AI at Red Rocks

A type of AI where systems learn from large amounts of data rather than being explicitly programmed with rules. The more data the system trains on, the better it becomes at recognizing patterns and making predictions. 

A type of machine learning model trained on vast amounts of text. LLMs generate responses by predicting what words are most likely to follow based on patterns in their training data — not by "understanding" in the human sense. 

AI systems that can produce new content — text, images, audio, code, video — rather than just classifying or analyzing existing content. Most AI tools students and faculty encounter today are generative AI systems. 

The input you give an AI system — a question, instruction, or piece of text that tells the AI what you want it to do. The quality of a prompt significantly affects the quality of the output. 

Expand any term to learn more. This glossary covers vocabulary you are likely to encounter in conversations about AI at RRCC and beyond.

Hallucination
When an AI model generates information that is fabricated or nonsensical — but presents it confidently as if it were accurate. The term specifically refers to content the model invents, such as made-up citations, non-existent sources or plausible-sounding but false claims. It does not cover cases where a model accurately repeats incorrect information that was present in its training data — that is a separate (and also real) limitation.

This is one of the most important things to understand about AI tools: they do not "know" whether their output is true. Always verify AI-generated claims against credible sources.

Training Data
The large collection of text (or other data) that an AI model learns from during development. LLMs are typically trained on billions of words drawn from the internet, books and other sources. The content and quality of training data directly shapes what the model knows, what biases it may carry and what it tends to produce.

Training data has a cutoff date, which means AI models may not have information about recent events or developments.

Bias in AI
AI systems can reflect and amplify biases present in their training data — including biases related to race, gender, culture, language and socioeconomic status. This can affect the quality and fairness of AI outputs in ways that are not always obvious.

Bias can also be introduced directly by the humans who design, train and fine-tune these models — through decisions about what data to include or exclude, how to define desired outputs and what feedback is used to shape the model's behavior. This means bias in AI is not only a data problem but also a human and institutional one.

Being aware of AI bias is part of using these tools critically and responsibly, especially in academic, research and professional contexts.

Prompt Engineering
The practice of crafting effective inputs (prompts) to get better outputs from an AI system. Good prompts are typically clear, specific and provide relevant context. Prompt engineering has become a recognized skill in both academic and professional settings.

For example, "Summarize this article in three bullet points for a non-specialist audience" is likely to produce a more useful result than "Summarize this."

Generative vs. Analytical AI
Generative AI creates new content — text, images, code, audio — based on a prompt. Examples include ChatGPT, DALL-E, and GitHub Copilot.

Analytical AI analyzes existing data to identify patterns, make predictions, or classify information. Examples include spam filters, recommendation algorithms and medical image analysis tools. Both types are widely used, but most of the current conversation in academic settings focuses on generative AI.

AI vs. Automation
Automation follows fixed rules to complete repetitive tasks — a calculator, a spell checker or a scheduling script. AI, by contrast, involves systems that can adapt, learn and generate outputs that weren't explicitly programmed.

The distinction matters because AI outputs are probabilistic and variable — the same prompt can produce different responses — while automated systems produce consistent, deterministic results.

Intellectual Property & AI
AI-generated content raises unresolved questions about copyright and intellectual property. In most jurisdictions, AI-generated content is not currently eligible for copyright protection because it lacks a human author. The legal landscape is evolving rapidly.

For academic purposes, the more immediate concern is attribution — being transparent about when and how AI contributed to your work. 

It depends on the tool and your account type. Free tiers of many major AI platforms — including some versions of ChatGPT, Google Gemini and others — have historically used conversation data to improve their models by default. This means your prompts and the AI's responses could become part of future training data.

Paid subscriptions often include stronger data protections, sometimes guaranteeing that your data will not be used for training. However, you are taking the company at its word — and several major AI companies are currently facing legal challenges related to how they sourced and used data for training. Review the privacy policy and terms of service for any tool you use regularly. 

Potentially, yes. LLMs are increasingly capable of inferring personal characteristics — location, profession, education level, first language, age — from writing style and content alone, even when no identifying information is directly provided. This process, sometimes called deanonymization, means that the content of your prompts can itself be identifying, independent of any account information.

Additionally, your IP address, device fingerprint and browser characteristics are typically collected even on sessions where you are not logged in. True anonymity on major commercial AI platforms is difficult to achieve. 

Some tools — such as DuckDuckGo's duck.ai — route your prompts through a proxy that strips identifying information before passing the request to the underlying AI model. This provides meaningful protection for your identity: the AI company may not know who is sending the prompt.

However, this does not protect the content of the prompt itself. The information you type still reaches the AI provider's servers. If your prompt contains sensitive personal, student or institutional data, a privacy proxy does not prevent that data from being processed or potentially retained by the model provider.

This kind of tool is a reasonable step toward identity protection, but should not be treated as comprehensive data protection. 

FERPA (the Family Educational Rights and Privacy Act) protects the privacy of student education records. It applies to RRCC and restricts how student information can be shared with third parties.

When a faculty or staff member submits student work, grades, names, ID numbers or other identifying information to a commercial AI tool, they may be sharing FERPA-protected data with a third party that does not have a formal data-sharing agreement with RRCC. This could constitute a FERPA violation.

The safest approach: do not enter student names, ID numbers, grades or any information that could identify a specific student into a commercial AI tool without first confirming that RRCC has an appropriate data-use agreement in place with that provider. Contact IT Services or the Office of Academic Affairs with questions. 

Institutional data that should not be entered into commercial AI tools without IT authorization includes:

• Student names, student ID numbers, grades or academic records
• Employee personnel records, compensation details or HR information
• Internal financial data, budget documents or procurement information
• Unpublished research data or proprietary institutional information
• Any data covered by a confidentiality agreement
• Medical or health-related information (HIPAA-protected data)

When in doubt, treat the information as sensitive and consult IT before proceeding. 

Many commercial AI platforms offer paid tiers that disable training on your data and may include enterprise-level data retention policies. This is a meaningful step up from free tiers.

The limitation is that you are relying on the company's contractual commitments. The AI industry has a complicated track record on data use and several major providers are currently involved in litigation over training data practices. Read the terms carefully and understand what you are agreeing to.

Appropriate for: general-purpose use with non-sensitive data where you want additional assurance. Not sufficient for student data or other FERPA/HIPAA-protected information without a formal institutional agreement. 

Tools like DuckDuckGo's duck.ai route your prompts through an anonymizing proxy before sending them to the underlying model. This helps conceal your identity from the AI provider, since the request does not include your account credentials or IP address.

The important caveat: the content of your prompt still reaches the AI provider. If your prompt itself contains sensitive data, a proxy does not protect that data — only your identity. LLMs are also capable of inferring identity from writing patterns even without direct identifiers.

Appropriate for: general tasks where identity protection matters but prompt content is not sensitive. 

Open-weight AI models (such as Meta's Llama series or Mistral) can be downloaded and run entirely on your own computer using software like Jan or Ollama. When running locally, your data never leaves your device — there is no server, no company receiving your prompts, and no risk of third-party data retention.

The tradeoff is hardware. Most of the processing happens on the GPU. A powerful, modern computer will handle many tasks well; lower-end machines may be limited to simpler tasks like brainstorming, drafting and basic Q&A. Computationally intensive tasks like large-scale coding or image generation may be slow or impractical.

Appropriate for: any situation where prompt content is sensitive and data must not leave your device. The best option for working with institutional data until formal RRCC-approved tools are in place.