Ownership and Access
When evaluating an AI tool, it is important to know who owns or developed it, because this affects how it is maintained, updated, and regulated. Access and usage rights also matter – some tools are proprietary and controlled by private companies, while others are open-source and freely available for adaptation and research. Ownership and access shape not only how the tool can be used but also what can be done with the data provided as prompts or its outputs, and whether users can inspect, export, or reuse the underlying datasets. Understanding the ownership and access model helps researchers and users see the tool’s flexibility, transparency, data control, and possible limitations.

Key questions to ask:
• Who owns or developed the AI tool?
• Who can access or use it?
• Is it proprietary or open-source?
• Can users export data, outputs, or models?
• Can users see or reuse the training datasets?

Transparency
Because most AI tools are trained on large datasets, it is important for users to question how the information was gathered, cleaned, and processed. However, most AI companies do not disclose details about their training data or and models which often raises legal debates around the potential improper use of copyrighted material.

Key questions to ask:
• How trustworthy is the information about the AI tool/technology?
• How clearly is the data provenance documented?
• Where does the training data come from, and how was it collected?
• How transparent are the model training processes?
• What methods were used, and how were decisions made during training?
• How was the model evaluated and tested?

Bias
The lack of transparency in these AI tools contributes to biased outputs. Because the training data is often not publicly accessible, it is impossible to know which types of content were ingested or which perspectives have been encoded into the model. Bias also arises from the methods used to process the data, often amplifying majority viewpoints while marginalising minority ones. As a result, many AI tools can perpetuate content that is toxic, inaccurate, or offensive, particularly in relation to vulnerable communities and groups.

Key questions to ask:
• Are the biases present in the training data and/or model development well documented and transparent?
• Does the documentation provide details on the type of bias, how it was detected, and any mitigation strategies applied?
• How might these biases affect the outputs or decisions made by the AI tool?

Reliability and Reproducibility
AI tools can be very useful for combining information from many sources, but their results are not always accurate or reliable. One common problem, especially with generative AI, is “hallucination,” where the AI produces information that sounds plausible but is wrong. The trustworthiness of AI is also affected by the general failure to provide clear explanations for its decisions or outputs. Without transparency about how the models process data and reach conclusions, users cannot fully assess the accuracy or validity of the results. If the training data is incomplete, outdated, or biased, the AI’s outputs will show the same problems.

Commercial models are often hosted on distant servers and only accessible through API calls.As a researcher, you have no control over access to these models, which poses significant reproducibility problems. Once these models are taken out of production, it becomes impossible to reproduce the research results. For example, it is very difficult, if not impossible, to cite or reproduce outputs from a GenAI model. Unlike traditional sources, generative AI produces responses dynamically, which means the same prompt can yield different answers at different times. This lack of consistency raises challenges for verification and academic referencing.

There is no clear-cut solution to these problems. Possible ways to obviate these issues is to document properly which model is used and when, save both the inputs (prompts) and outputs (the response of the model also called prompt completion) or use open-source models.

Key questions to ask:
• How accurate and reliable are the AI outputs?
• How there transparency about the data and methods used to train the AI tool?
• How can I access the model (through API or can I run it locally on my computer)
• Is there sufficient documentation for the AI tool that provides transparency about the datasets, model architecture, training procedures, and guidance on how to cite or reference its outputs?

Policy compliance and data security
Before adopting any AI tool, researchers need to consider not just what an AI tool can do, but also the wider risks it might bring. This includes institutional policies in place, such as the ones stated at University of London policy, legal frameworks, data compliance and information security, and even environmental impact.

It’s essential to ensure the tool keeps sensitive information safe and does not cause harm or create risk if people rely on it for critical decisions without proper checks. Of course, legal frameworks around AI are still evolving and this remains an area of ongoing development and scrutiny. This means that what is considered compliant or best practice today may change in the near future and researchers need to stay informed and adapt their practices accordingly.

Key questions to ask:
• Does the tool handle sensitive or personal data securely and in compliance with GDPR/UK Data Protection laws?
• Does it comply with wider institutional or funder (data) policies?
• Do the tool’s terms and conditions raise any issues concerning the Intellectual Property rights vested in both inputs and outputs?
• Could entered data be exposed, shared, or misused?
• Can it produce harmful, offensive, or misleading content?
• Could reliance on it cause errors or harm in real-world use?
• Is the particular tool the most environmentally sustainable option available?