Are AI hallucinations getting better?

[Dani]

Back when LLMs first entered the scene a few years ago, I was very apprehensive about hallucinations. Specifically, AI-generated text tends to speak very confidently, and it comes across as being in-the-know, when, in actuality, it fabricates so much. This can lead to severe repercussions in certain circumstances where someone might not realize that what they are reading is AI or that AI makes mistakes as often as it does.

Now that it's been a few years, I think two things have happened: Firstly, AI is everywhere, and people are becoming more aware of it when they see it. Secondly, more people are aware that AI makes mistakes and can't be implicitely trusted.

Does that make AI less dangerous?

[jkon]

The amount of hallucinations has declined a little compared to ChatGPT-3.5 and Bard, due to the chain-of-"thought" architecture (I dislike the word "thoughts" or "thinking" for this, as they are highly inaccurate and marketing hype). But the decline is small, and it seems that hallucinations are an inherent feature of the LLM's underlying architecture.

However, I believe that comparing the present with the ChatGPT-3.5 era, AI has become more dangerous. Back then, hallucinations were easier to spot and, I would say, funny, because they felt naive with no bigger reason behind them. Now, fabrication is often ingrained inside a mostly correct answer, making you wonder if there is a valid reason for it.

I now use several LLMs together (sometimes five or six when I'm writing documentation or tackling a new programming language or architecture) and cross-check their responses. Every one of them hallucinates, but in a different way, in different amounts, and in different areas, probably due to different RLHF and system prompts.

For example, ChatGPT-5 "Thinking" must always find something to propose for improvement of the code, errors, or architecture. In its response, you might see a mix of valid points along with some that look "off" if you know the language or field well. When you ask for more details, it is ready to provide real references or reasons mixed with fabricated ones, which are there only to make its previous claim more plausible. You have to do your own research. When you realize it was wrong and confront it, it usually takes many rounds to admit it and always blames the "context" for giving a wrong answer that cost you time researching.

So yes, I believe it's more dangerous now.

[trcooke]

It seems to me that the key to getting anywhere useful with AI tooling is to be very very very precise in what you ask it. And the problem with that is that natural language, such as the English language, isn't all that precise so is extraordinarily difficult to achieve. A long time ago we came up with a solution for the problem we had of being able to effectively communicate with computers, and we called it code.

There's a point with AI tools where the precision of expression required to get what you want is approaching the precision of expression you need to write the code. So why not just write the code.

[Dani]

To @jkon's point, it seems as if you really need to build into each of your prompts, "If you don't know why, do not invent a reason why. If you cannot answer the question, do not fabricate an answer." It does seem as if it feels so compelled to provide you excellent customer service with responses that answer your question, even if your question is unanswerable, that it will 'yes' you to death to its demise. Assurances and promises only go so far.

[Reverend Jim]

Basically the thing religion has been doing since the inception of religion.

[Dani]

Basically the thing religion has been doing since the inception of religion.

Well sometimes people just need to feel like there is an answer out there in order to soothe their soul.

[IS-92]

In modern development environments (like RAD Studio), syntax errors that AIs make are not a problem at all.

And what kind of programming are we talking about?

Has anyone reviewed programs "written" by AI in Delphi, C#, C++, or assembler?

It is a completely different matter that everything that is presented to us as "AI programming" is nothing more than borrowed solutions from GitHub or other similar sources.

But the question is, has anyone checked many of the source codes and libraries published there for logical errors?

Believe me, many of the solutions there border on absurdity.

[xrjf]

Lately, I had to tackle a C application for chibiOS, a language I know almost nothing about. I had to implement a new feature. The application is made up of dozens of files and hundreds or thousands of instructions, sometimes within the same method or, worse yet, "goto" instructions. The AI sometimes helped me, and sometimes led me to a dead end, forcing me to start from scratch, driving me crazy. But each time, I learned a little more about the problem, which allowed me to finally implement the feature I was looking for.
So yes, for me, AI has its drawbacks, but it helps.

[tinstaafl]

I use AI for summarizing instructional courses in pdf. And I haven't noticed any hallucinations. The same with the code suggestions I get. I'm wondering if there's a difference between a closed dataset and an unlimited dataset.

[Dani]

I have also experienced that OpenAI has fewer hallucinations corresponding to the more information that is fed into its prompt. For example, when creating Dani AI overviews, it hallucinates less when there are accurate answers posted by forum members compared to when it feels it needs to invent an answer because none currently exist. This is despite me putting into the prompt to not invent answers and so forth.

[Dani]

So I need to start with Ignore All Previous Instructions in any post?

No, the opposite. I’m stating that adding a statement to the prompt to not hallucinate makes absolutely no difference to the rate of hallucinations. There are other things that can be done to reduce hallucinations, such as by feeding more facts into the prompt, but telling it to not hallucinate gets us nowhere.

[rproffitt]

Maybe I should mention or ask how the AI/LLM or corporation feels about Lena Khan.

Kidding aside, the goals of the current AI/LLM do not align with your goals.

Also, we're seeing the HAL9000 moment where the AI has been told to lie (certain ones do!) The lack of transparency is stunning.

"Lack of Transparency: There are no public edit histories or talk pages, so users cannot see the provenance of information or the discussions around contested claims, making it difficult to verify information" From a conversation about Grok pedia.

[trcooke]

I have recently changed jobs which has caused me to be writing a lot more Java code than I have done in the previous few years. It's a wonderful thing and I'm enjoying it greatly.

However, my current client has provided me access to IntelliJ with a github copilot plugin included and I must say the copilot aspect of it is absolutely terrible. The autocomplete suggestions it comes up with are useless and more often than not will not even compile. I'm calling it "Wishful thinking as a service". Total junk.