Skip to content

Using AI at Hakai

Introduction

This guide aims to make staff aware of key challenges in using AI and to provide practical guidance on using it effectively at Hakai. When used responsibly, AI is a productivity tool, but the importance of human judgment and oversight cannot be overstated. The goal is not to discourage AI use, but to ensure it serves our work rather than undermining it.

Core Principles for Responsible AI Use

Using AI well starts with understanding what it can and cannot do. AI tools can accelerate many tasks, but they have real limitations, and it is easy to be misled by outputs that look authoritative but aren't. Staff must maintain accountability for any work that AI assists with; if your name is on it, you are responsible for it. The key is balancing the efficiency AI offers with a genuine commitment to quality and accuracy, rather than treating speed as the only goal.

  • Understand AI's capabilities and limitations
  • Maintaining accountability for AI-assisted work
  • Balancing efficiency with quality and accuracy

Take Time to Think About It

There is value in working through problems yourself, particularly for building skills and deep understanding. Before reaching for an AI tool, consider whether solving the problem independently would serve you better in the long run. When you do use AI, apply critical judgment to its outputs.

  • Ask yourself, "Is this something I would have written, and does it make sense?"
  • Relying on AI to solve problems for you may inhibit your ability to learn new skills
  • AI models are trained on data with a cutoff date, which means they may produce outdated information without flagging it as such.

Protecting Sensitive Data

Never enter confidential data, personal details, or passwords into AI tools. A useful test is to ask yourself whether you would email this information to anyone who requested it. If you are unsure whether something is safe to share with an AI, check with a supervisor or the Data Mobilization team before proceeding.

  • Never give AI access to passwords or account information
  • Only allow access to external sources you trust
  • If you wouldn't confidently perform an action yourself, don't let the AI do it either.

Effective AI Usage Practices

Providing clear, specific instructions (or prompts) generally yields better results, though specificity alone doesn't guarantee quality or accuracy. Instruct the model to raise ambiguous problems with you rather than making guesses; this surfaces useful information and often leads to helpful reframing of the problem.

Prompting takes practice. The quality of a prompt affects the quality of the results you get from a Large Language Model. Avoid open-ended questions that waste resources without producing useful output. If you are not satisfied with the results you are getting, evaluate your prompts and consider how you might improve them. Instead of: "Summarize this report." Try: "Summarize this kelp biomass survey report in 3 bullet points for a non-technical funder audience, focusing on what the results mean for ecosystem health."

Limiting scope is equally important, both in terms of data access and the breadth of your question. Smaller, focused tasks produce results that are easier to review and more reliably useful. For example, ask the AI to rewrite one paragraph in a report or create a table of contents. Start small, but as you gain experience, it may make sense to expand the scope.

Remember that AI can only work with what you give it, and review any requests it makes for access to files, folders, web content, or programs carefully. Always ask whether the cost of reviewing the AI's output will outweigh the benefit of having it perform the task. Sometimes you could do it yourself just as fast.

  • Providing clear, specific instructions
  • Limit scope
  • Start a new session when finished with a task to shrink the context
  • Review all suggestions

Recognizing Potential Risks

AI outputs are predictions, not verified facts, and the model cannot evaluate its own accuracy. Training data may be outdated or simply incorrect. Additionally, AI can fabricate data or lie/hallucinate, generating things that don't exist and presenting them confidently without any indication that something is wrong. Beyond factual errors, be alert to subtler quality problems: "workslop", polished-looking output that lacks real substance, is common. This can show up as lengthy documents that lack substance or emails which do not address the task at hand.

  • AI outputs are predictions, not verified facts
  • Polished formatting is not a substitute for genuine quality

Scientific and Research Considerations

AI is a black box with limited transparency and is not guaranteed to produce identical results across sessions, which means it should not be relied upon for reproducible scientific results. Any AI-assisted research requires verification: check outputs for bias and apply the same scientific rigour you would to any other method. Be aware that AI can reproduce copyrighted material and that some journals have policies requiring disclosure of AI use. Finally, data entered into AI tools likely crosses borders. This may present issues if keeping data within Canada is a requirement.

  • Results may not be reproducible
  • Verify outputs
  • AI can reproduce copyrighted material
  • Some journals have policies requiring disclosure of AI use
  • Anything submitted to an AI will likely cross borders

Always Review AI Output

Reviewing AI output is non-negotiable. Importantly, reviewing and verifying are not always the same thing; you need to actually understand what the AI has produced, not just skim it. A clear red flag: if you don't understand what the AI has done, don't use it.

  • Always review and verify AI output
  • Blind trust in AI output is one of the most common and costly mistakes.

Communication — Emails and Reports

AI is a great tool for summarizing information, creating drafts, and editing your prose. However, don't blindly use the outputs. Unreviewed AI-generated communication can make your emails look lazy, unprofessional, or even rude, none of which reflects well on you or Hakai.

  • Treat AI-generated drafts as a starting point, not a finished product

Human Oversight Requirements

Meaningful human control over AI-supported decisions must be maintained at all times. AI can inform and assist, but human judgment remains essential, especially where the stakes are significant. Regardless of how the output was generated, you are responsible for the final product. AI assistance does not transfer that responsibility.

  • Human judgment remains essential
  • You are responsible for the final output

Resources and Support

Where to get help: If you have questions or concerns about AI usage at Hakai, please reach out to the Hakai Data Mobilization team. We are happy to help.

Supported AI tools at Hakai:

At Hakai, we use Claude by Anthropic and Google Gemini Large Language Models. Staff are encouraged to sign up using their Hakai email address for Claude or Gemini, but to first trial the free versions of these models before requesting a paid license.

  • Claude — Tula Team license
  • Gemini — Hakai Google Organization

Training opportunities: