Anyone trying to integrate Generative AI using Large Language Models into some commercial or professional business process should understand the dangers of so-called hallucinations. Because LLMs are generating the next text, given a corpus of training text and prompt data, all activity by these models is the same – prediction of the most highly likely completion text. Some outputs may appear to be hallucinatory only to those of us tethered to a real world. Anyone untethered – and this includes the LLMs themselves – will not be able to distinguish so-called real from so-called hallucinations, just as a sleeper cannot tell if his or her dreams are plausible or not.

Eventually, with clever design of wrapper processes by and including humans, we will be able to deal with these so-called hallucinations. I was reminded of the situation with the invention of printed books by Gutenberg. For a century or so after the invention of printing, printed books were, on the whole, much less accurate than hand-copied manuscripts. This is because hand-copiers, recognizing the potential for errors, had created sophisticated processes for checking and proof-reading hand-copied manuscripts. It took the nascent printing industry 100 years or so to re-invent similar wrapping processes to ensure comparable levels of accuracy for printed books.

Computer Scientists have done something similar before, with the creation of the TCP Protocol: Designing a portfolio of processes to ensure the reliability and accuracy of messages transmitted over the Internet, in the face of and despite the unreliability of the components transmitting those messages. The task with TCP was syntactical, whereas the task with LLMs is generally semantic, so it will be more difficult to devise effective wrapper processes. LLM Reliability will likely have to to be tackled domain by domain. What is dangerous fantasy in, say, the law, may well be fruitful scenario planning in business strategy development.