Introduction

This guide describes my responsible approach to using AI tools. It is intended for anyone who uses AI directly through chat or an API, or indirectly through a third-party tool.

Using AI responsibly means being transparent about its use, overseeing its output, protecting sensitive and personal information, avoiding harm to individuals and groups, and being mindful of the environmental and social cost.

Overview

AI large language models (LLMs) generate text by predicting the most likely next word based on patterns learned from large volumes of training data. LLMs approximate reasoning statistically and do not understand context in the way humans do. Its frequently confident output is not evidence that AI is intelligent or accurate.

AI models reflect training data bias. Training data is typically scraped from large volumes of public licensed content, which results in the model inheriting factual errors, biases, and data gaps. Many AI vendors have also been accused of using pirated content for training, introducing additional ethical concerns.

Models are trained on data up to a fixed point in time, so its knowledge has a cutoff date and may not reflect recent developments. Different model versions can produce meaningfully different outputs. Users should verify which model they are using and treat any time-sensitive information with skepticism.

LLM Platform Types

Each vendor implements its LLM platform differently, and these choices can affect privacy, data security, cost, environmental impact, and the degree of control you have over your data and the model itself.

Type	Description	Examples	Key Considerations
Closed-source cloud	Vendor hosted proprietary models	ChatGPT, Claude, Gemini	Privacy depends on tier; prompts may be used for training
Open-source cloud	Third party hosted open-weight models	Llama, Mistral	Model is inspectable; hosting provider has access to your data
Open-source local	Models run locally	Ollama, LM Studio	Maximum privacy; no data leaves your device; limited by local compute
Enterprise / private deployment	Vendor-hosted with contractual data protections	Claude Enterprise, Azure OpenAI	Stronger privacy guarantees; higher cost; suitable for sensitive data
API access	Direct programmatic access to models	Anthropic API, OpenAI API	Generally stronger privacy than consumer tiers; short retention

When choosing a platform, consider:

• Platform safety: Does the provider publish transparency reports, pre-release “red team” testing results, or usage policies?
• Data privacy: Are prompts used for training? What is the data retention period? Is opt-out available?
• Training data sourcing: Has the provider disclosed what data was used to train the model and its licensing status? Is there evidence that the platform vendor used pirated or stolen data for training?
• Energy use: Does the provider publish emissions data or use renewable energy? 
• Open-weight vs closed-source: Open-weight models allow independent inspection of trained parameters and local deployment; closed models offer less transparency but more capability.
• Local vs cloud: Local models offer maximum privacy but require capable hardware and technical setup.

Responsible Use

Using AI as a Research Tool

Using AI is appropriate at various stages of drafting and research, including generating initial outlines, summarizing sources, exploring ideas, coding, or producing a first draft, but should not produce the final product. LLMs, used thoughtfully, are a time-saving research tool but frequently insert incorrect data in its output. Responsible use of AI as a research tool include:

• Asking the model to provide a reference for every new claim in its output
• Following links to reference material and read it in context of claims being made
• If a reference seems incorrect, asking the model what point it is supporting and why it was chosen
• Independently cross-checking references against primary sources before citing them
• Confirming all statistics, dates, names, or technical claims

Data Privacy

Some LLMs, including many free tiers, use prompts and output for training. This means that information you provide may be used in output for the platform’s other users.

• Do not enter personally identifiable information (PII), proprietary, confidential, or legally protected data into AI systems unless the platform is locally hosted
• Treat all interactions with cloud-based AI as potentially logged, reviewed, or retained
• Read and understand the platform’s privacy and data handling policies.

Vulnerable Users

Some users face heightened risks when interacting with AI:

• Supervise children’s use of AI. AI systems are not vetted for child safety without controls in place
• AI should supplement, not replace, human contact and support for the elderly
• AI is not a substitute for clinical care for people with mental health conditions and may produce destabilizing responses
• AI should not be a first point of contact for people in crisis situations

Resource Use

Every AI query consumes resources, including water and energy. Some vendors have built their own electric generators in residential areas that run on polluting fossil fuels. Longer, less-focused prompts consume more energy than a standard web search. While this cost is negligible at an individual user level, it is significant when scaled across millions of users and billions of prompts. Mindful, efficient prompting is a responsible practice.

• Unnecessary filler phrases, like “Please,” “Thank you,” “How are you?” and similar social language, have no effect on output quality and consume tokens for no purpose
• Vague prompts generate longer, less useful responses
• Provide only the relevant portion of a document, not the entire file
• Set output format and length expectations (e.g., “respond in bullet points, under 200 words”)
• Batch related sub-tasks into a single prompt
• Avoid regenerating responses unnecessarily

Task-Appropriate Models

Not every task requires the most powerful available model. Using a smaller or faster model for simple tasks reduces cost and energy use without meaningful loss of quality.

Task Type	Recommended Approach
Simple Q&A, summarization, drafting	Use a smaller or faster model (eg, Claude Haiku, GPT-5.4 mini, Gemini Flash, Llama 3.3)
Complex reasoning, code generation, research	Use a larger model where quality matters (eg, Claude Opus, OpenAI GPT-5.4, Gemini 3.1 Pro)
Repeated or automated tasks	Use API access with a small model
Sensitive or private data	Use local models or enterprise-tiers

Bias

AI models can reflect and amplify the biases present in their training data. This includes:

• Defaulting to majority-culture frames of reference reflects demographic bias
• Underrepresenting languages, cultures, or communities introduces gaps in generated content 
• Names or demographic signals in a prompt may affect output quality
• Political viewpoints in training data may bias output

Asking a model to “be neutral” or “avoid bias” does not reliably work since bias is embedded in the model’s weights and cannot be overridden by a prompt. Users should treat bias mitigation as an ongoing review practice, not a one-time instruction.

Effect on Labor

AI systems are already displacing certain categories of work, including routine writing, image generation, customer support, data entry, and aspects of software development. Ethical dimensions include:

• Whether AI-generated content fairly compensates human creators whose work was used to train the model
• How organizations manage the transition for workers whose roles are affected
• Whether AI augments human capability or primarily serves to reduce headcount and labor costs

Conclusion

Responsible AI Use in Practice

Using AI responsibly is an ongoing practice of disclosure, verification, critical review, and informed choice. The most important habit is staying alert to what AI is producing, how it is being used, and who might be affected.

Evolving AI Landscape

AI capabilities, platforms, and policies change rapidly. What is true of a model or provider today may not be true in six months. AI platforms will address some ethical concerns as they inevitably become more energy efficient and regulated. Ethical users will stay informed by: 

• Following provider transparency reports, privacy policies, and terms of service updates
• Revisiting internal policies on AI use at least twice a year
• Monitoring regulatory developments in your jurisdiction
• Treating this guide as a living document subject to revision