AI Hallucinations: Are Reasoning Models Worse?

⬇️ Prefer to listen instead? ⬇️

• OpenAI’s reasoning model o3 hallucinates more often than GPT-4-turbo despite better logic performance.
• When AI reasons more, it might get facts wrong because it depends less on its training data.
• Developers risk their name more when facts are wrong in important jobs like law and finance.
• With models where the code is secret, researchers have a hard time figuring out why hallucinations happen.
• Retrieval-augmented generation and checking with multiple AI systems are good ideas to lessen hallucinations.

OpenAI’s newer AI models, like o3 and o4-mini, focus on logic. They reason better than past models. But along with better thinking comes more hallucinations. These are times when LLMs confidently say things that aren’t true. This is confusing for AI makers and people using the AI. People ask: Why are smarter models worse at telling the truth? How can we stop this?

Advertisement

What Are AI Hallucinations, and Why Should You Care?

AI hallucinations happen when large language models (LLMs) give answers that seem okay, but are wrong, don’t matter, or are totally made up. These wrong answers can be small mistakes, like saying the wrong person said something. Or they can be big made-up things, like inventing court cases or science ideas. Hallucinations aren’t just math errors. They sound just as smooth and sure as true facts, making them very tricky.

This is a big worry. Hallucinations cause real problems, especially as AI tools get used in important jobs like:

• In healthcare, wrong advice could mean the wrong sickness is found or cause bad treatment.
• In law, made-up cases or rules can mess up court plans.
• In finance, wrong market numbers or made-up statistics can lead to bad money choices.

As LLMs like OpenAI’s are used more in work tools, small mistakes can spread widely. This could cause big problems across systems if no one checks them.

From Chatbots to Logical Thinkers: What Sets o3 and o4-mini Apart?

OpenAI’s models that focus on reasoning, especially o3 and the smaller o4-mini, show a main change in how AI is being made. Past models focused on smooth talk, chat skills, or doing tasks on their own. But these new models are fine-tuned to solve harder reasoning problems.

OpenAI says these models are very good at:

• Solving math problems with many steps
• Testing ideas in science
• Using logic to figure things out
• Breaking down big problems into smaller parts

On paper, this makes models like o3 appear “smarter” or more capable. But there’s a main trade-off: these models can, and are even pushed to, guess, make sense of things, and try new ways of thinking. This freedom to create is key for solving hard, unclear problems. But it also makes hallucinations more likely, especially when the model doesn’t have clear training data or the question isn’t clear.

Chatbots usually stick close to known facts or what’s in databases. But reasoning models go into less known thinking areas. This makes it more likely they’ll say something wrong with confidence.

How OpenAI Measures AI Hallucination Rates

To know when and how these hallucinations happen, careful testing is needed. OpenAI and other AI companies use ways to test how models act over time:

• People checking: They look at what the model says and check it against trusted sources or fact lists to see if it’s right.
• Test Data: They use things like TruthfulQA and FactScore to see how true the model’s answers are in different areas.
• Checking against old models: OpenAI watches if new models do better or worse than old ones at reasoning and getting facts right.

And what they found is clear: OpenAI’s own look at this, which TechCrunch wrote about, says that o3 does hallucinate more than GPT-4-turbo. They haven’t shared the exact numbers. But talks inside the company and test results shared outside point to more hallucinations happening. This is especially true for tasks where the answer isn’t fixed or the question isn’t clear.

This step backward in being trustworthy, even with better reasoning, makes people worry if today’s AI models are really safe to use where mistakes can be bad.

Why Would Better Reasoning Lead to More Errors?

It might seem strange that better reasoning causes more sure errors. But it makes sense when you look at how LLMs work.

More Freedom to Create

Better reasoning gives the model more freedom to figure out questions, link hard ideas, or build logical points. But this freedom can also disconnect the model from its checked training data. This makes it more likely to create answers that sound right but are made up.

Example: Ask an o3 model to “explain how quantum gravity could affect how planets move.” It might write something that sounds real. But there might be no real studies backing up what it says. It is reasoning, but it’s wrong.

Not Tied to Reality

LLMs, including o3 and o4-mini, are not “aware” of what is real. They don’t use a model of the real world. They work based on patterns they learned from their training data. When reasoning tasks go beyond the data they were trained on, the model often just guesses to fill in what it doesn’t know.

Limits in Training Data

OpenAI has trained its models on a huge amount of text. But there are still things they don’t know and ways they lean. For hard tasks with not many real-world examples—like new science ideas or tough moral questions—AI models are more likely to just make things up.

Less Stickiness to Facts

Older OpenAI models, like GPT-3.5 or GPT-4-turbo, often used tighter data checks and safety steps to give safer answers. With new models like o3, they might have cared more about strong reasoning. This could have loosened how much they stick to facts. The choice between being creative and being right is clearer now.

The Growing Impact on Developers and Users

Developers putting models like o4-mini into other apps face a problem: hallucinations aren’t just tech quirks, they’re a risk. Bad results could spread across different jobs:

These hallucinations can hurt how good a product is and how much people trust the brand. And then, they make work less smooth. Now developers have to add extra checks—like tools that look up facts, people who check answers, or backup plans—to make sure things are right.

A big point is that as models seem smarter, people trust them more without asking questions.

This trust can cause:

• People using wrong answers without checking
• Making bad choices in important situations
• Wrong ideas getting stuck in how people work

Trust can be tricky. If an AI seems really good, its wrong answers can be more harmful.

Transparency Battles: Why Secrecy Adds to the Problem

The problem with hallucinations gets worse because making AI is often hidden. OpenAI, and other big labs, show fewer technical details now. Things like:

• How the models are built in detail
• What data is used for training
• How they fine-tune the models
• How they add safety checks

Moving to models where the code is secret makes it harder for outside researchers to find out why hallucinations happen. This makes it harder to check the work and for groups to work together to fix things.

People who want more openness say this secrecy really slows down trust and progress:

• Developers don’t know how specific models tend to hallucinate
• News reporters have a hard time checking what AI says
• Rule-makers have trouble making sure companies are responsible

More and more, people want not just models that work better. They want models that you can explain and check. You need this before you use AI in places where mistakes can be dangerous.

Ways to Lessen AI Hallucinations

Getting rid of all hallucinations might not be possible yet. But people are trying different tech methods and steps to make them happen less often.

Retrieval-Augmented Generation (RAG)

RAG puts together language models with a system that can look up documents. This lets models find current, checked facts while answering a question. Instead of just using what it learned, the AI can add details from real documents or sources.

🔧 Example: Ask o4-mini to sum up a new law. It can look it up in a law database instead of making up rules from its training data.

Training with Truth in Mind

They are using data like TruthfulQA or comparing with knowledge graphs more and more. This is to give models a penalty during training when they make up facts. After doing this many times, the model can slowly learn to avoid saying things that aren’t true or using wrong logic.

🚫 Idea: If the model says “Einstein wrote ‘The Theory of Everything’,” it gets a penalty. This teaches it not to link those things, since it’s wrong.

Showing How Sure It Is

Some new models are trying out ways to show how sure they are about their answers. Users get a warning or a flag when the model isn’t very sure about the info it gives.

🧭 Good side: People can check answers that seem off or ask for things to be made clearer.

Checking with Many AIs

In systems with many AIs, one AI checks the answer of another. Having AIs check each other can make facts much more right. This is extra helpful for longer or special questions.

🤖 + 🤖 = ✅ Enhanced accountability and content vetting

Building Trust Again When Hallucinations Happen

Models like o3 and o4-mini show the good and bad sides of newer AI that makes things. On one hand, they make possible new things in solving problems, being creative, and using logic. On the other hand, right now they do this while trying to balance being complex and twisting the truth.

Rules being made show this struggle:

• 🇪🇺 The European Union’s AI Act puts trustworthy AI first. It clearly points to hallucinations as dangers in models used by the public.
• 🇺🇸 The White House’s plan for AI rights says people have the right to know why AI does things, to check facts, and to control AI systems.

Groups looking ahead have already started checking their AI inside, having others check models, and using plans for using AI fairly. This is to lower the chance of truth problems.

It looks like things will go this way: Security became a basic need for software. Being factually right will become a basic need for AI.

Moving Ahead Carefully

When models focused on reasoning, like o3 or o4-mini, hallucinate more than older ones, don’t see it as a failure. See it as signs of things getting better on the way to smarter AI. The freedom that helps them reason better also makes it more likely they’ll go wrong.

People know more about AI hallucinations now. So we all need to try harder to:

• Show more about how models are made
• Agree on safety rules across different jobs
• Come up with new tech ways to tie AI better to facts
• Help people learn how to tell if AI info is true

Machines will keep getting better at thinking. But we must also keep asking the hardest question: not just “Can this AI sound smart?” but “Can this AI be trusted?”