Mitigating AI ethical issues in scientific process by supporting knowledge triangulation and AI limitation awareness

The integration of Artificial Intelligence (AI) tools into scientific communication in chemistry presents both opportunities and ethical challenges. AI tools – such as generative models, natural language processing algorithms, and automated data analysis tools – offer the potential to enhance the efficiency of literature reviews, facilitate data-driven insights, and accelerate innovation. However, their use also raises critical ethical concerns, particularly regarding transparency about how AI tools reach a conclusion or an outcome and the underlying bias of the models. Researchers and practitioners may rely on AI-generated outputs without fully understanding how conclusions were derived, which can undermine the reliability of scientific discourse. For example, we have been seeing reports on several problematic cases associated with how LLMs (Large Language Models) hallucinate in various scenarios. This is a well-known limitation for this kind of AI tool. We cannot ignore the potential AI tools have for the scientific process and the opportunities they may offer. Nevertheless, we need resources to deal with their ethical challenges and limitations. In a previous work, at the ACS Fall 2023, our group presented the Discovery Workbench, a framework which facilitates human-AI co-creation. This framework relies on a knowledge base containing domain and process knowledge, and user-interaction components to acquire knowledge and advise subject-matter experts (SMEs). In this work, we present how the workbench can support workflows with AI-based tasks and SME checkpoint tasks allowing triangulation of the AI tools outputs with domain and process knowledge. This workflow support provides resources for research traceability, which is crucial for scientific communication and transparency. In addition, we present user experience features that highlight where AI tools participated in the process and user interaction strategies to keep the SME aware of AI tools limitations not only when the tools are used, but also as an impact factor for the whole process.