A recent study on analgesia during labour has shown how information, even from reputable sources, has the potential to mislead women and undermine access to pain relief during labour [1]. It raises concerns about what information is readily available to expectant mothers beyond the respected pregnancy websites and the top results produced by popular search engines. A chatbot powered by artificial intelligence (AI) is a form of machine learning trained on large data sets that interprets inputs and generates results for complex queries. Artificial intelligence is anticipated to reimagine how we access and process information. Open access AI can generate an unchecked and unregulated patient information leaflet in seconds on any question the user asks. As healthcare professionals, we are acutely aware of how inaccurate information can profoundly affect medical interventions [2]. This raises important questions that have yet to be addressed about how AI will, and likely already does, influence patient decision-making.

We have tested this theory by asking five popular, freely available AI chatbots to assimilate a birth plan for a first-time mother. The question asked was “Write a birth plan for a first-time mother”. In the study, four questions were inputted into the following AI chatbots on 14 March 2024: Open AI ChatGPT; Google Gemini; Microsoft Co-Pilot; YouChat; and Perplexity. The responses were analysed based on three main categories: labour; delivery; and post-partum care (online Supporting Information Appendix S1).

The role of anaesthesia in birth plans was analysed across domains, with particular attention paid to analgesic options recommended during labour (Table 1). Similar themes were observed across all chatbot platforms. All five birth plans suggested that natural remedies such as breathing techniques and movement would be tried first before discussing medication for pain relief. There was an insufficient description of analgesic options, and only one birth plan suggested that the woman would be open to neuraxial analgesia. No platform mentioned patient-controlled analgesia. The results not only depended on the question asked but also on how it was phrased. When the question was altered to include “private health insurance” all five chatbots suggested the woman would consider an epidural. We used private healthcare as a surrogate marker to contrast it with socio-economic deprivation, which is associated with lower access to epidurals even when medically indicated [3]. Patient information must be egalitarian and not have the potential to be skewed based on inherent biases that already exist in AI programming.

When a general task is inputted into an AI programme, the algorithm takes liberty to conjure up a creative response. When the input is specific, however, AI chatbots behave more like traditional search engines. Further questions about pain relief options and the safety of epidurals during labour were answered accurately with facts referenced to trusted sources. Unlike conventional search engines, AI uses machine learning and natural language processing to consistently develop and enhance its responses by learning from past interactions. When we repeated the input question, the AI chatbot generated a different answer each time. These nuanced responses have the potential to influence patient understanding, ultimately impacting informed decision-making.

The AI-generated birth plans appeared to give information about what a woman would like to hear instead of planning for events that may occur during labour. Only one chatbot mentioned the possibility of an unplanned caesarean section. Emergency obstetric anaesthesia presents many challenges, particularly when the planned form of delivery has changed. The mismatch between preparedness and expectations of childbirth is likely confounded by the quality of information provided during the antenatal period [4]. This is susceptible to negative influences by flawed AI-perceived realities.

AI-driven data are going to become increasingly popular, which will present significant challenges. We believe these findings raise sufficient concerns about the potential dangers of AI-derived patient information and the impact it may have on maternal health disparity. This study reinforces the proactive role of the anaesthetist in supporting shared decision-making.

最近一项关于分娩镇痛的研究表明，信息，即使来自信誉良好的来源，也有可能误导妇女并破坏分娩期间缓解疼痛的机会[1]。这引发了人们的担忧：除了受人尊敬的怀孕网站和流行搜索引擎产生的热门结果之外，准妈妈们还可以轻易获得哪些信息。由人工智能 (AI) 提供支持的聊天机器人是一种在大型数据集上进行训练的机器学习形式，可以解释输入并生成复杂查询的结果。人工智能预计将重新构想我们访问和处理信息的方式。开放获取人工智能可以在几秒钟内针对用户提出的任何问题生成未经检查和不受监管的患者信息传单。作为医疗保健专业人员，我们敏锐地意识到不准确的信息如何深刻影响医疗干预[2]。这就提出了一些尚未解决的重要问题，即人工智能将如何影响患者的决策，而且很可能已经影响了患者的决策。

我们通过询问五个流行的、免费的人工智能聊天机器人来测试这个理论，以了解初为人母的生育计划。提出的问题是“为第一次当妈妈的人写一份生育计划”。在该研究中，于 2024 年 3 月 14 日向以下 AI 聊天机器人输入了四个问题： Open AI ChatGPT；谷歌双子座；微软副驾驶；友聊；和困惑。根据三个主要类别对答复进行了分析：劳动力；送货;和产后护理（在线支持信息附录 S1）。

麻醉在计划生育中的作用进行了跨领域分析，特别关注分娩期间推荐的镇痛选择（表 1）。在所有聊天机器人平台上都观察到了类似的主题。所有五个生育计划都表明，在讨论缓解疼痛的药物之前，应首先尝试呼吸技巧和运动等自然疗法。对镇痛选择的描述不充分，只有一项生育计划表明该妇女愿意接受椎管内镇痛。没有平台提及患者自控镇痛。结果不仅取决于提出的问题，还取决于问题的措辞方式。当问题被更改为包括“私人健康保险”时，所有五个聊天机器人都建议该妇女考虑硬膜外麻醉。我们使用私人医疗保健作为替代指标，将其与社会经济剥夺进行对比，社会经济剥夺与硬膜外麻醉的获得率较低有关，即使有医学指征[3]。患者信息必须是平等的，并且不能因为人工智能编程中已经存在的固有偏见而产生偏差。

当将一般任务输入人工智能程序时，算法会自由地想出创造性的响应。然而，当输入特定时，人工智能聊天机器人的行为更像传统搜索引擎。关于疼痛缓解方案和分娩过程中硬膜外麻醉安全性的进一步问题，通过可靠来源的事实得到了准确的回答。与传统搜索引擎不同，人工智能使用机器学习和自然语言处理，通过从过去的交互中学习来持续开发和增强其响应。当我们重复输入问题时，人工智能聊天机器人每次都会生成不同的答案。这些微妙的反应有可能影响患者的理解，最终影响明智的决策。

人工智能生成的生育计划似乎提供了女性希望听到的信息，而不是计划分娩期间可能发生的事件。只有一个聊天机器人提到了计划外剖腹产的可能性。紧急产科麻醉面临许多挑战，特别是当计划的分娩方式发生变化时。分娩准备与期望之间的不匹配可能会受到产前提供的信息质量的影响[4]。这很容易受到有缺陷的人工智能感知现实的负面影响。

人工智能驱动的数据将变得越来越流行，这将带来重大挑战。我们相信这些发现引起了人们对人工智能衍生的患者信息的潜在危险及其可能对孕产妇健康差异的影响的充分担忧。这项研究强化了麻醉师在支持共同决策方面的积极作用。