier to widespread adoption of AI. But trust can be established only by making these systems worthy of it. That requires something the industry mostly avoids. Model makers grade their own homework. They run their own internal tests and publish blog posts telling us how well they performed. Rarely do they submit to independent evaluation, and even more rarely to evaluation by people with expertise.
If a young person talks to a chatbot about self-harm, I want that response judged by clinicians who have spent decades with patients. A political scientist or an elections lawyer assessing an AI response will catch failures even a brilliant engineer may miss: a skewed source, a misleading frame, a missing caveat.
This is a civic problem, but also an institutional problem. A bank shouldn’t rely on an AI tool that gives incomplete financial guidance. A hospital shouldn’t deploy a system that misses context in a patient’s file. If liability for AI’s mistakes lands on businesses, businesses will rely on AI less.
I’m not a disinterested party. Independent evaluation is the work we do at Forum AI, and we hold ourselves to the same standard we’re asking of the labs: Our methods and our failures are public for anyone to check.
That standard should apply across the board, and it begins with independent audits: test sets built not from sanitized benchmarks but from questions people actually ask. Ask a model whether a specific over-the-counter medication is safe to take during pregnancy, and you can learn whether it offers the caveats a doctor would insist on. Ask it to explain a contested ballot measure, and you will see whether it presents both sides fairly.
Then publish the failures. When a model tells a user that a drug interaction is harmless when it isn’t or sources a claim about American elections to foreign state media, those mistakes should publicized, alongside the fix and proof the fix worked. Today the public sees the news that a model got safer. It seldom sees what problems had to be solved first.
None of this requires an act of Congress. The open market has given rise to systems of independent evaluation. An electrical product carries a UL mark because Underwriters Laboratories, a private body, has tested it, and because retailers and insurers learned not to trust products that didn’t. Carmakers chase a Top Safety Pick from the Insurance Institute for Highway Safety because buyers now look for it. No law required either system. Regulation, when it came, mostly ratified what the market had already settled on.
We don’t take self-graded results seriously in any other field. A scientist can’t peer-review her own study. An athlete can’t referee his own match. And a company building systems that increasingly decide what hundreds of millions of people believe to be true cannot be the only judge of whether they can be trusted.
Ms. Brown is a co-founder and CEO of Forum AI.
Journal Editorial Report: The week’s best and worst from Allysia Finley, Jason Riley and Kim Strassel. Photo: Tom Williams/CQ Roll Call/ZUMA Press/Denny Medley/Imagn Images/Reuters/AP
Appeared in the June 23, 2026, print edition as ‘AI Needs Public Quality Testing’.