By Joe Gladd. Part of MyEssayFeedback.ai's Critical AI Literacy series. Written for students with no machine learning background.
You'll learn: the difference between human writing and machine-generated text, how NLP works (tokenization and embedding), how LLMs are aligned (RLHF vs Constitutional AI), and the risks of bias, censorship, and hallucinations.
Human writing involves free writing, brainstorming, research, analysis, drafting, feedback, and revision — a complex overlay of voice, other voices, and pattern recognition. LLMs skip all of that. They rely entirely on pattern recognition — vastly accelerated. ChatGPT produces text that looks like a college essay because it excels at genre conventions, but lacks worldview, experience, or beliefs.
LLMs are trained on enormous text: Wikipedia, scholarly essays, Reddit, books. Tokenization converts text to numbers. As OpenAI explains, GPT models "understand the statistical relationships between tokens, and excel at producing the next token in a sequence." Embedding assigns probabilities to where words belong based on their distribution — language generates meaning by association.
Raw LLMs can produce toxic content. OpenAI uses Reinforcement Learning from Human Feedback (RLHF) — human trainers rank outputs and the model learns preferences. Anthropic's Constitutional AI trains the model to evaluate its own outputs against written principles, reducing reliance on human labelers.
Bias: LLMs reflect training data biases — political, linguistic (favoring English), cultural. Censorship: alignment can overcorrect, suppressing viewpoints. Hallucinations: LLMs generate confident fabrications because they predict likely words, not accurate ones. They invent fake sources, statistics, and facts in authoritative prose.
Your unique perspective is irreplaceable. Overreliance on AI dilutes your voice. AI is a tool, not a substitute. Distinguish what LLMs do well (idea generation, pattern matching) from what they cannot do (judge quality, understand your purpose). See our template phrases for reflecting on AI feedback.