What was a key limitation of early language models that focused on word prediction?

What was a key limitation of early language models that focused on word prediction?

• A. High accuracy and alignment to human-like behavior
• B. Generated only code
• C. Memorized entire documents
• D. Word prediction without ensuring accuracy or alignment to human-like behavior

Answer: (D)

The main idea is that predicting the next word is not the same as guaranteeing truth or safe, human-aligned behavior. Early language models were trained to choose the most likely next word to make the text read fluently and coherently. This objective yields impressive-sounding text, but it doesn’t require the output to be accurate, reliable, or aligned with human goals or values. So they can generate convincing yet incorrect information, or produce content that isn’t appropriate or safe, because the training signal is about language patterns, not factual correctness or ethical considerations. That’s why the best answer highlights a limitation specific to word prediction: without a mechanism to ensure accuracy or alignment to human preferences, the models may produce plausible but fallible results. The other statements don’t capture this fundamental constraint—early models aren’t restricted to code, they don’t inherently memorize entire documents as their defining limit, and though memorization can occur, the core issue for word-prediction-based models is the gap between fluent generation and verified, aligned output.

The main idea is that predicting the next word is not the same as guaranteeing truth or safe, human-aligned behavior. Early language models were trained to choose the most likely next word to make the text read fluently and coherently. This objective yields impressive-sounding text, but it doesn’t require the output to be accurate, reliable, or aligned with human goals or values. So they can generate convincing yet incorrect information, or produce content that isn’t appropriate or safe, because the training signal is about language patterns, not factual correctness or ethical considerations.

That’s why the best answer highlights a limitation specific to word prediction: without a mechanism to ensure accuracy or alignment to human preferences, the models may produce plausible but fallible results. The other statements don’t capture this fundamental constraint—early models aren’t restricted to code, they don’t inherently memorize entire documents as their defining limit, and though memorization can occur, the core issue for word-prediction-based models is the gap between fluent generation and verified, aligned output.