英国税务机关担心人工智能正在给人们提供错误建议。英国税务海关总署（HM Revenue & Customs，简称HMRC）正试图遏制来自人工智能和社交媒体的错误信息洪流，计划部署“数字大使”，在Reddit等在线论坛明确标示正确建议，以免浏览者感到困惑。该机构还在采取措施，使官方信息更容易被大语言模型抓取。
The UK’s tax authorities are worried that artificial intelligence is giving people bad advice. HM Revenue & Customs is trying to stem a flood of misinformation from AI and social media, with a plan to deploy “digital ambassadors” to clearly flag correct advice in online forums such as Reddit so that browsers don’t get confused. It’s also taking steps to make official information more easily picked up by large language models.
“当信息来源相互冲突时，信任度下降，昂贵错误的风险上升，”HMRC数字传播主管詹姆斯·贝克（James Baker）对彭博社表示。“该部门的结论是，主动干预，在正确的地方提供正确的内容，可以塑造更广泛的信息生态系统对错误信息的反应和纠正速度。”人工智能模型一直在抓取关于税收的错误信息，这些信息往往源自社交媒体，同时遗漏了HMRC的官方建议。这些错误信息随后被提供给寻求建议的英国民众，或在TikTok等网站上被放大。
“Where sources conflict, trust falls and the risk of costly errors rises,” James Baker, head of digital communications at HMRC, told Bloomberg. “The department’s conclusion is that proactive intervention, with the right content in the right places, can shape how the wider information ecosystem responds to and corrects misinformation quickly.”AI models have been picking up erroneous information about tax, often from social media, and missing official advice from HMRC. That is then given to Britons searching for advice or amplified on sites such as TikTok.
今年早些时候就出现过这样一个例子：搜索引擎和人工智能生成的摘要给出了关于企业增值税截止日期的错误建议。HMRC的调查将该错误追溯到一个网站，该网站的内容被复制到其他网站，在社交媒体上广泛传播，随后被人工智能抓取。此类虚假信息可能导致纳税人错过截止日期并面临罚款。
One such example came earlier this year when search engines and AI-generated summaries were giving wrong advice about value-added tax deadlines for businesses. An investigation by HMRC traced the error to a website which had been copied to other sites, spwidely on social media and then picked up by AI. False information such as this could have led to taxpayers missing deadlines and receiving penalties.
贝克表示，“这一事件也表明，公民获取信息的旅程变得多么碎片化，人们在搜索引擎、人工智能工具、应用程序和社交平台之间快速切换。”虽然人工智能公司已警告过幻觉或错误的风险，但英国民众越来越依赖聊天机器人获取财务建议和进行投资。据比价网站Finder的一项调查，约三成英国人曾求助于此类工具获取财务建议。社交媒体影响者可能通过发布教观众如何减少税单的视频，放大了错误建议。
Baker said the “episode also demonstrated how fragmented the citizen information journey has become, with people moving rapidly between search engines, AI tools, apps and social platforms.”While AI companies have warned about the risk of hallucinations or mistakes, Britons are increasingly relying on chatbots for financial advice and investments. Around three in 10 Britons have turned to such tools for financial advice, according to a survey by comparison website Finder. Social media influencers may be amplifying bad advice with videos on how viewers can cut their tax bills.
更多内容：让我们聊聊你的人工智能财务顾问——《梅林谈钱》。错误建议让英国税务当局头疼，因为他们正试图缩小该国590亿英镑（790亿美元）的税收缺口——即应收税款与实际征收税款之间的差额。小企业是这一缺口迄今为止最大的原因，但它们可能也缺乏发现网络建议中错误之处的专业能力。
More: Let’s Talk About Your AI Financial Adviser — Merryn Talks Money Erroneous advice is a headache for UK tax authorities as they look to close the country’s £59 billion ($79 billion) tax gap — the difference between what it is owed and what it takes. Small businesses are by far the largest reason for the gap but they also may not have the expertise to spot inaccuracies in online advice.
英国税务海关总署计划派遣官员作为受信任的数字大使，在在线论坛和社交媒体上与用户交流，标注正确信息。贝克表示，这些管理人员将“相当于该署广播发言人的主动数字版”，并“帮助指明正确信息、建立信任”。官员们还制定了一项新的“英国税务海关总署认证”策略，这是一个旨在对抗网络虚假信息、引导大语言模型给出正确建议的警示系统。据该署称，大语言模型已对这些变化作出回应，并突出了正确信息。
HMRC plans to deploy officials as trusted digital ambassadors to speak to users on online forums and social media, flagging correct information. Baker said the managers will be “a proactive digital equivalent to HMRC’s broadcast spokespeople” and “help to sign post correct information and build trust.”Officials have also set up a new “HMRC Verified” strategy, an alert system aimed at countering online misinformation and guiding LLMs toward the correct advice. The large language models have responded to the changes and highlighted the correct information, according to HMRC.
英国税务海关总署还在考察一款新的人工智能辅助工具，该工具可以查找并核对其官方建议与网上所发说法是否一致，然后再起草回复供人工审核和签批。在该工具推出之前，该署已向员工发布指引，要求将内容制作成“大语言模型可用”的形式。
HMRC is also scoping out a new AI-assisted tool that could find and check claims made online against its official advice before drafting a response for a human to review and sign-off. It has issued guidance to staff to make content “LLM ready” before its launch.
“权威不再只存在于你的网站上，”贝克补充道。“它存在于整个机器可读生态系统中——从谷歌的侧边栏到聊天机器人的摘要，再到Reddit的帖子——如何忠实地复述你实际说过的话。”
“Authority no longer lives on your website alone,” Baker added. “It lives in how faithfully the entire machine-readable ecosystem, from Google’s sidebar to a chatbot’s summary to a Reddit thread, repeats what you actually said.”