A.I. Is Getting More Powerful, but Its Hallucinations Are Getting Worse

[John F Sowa]

Following is a note I sent to another email list, and I thought it would also be useful for the Ontolog and CG lists.  In fact, the three links below are to talks and slides that had been discussed on these lists.  But a reminder may be useful, since these issues are still hot topics.

John

 

---

Katie,

That point about hallucinations is true for AI systems that are based ONLY on LLMs.   But many AI systems have been producing excellent results based on the 60+ years of symbolic AI that reached a high level of sophistication long before LLMs were invented.

A combination of LLMs + symbolic AI can support the best of both worlds.   There are quite a few such systems, but the news about them is drowned out by people who are promoting LLM technology by itself without using symbolic (AKA logic-based) methods.

After my 30 years with IBM, I have been a cofounder of two AI startup companies:  VivoMind LLC from 2000 to 2010 and Permion Inc. more recently.  For the slides of a talk that describes that technology and several very powerful applications, see the slides for "Cognitive Memory for Language, Learning, and Reasoning,"      https://www.jfsowa.com/talks/cogmem.pdf .

For three sample applications, skip to slide 44.  None of those applications (and many others) could be implemented by LLMs alone.  That is because all three of them require high precision, absolute accuracy, and no hallucinations.  All of them accept natural language input in English and several other languages -- the usual European languages + Russian, Chinese, and Arabic + artificial languages such as names and notations for organic chemistry, computer notatioTns, and any specialized notations for which a grammar can be written.   See the examples of the applications starting at slide 44.

None of those examples could be implemented with LLMs.  But our new Permion Inc. company combines LLMs with an upgraded version of the technology of LLMs.  It can do the kinds of applications that use LLMs, the kinds of applications that used Cognitive Memory, and applications that combine both.

Most importantly, Permion avoids the hallucinations by using LLMs ONLY for two purposes:  (1) translation from one language to another (either or both may be natural or artificial); and (2) abduction (educated guessing),   Abduction is the most creative result of LLMs, but it is also the process that produces hallucinations (stupid guesses).  The reasoning methods of deduction and induction (the great strength of the symbolic methods) can detect and correct the stupid guesses that are often caused by LLMs.

For a YouTube talk that I presented in 2024, see "Without Ontology, LLMs are  Clueless."      https://www.youtube.com/watch?v=t7wZbbISdyA .

That talk got more than 10K downloads.  For another talk with my colleague Arun Majumdar in 2025, see https://www.youtube.com/watch?v=zRYJE6QJZx0&t=45s ,  Arun discusses the ways our methods for using LLMs are similar and different from methods used by DeepSeek.

In the Q/A discussion at the end of this last talk, Arun and other participants discuss issues of human interpretation that won't be found by LLMs or by a search of data on the WWW.  Humans are not obsolete.

John

---

From: "Katie Byl" <kati...@gmail.com>

Any thoughts / viewpoints on this, from expert(s)?

https://www.nytimes.com/2025/05/05/technology/ai-hallucinations-chatgpt-google.html?unlocked_article_code=1.E08.cvbT.1y7bN1jyxrzi&smid=url-share
(e.g., apparently, AI hallucination rates are actually increasing, despite kind of overall "seeming better" all the time? I guess, more or less?)

Anecdotally, I recently feel as if my responses (e.g., "are you sure?" vs "oh makes sense!" vs "wait but what about x?") has more impact than it once did, on the "choose your own adventure" that is a conversation with some LLM-based agent.  (But perhaps I am hallucinating there!)

-Katie

[Ravi Sharma]

John

Many thanks for the overview and deeper observations.

What happens to LLMs accuracy and is hallucination greatly reduced if we limit the store of LLMs by focussing on predetermined related info filters of some sort, somewhat like what Arun showed in Architecture figure during your talk.

I think hallucination happens more often when logic on clusters or learning sets is not able to be decisive.

Thus I think private agents based solutions in AI are emerging which are double edged swords hiding knowledge.

Kindly comment on how open source AI can stay effective with less hallucination, and thus handle inherent ambiguities with fact checks builtin to reduce errors?

Also when will you include Hindi as a language also if not already? to reach or use 500M people's Knowledge base of course much in Sanskrit.

Regards,

Thanks.

Ravi

(Dr. Ravi Sharma, Ph.D. USA)

NASA Apollo Achievement Award

Former Scientific Secretary iSRO HQ

Ontolog Board of Trustees

Particle and Space Physics

Senior Enterprise Architect

SAE Fuel Cell Standards Member

--
All contributions to this forum are covered by an open-source license.
For information about the wiki, the license, and how to subscribe or
unsubscribe to the forum, see http://ontologforum.org/info
---
You received this message because you are subscribed to the Google Groups "ontolog-forum" group.
To unsubscribe from this group and stop receiving emails from it, send an email to ontolog-foru...@googlegroups.com.
To view this discussion visit https://groups.google.com/d/msgid/ontolog-forum/27f45d2c5c544fce83069e96dcc72a43%400141aca7e38c4fc69d65085ba40edfdd.

[John F Sowa]

Ravi,

Combining LLMs with methods for evaluating the results requires a major amount of work.  Our Permion system was able to do that because the foundation had been developed about a decade ago -- and that requires many years of work.

To answer your question about Hindi:  Each new language requires a great deal of effort.   Our Permion company would add it if and when somebody wants to (1) do it or (2) pay for it..

John

 

---

From: "Ravi Sharma" <drravi...@gmail.com>