控制AI幻觉：在医学信息和医学写作工作流程中建立以证据为先的信任
Controlling AI Hallucinations: Building Evidence- First Trust in Medical Information and Medical Writing Workflows
Ome Ogbru, PharmD，AINGENS CEO兼创始人信任才是真正的问题AI的输出是否值得信任？这是医疗专业人员针对生成式AI提出的最重要问题之一。这一问题对于医学信息和医学写作尤为重要，因为响应的价值取决于其清晰度，以及每一项陈述是否都有可验证的证据作为依据，并且能够经受专家审查。这一点之所以重要，是因为这些工作流程会影响临床判断、科学交流以及受监管的医疗沟通。
AI正在取得进展的领域
对于临床医生而言，医学信息通常意味着通过查阅文献、核对治疗指南，或综合多个来源的信息，为护理决策寻找并解读证据，以支持治疗决策。对于制药企业医学信息团队而言，这意味着制定准确、平衡且可追溯的科学响应，从而影响医疗专业人员如何解读数据。医学写作决定了证据如何在研究、科学交流和受监管环境中被总结、置于背景中并进行沟通。
生成式AI在所有这些场景中都具有天然吸引力，因为相关工作以文本为主、劳动密集、耗时，并且难以高效扩大规模。AI解决方案可以加快文献发现、创建草稿，并帮助专业人员更快地完成医学沟通工作流程。
然而，这些益处也伴随着AI幻觉这一重大风险。如果不加以控制，AI幻觉可能影响科学信息的获取方式，影响与医疗护理相关的决策，并助长错误信息。
临床医生面临的AI信任问题
对于临床医生而言，问题在于AI是在帮助澄清证据，还是让不确定性看起来像事实。如果AI系统提供了简洁的摘要，但结论的来源并不明确，那么临床医生就无法对该输出进行适当审查，却被要求予以信任。在临床决策工作流程中，这并不是效率，而是隐藏的风险。
临床医生需要的是一个能够帮助他们快速找到可信证据，并显示信息来源的系统。
医学信息与医学写作中的AI困境
对于制药企业医学信息团队和医学写作者而言，这一挑战同样严峻。这些团队在高度受监管的环境中开展工作，准确性、平衡性和可追溯性是不可妥协的要求。
AI可以通过加快文献审查、发现相关证据和生成初稿来提供帮助。但如果模型混淆不同研究的发现、夸大结论、遗漏背景信息，或编造支持性参考文献，那么输出就会变得不那么有用，并可能带来风险。
解决方案始于系统设计
幻觉并不意味着AI不可用，而是意味着系统设计更加重要。用于医学和科学工作流程的AI系统应以证据为先，包含稳健的检索管线，使用强大的模型，并提供内置的来源可追溯性，使用户能够准确看到陈述的来源。这些系统应鼓励用户将平台指向可信且经过审核的资源，而不是依赖模型的一般训练知识。
这些系统应使审查更加容易，并帮助用户检查源材料、编辑输出、识别响应何时没有证据依据，以及在缺少可信源材料时明确说明。模型应能够直截了当地回应：“我不知道”，或者“我没有足够的信息来提供可信的答案”。
这种克制和透明度是AI平台的优势，而不是弱点。
工作流程设计仍然需要人类专业知识
仅靠技术是不够的。安全且有效地使用AI还取决于工作流程设计。人的责任包括定义任务、决定是否应使用AI、选择合适的平台、设置防护措施、识别可信的知识来源、提供清晰的指令，以及审查和修订输出。在具备资质的人员对其进行审查并承担责任之前，AI输出都应被视为草稿，而不是最终成果。
这正是把AI作为自动化捷径使用，与把它作为受管理的生产力工具使用之间的区别。
用户培训、实践和预期同样重要
用户培训、实践，以及愿意尝试AI解决方案的意愿至关重要。许多用户可能会觉得，学习新系统所需的认知负担不值得付出，尤其是在他们还必须审查输出的情况下。随着用户对平台越来越熟悉，有效使用平台所需的努力会减少，而生产力的提升会显著增加。
人们普遍认为，审查AI输出并通过多轮提示对其进行完善，会抵消节省的时间。这种比较和结论并不合乎逻辑。一份由人撰写、需要数天准备且仍然包含缺口或错误的草稿，通常要经过多轮审查才能定稿。如果AI解决方案能够帮助用户在几秒钟内找到相关文献并在几秒钟内生成一份可用的草稿，而用户随后只需花几分钟进行审查和完善，那么净效率提升仍然是非常可观的。
用户应对AI解决方案的能力有现实的预期。一个常见错误是认为AI模型能够在每个领域都充当专家。受过训练的专业人员仍然是专家，而AI平台是支持他们的工具。
在医学信息、医学写作和临床工作流程中成功使用AI，取决于平台以及了解其能力和局限性、并愿意承担适当使用责任的用户。
领导者应优先考虑什么
对于决策者而言，方法应当是直接明了的。组织不应主要依据速度、受欢迎程度或雄心勃勃的宣传来评估用于医学信息、医学写作或临床工作流程的AI。他们应优先选择具备以下特点的系统：
以证据为依据透明可审计针对特定工作流程构建旨在支持人工审查系统只是其中一部分。领导者必须为用户配备适当的培训，提供支持，设定现实的预期，建立治理机制，并确保对AI支持的工作承担责任。
信任必须经过设计，而不是想当然
幻觉并不是避免在医学信息和医学写作中使用AI的理由。当合适的平台与合适的工作流程相结合时，幻觉可以得到最小化，并在影响最终输出之前被发现。最佳方法既不是盲目信任AI，也不是一概拒绝AI，而是通过以证据为先的系统设计、治理、有效的工作流程，以及人工监督和问责来建立信任。
关于Ome Ogbru, PharmD
Ome Ogbru, PharmD，是AINGENS的CEO兼创始人。AINGENS是一家生命科学软件公司，正在为科学和医学工作流程构建以证据为先的AI平台。凭借在制药、biotech和医疗保健领域超过20年的经验，他曾担任临床药师、教授和全球医学信息负责人，在科学、监管和内容创作的交叉领域开展工作。
Ogbru博士亲身经历了循证内容工作流程中的低效问题，因此创立AINGENS，旨在开发实用、可用于企业的解决方案，改善科学信息的创建、审查和交付方式。通过其旗舰平台MACg（Medical Affairs Content Generator），他专注于在不牺牲准确性或合规性的前提下，实现更快速、更可靠的医学和科学沟通。
Ome Ogbru, PharmD, CEO and Founder of AINGENS
Trust Is the Real Question Can AI outputs be trusted? This is one of the most important questions healthcare professionals ask about generative AI. This question is especially important in medical information and medical writing because the value of a response depends on clarity and whether every statement is grounded in verifiable evidence and can withstand expert scrutiny. This matters because these workflows impact clinical judgment, scientific exchange, and regulated healthcare communication.
Where AI Is Gaining Ground
For clinicians, medical information often means finding and interpreting evidence for care decisions by reviewing literature, checking treatment guidelines, or synthesizing information from multiple sources to support a treatment decision. For pharmaceutical medical information teams, it means developing accurate, balanced, and traceable scientific responses that shape how healthcare professionals interpret data. Medical writing determines how evidence is summarized, contextualized, and communicated across research, scientific exchange, and regulated environments.
Generative AI is naturally attractive in all of these settings because the work is text-heavy, labor-intensive, time-consuming, and expensive to scale efficiently. AI solutions can speed literature discovery, create drafts, and help professionals move through medical communication workflows faster.
However, those benefits come with a meaningful risk of AI hallucinations that can affect how scientific information is consumed, influence care-related decisions, and contribute to misinformation if left unchecked.
The Clinician’s AI Trust Problem
For clinicians, the question is whether AI helps clarify evidence or makes uncertainty seem factual. Clinicians are asked to trust an output that cannot be properly reviewed if an AI system offers a concise summary, but the source of the conclusion is unclear. In a clinical decision workflow, that is not efficiency. It is hidden risk.
Clinicians need a system that helps them quickly find trusted evidence and shows where the information came from.
The Medical Information and Medical Writing AI Dilemma
The challenge is equally serious for pharmaceutical medical information teams and medical writers. These teams operate in highly regulated environments where accuracy, balance, and traceability are nonnegotiable.
AI can help by accelerating literature review, surfacing relevant evidence, and producing first drafts. But if a model confuses findings from different studies, overstates conclusions, omits context, or invents supporting references, the output becomes less useful and potentially risky.
The Solution Starts With System Design
Hallucinations do not make AI unusable. They make system design more important. AI systems for medical and scientific workflows should be evidence-first, include a robust retrieval pipeline, use strong models, and provide built-in source traceability so users can see exactly where statements came from. They should encourage users to point the platform to credible, vetted resources rather than relying on the model’s general training knowledge.
These systems should make review easier and help users inspect source material, edit outputs, recognize when a response is not based on evidence, and clearly say if credible source material is missing. The model should be able to respond plainly: I do not know, or I do not have enough information to answer credibly.
That kind of restraint and transparency is a strength, not a weakness of the AI platform.
Workflow Design Still Requires Human Expertise
Technology alone is not enough. Safe and effective AI use also depends on workflow design. Human responsibility includes defining the task, deciding whether AI should be used, selecting the right platform, setting guardrails, identifying credible knowledge sources, giving clear instructions, and reviewing and revising outputs. AI outputs should be treated as drafts, not final work, until a qualified human has reviewed them and accepted responsibility.
This is the difference between using AI as an automated shortcut and using it as a managed productivity tool.
User Training, Practice, and Expectations Matter
User training, practice, and a willingness to experiment with AI solutions are essential. Many users may feel that the cognitive load of learning a new system may not be worth the effort, especially if they have to review the outputs. As users become more familiar with a platform, the effort required to use it effectively decreases, while the productivity gain increases significantly.
There is a common assumption that reviewing AI outputs and refining them over several prompts eliminates the time savings. That comparison and conclusion are not logical. A human-written draft that takes days to prepare and still contains gaps or errors typically goes through multiple rounds of review before it is final. If an AI solution can help users find relevant literature in seconds and produce a workable draft in seconds, and the user then spends minutes reviewing and refining it, the net efficiency gain is still substantial.
Users should have realistic expectations about what AI solutions are capable of. One common mistake is assuming that AI models can function as experts in every domain. Trained professionals remain the experts, and the AI platform is a tool that supports them.
Successful use of AI in medical information, medical writing, and clinical workflows depends on the platform and users who understand its capabilities, limits, and assume responsibility for using it appropriately.
What Leaders Should Prioritize
The approach should be straightforward for decision-makers. Organizations should not evaluate AI for medical information, medical writing, or clinical workflows based primarily on speed, popularity, or ambitious claims. They should prioritize systems that are:
Evidence-grounded Transparent Auditable Built for the specific workflow Designed to support human review The system is only part of the equation. Leaders must equip users with the right training, provide support, set realistic expectations, establish governance, and ensure accountability for AI-supported work.
Trust Must Be Designed, Not Assumed
Hallucinations are not a reason to avoid AI in medical information and medical writing. When the right platform is paired with the right workflow, hallucinations can be minimized and caught before they affect the final output. The best approach is not blind trust in AI, and not blanket rejection of it, but trust built on evidence-first system design, governance, effective workflows, and human oversight and accountability.
About Ome Ogbru, PharmD Ome Ogbru, PharmD, is the CEO and Founder of AINGENS, a life sciences software company building evidence-first AI platforms for scientific and medical workflows. With over 20 years of experience across pharma, biotech, and healthcare, his background includes roles as a clinical pharmacist, professor, and global medical information leader, where he worked at the intersection of science, regulation, and content creation.
Driven by firsthand experience with the inefficiencies of evidence-based content workflows, Dr. Ogbru founded AINGENS to develop practical, enterprise-ready solutions that improve how scientific information is created, reviewed, and delivered. Through its flagship platform, MACg (Medical Affairs Content Generator), he focuses on enabling faster, more reliable medical and scientific communication without compromising accuracy or compliance.
Reader Interactions