What currently passes for AI - large language model chat bots - simply find likely patterns of words. They do not apply reasoning nor do they maintain an internal model of reality. I participate in a Parkinson's patients' forum and find that I have to regularly correct wrong AI results. All examples given here are from posts or comments that people actually made, mistakenly accepting a false result given by AI. Before I continue, a bit of background is in order. Per Wikipedia:
Decarboxylation is a chemical reaction that removes a carboxyl group and releases carbon dioxide (CO2)... Enzymes that catalyze decarboxylations are called decarboxylases
First example, as reported by the user:
I asked Gemini AI: Does Ceylon cinnamon inhibit decarboxylation?
AI's answer: “Yes, certain compounds found in Ceylon Cinnamon, particularly cinnamaldehyde, have been shown to inhibit amino acid decarboxylase activity.”
Gemini gave a reference. Here is everything the reference has to say regarding cinnamaldehyde and decarboxylation:
As observed with eugenol, the possibility that the mechanism of cinnamaldehyde activity involves inhibition of cell wall synthesis (2.36 mM, B. cereus) (13) or inhibition of biosynthetic enzymes (>7.5 mM, histidine decarboxylase) (29) is unlikely because of the rapidity of ATP inhibition or depletion.[emphasis added]
Gemini did not understand that the "is unlikely" negates what preceded it. This is what you get from a system that does not understand or reason - “is unlikely" is merely another text string.
Unfortunately reporting on AI is made difficult by the fact that it is a moving target. Gemini has changed its output and now says there is a lack of evidence on the subject. See below for a discussion of the results of most current attempts to improve AI.
What ensued in the comments on this post was much of AI boosting by AI fans, which culminated in a comment that included:
Antibacterial effects: Research by Zainal-Abidin et al. (2013) showed that cinnamaldehyde in cinnamon bark oil inhibits amino acid decarboxylase activity in oral bacteria, contributing to its antimicrobial properties. This is because decarboxylases are involved in bacterial amino acid metabolism and survival. [emphasis added]
Evidently the poster of the comment did not check the Gemini provided reference, entitled: Anti-Bacterial Activity of Cinnamon Oil on Oral Pathogens, by Zainal-Abidin et al. It can be found here. It is a very fine paper, and I was pleased to discover that cinnamon has antibacterial properties against caries causing bacteria. The only problem with this reference; there was no discussion of decarboxylase activity. This study did not say what Gemini claimed it did. This is like the episode of How to Use ChatGPT to Ruin Your Legal Career, where AI made up case law, much to the regret of the attorneys who submitted it.
So the AI boosters, in their attempt to tout the benefits of AI, demonstrate again that AI gives wrong answers. Plus, those who should know better often fail to check, even at risk of peak embarrassment.
AI is getting worse instead of better
The more sophisticated AI models get, the more likely they are to lie:
AI models are not really intelligent, not in a human sense of the word. They don’t know why something is rewarded and something else is flagged; all they are doing is optimizing their performance to maximize reward and minimize red flags. When incorrect answers were flagged, getting better at giving correct answers was one way to optimize things. The problem was getting better at hiding incompetence worked just as well. Human supervisors simply didn’t flag wrong answers that appeared good and coherent enough to them. In other words, if a human didn’t know whether an answer was correct, they wouldn’t be able to penalize wrong but convincing-sounding answers.
Schellaert’s team looked into three major families of modern LLMs: Open AI’s ChatGPT, the LLaMA series developed by Meta, and BLOOM suite made by BigScience. They found what’s called ultracrepidarianism, the tendency to give opinions on matters we know nothing about. It started to appear in the AIs as a consequence of increasing scale, but it was predictably linear, growing with the amount of training data, in all of them. Supervised feedback “had a worse, more extreme effect,” Schellaert says. The first model in the GPT family that almost completely stopped avoiding questions it didn’t have the answers to was text-davinci-003. It was also the first GPT model trained with reinforcement learning from human feedback....
Instead, in more recent versions of the AIs, the evasive “I don’t know” responses were increasingly replaced with incorrect ones. And due to supervised training used in later generations, the AIs developed the ability to sell those incorrect answers quite convincingly. Out of the three LLM families Schellaert’s team tested, BLOOM and Meta’s LLaMA have released the same versions of their models with and without supervised learning. In both cases, supervised learning resulted in the higher number of correct answers, but also in a higher number of incorrect answers and reduced avoidance. The more difficult the question and the more advanced model you use, the more likely you are to get well-packaged, plausible nonsense as your answer."
Can it get even worse? Oh yes.
What could go wrong?
Some background is in order here. Vitamin B6 participates as a coenzyme in over 100 enzymatic reactions. In the process it is inactivated. It is recycled to the active form in a loop that requires vitamin B2. For healthy people, the minute quantities provided in the diet are adequate. Levodopa medication, the mainstay for Parkinson's patients, depletes vitamin B6. For patients taking large amounts of levodopa medication, vitamin B6 supplementation is essential. Otherwise B6 can be depleted with dire results. Pyridoxine, the inactive form of vitamin B6, is common in small quantities in cheap vitamin supplements. In large quantities it is toxic because it substitutes for the active form of B6. The active form of vitamin B6 is known as P5P, or PLP. It is nontoxic and is the form that occurs naturally in the diet. It is also readily available as a supplement. All links in this paragraph are to my writing, supported by references to primary sources - research reported in peer reviewed medical journals. I provide an overview of vitamin B6 in Parkinson's here.
The “turnto” AI says:
Why should vitamin B6 be avoided in Parkinson's disease?
As stated above, failure to supplement B6 in Parkinson's patients taking levodopa medication can have dire consequences, including intractable epileptic seizures. It continues:
Vitamin B6, also known as pyridoxine...
As discussed above, this is only one version of vitamin B6, the toxic one, which should not be used. Then:
Vitamin B6 can interfere with the effectiveness of levodopa.
Wrong. Reality: vitamin B6 is required for conversion of levodopa into dopamine: “Vitamin B6 drives the conversion of levodopa to dopamine.”
Next:
when levodopa is combined with carbidopa, vitamin B6 supplements may be taken safely.
Wrong. Vitamin B6 taken with carbidopa results in the two combining together, inactivating both: “carbidopa irreversibly binds Vitamin B6” They should not be taken at the same time. Separate by 2 hours or by a meal.
So we have four wrong instances of medical advice, shamelessly provided, all in a single AI post. Advice that, if taken, could have dire consequences
From an article at Ars Technica:
Derec01
I work in the field and get very frustrated at the framing of the question "Why do LLMs make stuff up?". It presupposes that "making stuff up" is an aberrant choice or bug, and that if we just fix a few bugs, the output would be truthful naturally.
A much more valid to ask "Why would the output of an LLM be close to truth?" They are distributions over outputs, extrapolating from anything they've been trained on. It's a miracle they are often as accurate as they are. Bringing source data, knowledge graphs, etc. into context and extended compute can be a start, but you have to rigorously drive these things towards truth, no one should be surprised that extrapolating text alone would be inaccurate. [Emphasis added]
Conclusion
What we know of as AI is a threat, not because it is smart, but because it is dumb. It pollutes our knowledge with lies. If you want to do research and get correct answers, avoid AI. I use Google Scholar. If you use Google Search, always append your search terms with "-ai" to keep the garbage out.
KOSABILITY IS BY AND FOR
- PEOPLE LIVING WITH DISABILITIES;
- WHO LOVE SOMEONE WITH A DISABILITY; OR
- WHO WANT TO KNOW MORE ABOUT THE ISSUES.
OUR DISCUSSIONS ARE OPEN THREADS IN THE CONTEXT OF THIS COMMUNITY.
FEEL FREE TO
- COMMENT ON THE DIARY TOPIC;
- ASK QUESTIONS OF THE DIARIST OR GENERALLY TO EVERYONE;
- SHARE SOMETHING YOU'VE LEARNED; OR
- GRIPE ABOUT YOUR SITUATION.
OUR ONLY RULE IS TO BE KIND.
BULLIES WILL BE IGNORED AND/OR OBLITERATED.
FOR MORE ELABORATION ON OUR GROUP RULE PLEASE READ THIS STORY.
AT KOSABILITY WE AMICABLY DISCUSS ANY AND ALL MATTERS PERTAINING TO HEALTH. OUR DISCUSSIONS ARE NOT MEDICAL ADVICE. MEDICAL ADVICE CAN ONLY BE PROVIDED BY A QUALIFIED PHYSICIAN WHO HAS EXAMINED THE PATIENT. IF YOU HAVE WORRISOME SYMPTOMS PLEASE SEE YOUR DOCTOR!
All health related discussion fair game!
