Three California community colleges spend up to $500,000 yearly each on AI chatbots for student services. The chatbots don't work well. East LA College's bot couldn't name its own president. This reveals the AI promise gap.
The Failure
CalMatters reports chatbots handle general questions but struggle with specific ones. Students asking about their financial aid status or transfer requirements get wrong or vague answers. At $500K per institution, that money could pay for several full-time advisors who'd actually answer correctly.
Why They Fail
Training data mismatch. Fine-tuned on official docs, but students ask messy contextual questions. Specificity problem. "What is financial aid?" is easy. "I filed FAFSA late with a dependent, am I Cal Grant eligible?" requires reasoning across policies. No accountability. When humans err, there's pushback and correction. Chatbot errors persist unchecked. False confidence. Chatbots present wrong info with the same certainty as correct info.
What Would Work
Hybrid systems. AI for general questions, humans for specific ones. Retrieval-grounded responses. Present relevant documents rather than generated interpretations. Confidence calibration. Say "I'm not sure" when uncertain. Continuous evaluation. Track questions and accuracy; most deployments skip this.
The Cost Question
$500K hires 5 to 7 full-time advisors providing better service. Chatbots offer 24/7 availability, but wrong answers at midnight are worse than no answers. Students think they have answers and stop looking.
AI chatbots work in education when deployed honestly: clear limitations, human fallbacks, continuous improvement. These failures are about institutions buying without understanding limitations, and vendors selling capabilities that don't exist in practice.