Two and a half years since the first ChatGPT release, and researchers, developers, and business leaders are reluctantly coming to the same conclusions:
- When asked to do any kind of work that requires accessing “known knowledge,” LLMs (large language models) continue to have a high degree of hallucinations – simply inventing answers that just are not there (Mauran, 2025; Goldman, 2025),
- Nothing that we’ve done to make LLMs work has really been effective – they continue to fail when it comes to giving us “right answers,” and
- {* To be completed *}.
So the real and compelling question is: what’s next?
Actually, there are two questions:
- Why don’t LLMs work? And why CAN’T we make them work? (Despite the best minds that we can throw at this problem, and lots and lots and LOTS of money.)
- Given that we can’t make them work, what’s next?
This last question leads us in all KINDS of directions, ranging from:
{* Work in progress *}
References and Resources
Problems with LLMs
- Mauran, Cecily. 2025. “More Concise Chatbot Responses Tied to Increase in Hallucinations, Study Finds.” Mashable.com (May 11, 2025). (Accessed May 16, 2025; available at More concise chatbot responses tied to increase in hallucinations, study finds | Mashable.)
Problems with Chatbots and Agent-Based Systems
- Goldman, Sharon. 2025. “A Customer Support AI Went Rogue—and It’s a Warning for Every Company Considering Replacing Workers with Automation.” Fortune (April 19, 2025). (Accessed May 16, 2025; available at A customer support AI went rogue—and it’s a warning for every company considering replacing workers with automation; available also at MSN.com.)
Reinforcement Learning
- Tewari, Ambuj. 2025. “What is Reinforcement Learning? An AI Researcher Explains a Key Method of Teaching Machines.” Tech Xplore The Conversation (April 7, 2025). (Accessed May 16, 2025; available at What is reinforcement learning? An AI researcher explains a key method of teaching machines.)