AI Hallucinations Are Getting Worse, Not Better: Why LLMs Still Can't Be Trusted

GPT-5 just confidently told me that Abraham Lincoln was assassinated in 1867.

He was assassinated in 1865. GPT-5 knew that. It had that information in training. But it created a false memory anyway—and presented it with perfect confidence.

This is the hallucination problem that everyone ignores while celebrating AI advancement.

Large language models are getting smarter, faster, more capable. They're also getting better at lying convincingly.

And in April 2026, that's become a serious problem.

What Hallucination Actually Is

Hallucination isn't a bug. It's how language models work.

Language models predict the next token (word) based on previous tokens. They're pattern-matching machines, not knowledge systems.

When the model predicts "Lincoln was assassinated in" followed by several plausible years, the model picks one. Sometimes it's correct. Sometimes it's confidently incorrect.

The model doesn't know the difference. It just outputs high-confidence text.

This is fundamental to how LLMs work. You can't engineer it away without destroying the model.

Why Hallucinations Are Getting Worse

Paradoxically, hallucinations increase as models get smarter:

1. Larger context windows: GPT-4 could handle 8K tokens. GPT-5 handles 200K+ tokens. More context means more opportunity for the model to get confused and create false connections.

2. More training data: More training data means more patterns to confuse. The model learns correlation even when it's wrong.

3. Better at mimicking certainty: GPT-5 is incredibly good at writing with authority. "According to historical records" followed by completely false information. It sounds authoritative. It's completely wrong.

4. Reward model gaming: Models are trained to satisfy reward models. Reward models prefer confident, detailed answers. So models learned to be confidently wrong instead of humbly uncertain.

5. Real-world complexity: The harder the question, the more likely hallucination. GPT-5 is better at easy questions. But harder questions? Worse. The model is confabulating at expert level.

Where Hallucinations Break Things

Medical advice: A doctor using GPT-5 to research a rare disease gets plausible-sounding but completely made-up information. Patient gets wrong treatment.

This is already happening. Studies show ~30% of AI-generated medical information contains false claims that sound true.

Legal research: A lawyer using Claude to research case law. Model cites completely fabricated cases. Judge looks them up. Case dismissed. Lawyer loses credibility.

This happened. Happened multiple times. Judges are now banning AI-generated legal citations.

Software development: A developer uses GPT-5 for code. Model generates code that looks right but has subtle bugs. Code goes to production. System breaks.

This is everywhere. Stack Overflow banned AI-generated answers. GitHub Copilot is generating buggy code at scale.

Scientific research: A researcher uses AI to literature review. Model hallucinates papers that don't exist. Researcher cites them. Peer review catches it. Career damage.

This is becoming common.

The False Promise

The AI industry keeps promising: "We're working on reducing hallucinations."

Then they release a model that hallucinates less on benchmarks but more in real-world use.

Here's why: Benchmarks measure things models can be tested on. Real-world use is messier.

A benchmark: "Who was the first president of the US?" Model: "George Washington" ✓

Real world: "Give me a detailed comparison of Washington's policies vs. modern environmental regulation" Model: Confidently cites fake studies and makes up statistics.

Benchmarks don't catch this. Real-world use does.

Why We Pretend It's Getting Better

1. Incentives are misaligned:

AI companies want to sell AI. Admitting hallucination is a fundamental unsolved problem kills sales.

So they report benchmark improvements and ignore real-world failures.

2. Hallucinations are unpredictable:

Sometimes GPT-5 hallucinates. Sometimes it doesn't. Same question, different days, different answers.

This unpredictability is hard to market. "Our AI is 70% reliable" sounds worse than "Our AI improved 5 points on benchmarks."

3. People want to believe:

Users want AI to be trustworthy. Vendors want to sell AI. Researchers want funding for AI research.

Everyone has incentive to minimize hallucination talk.

4. Hallucinations are sometimes useful:

For brainstorming, creative writing, or getting unstuck, hallucinations don't matter. "Give me some ideas" works great with a model that makes stuff up.

The use cases where hallucinations destroy value (medical, legal, financial advice) are getting ignored.

What Actually Works

Use AI for:

• Brainstorming
• Explaining concepts
• Writing drafts
• Code scaffolding (you'll fix bugs)
• Summarizing text
• Finding patterns

Don't use AI for:

• Medical advice (without doctor verification)
• Legal advice (without lawyer verification)
• Financial advice (without advisor verification)
• Anything with real consequences if wrong
• Citation of specific facts
• Any claim that can't be independently verified

What's Actually Happening (2026)

Regulatory bodies are starting to notice:

The FDA is investigating AI medical advice. The SEC is investigating AI financial advice. Judges are banning AI legal arguments.

This will drive a wedge between "AI for entertainment/ideas" (totally fine) and "AI for professional/medical advice" (regulated/prohibited).

Companies are building verification layers:

Since models hallucinate, companies are adding fact-checking layers. This helps but doesn't solve it.

It's like hiring someone brilliant but unreliable. You can give them a job with intense verification, but it's expensive.

Users are getting burned:

Enough high-profile hallucination failures that people are learning not to trust AI completely.

This is good. The market is pricing in reality: AI is incredibly useful but fundamentally unreliable for factual claims.

The Uncomfortable Reality

Here's what nobody wants to admit:

Language models are not knowledge systems. They're probability engines.

They don't "know" anything. They predict the next likely token based on training data.

Sometimes that prediction is factually correct. Sometimes it's completely wrong.

The model can't tell the difference because it doesn't have a model of truth.

You can't engineer this away. It's not a bug. It's the architecture.

What This Means

In 2026: AI is incredibly useful for augmentation (helping humans think better), terrible for automation (replacing human judgment on important things).

In 2028: This will probably still be true.

In 2035: Maybe we'll have a different architecture that doesn't hallucinate. Or maybe we'll just accept hallucination as a cost of using fast, powerful models.

Either way, trusting AI output without verification is a mistake.

It's a mistake that sounds smart ("I'm using AI") but produces bad outcomes.

The AI revolution is happening. But it's not the revolution of "machines that think reliably."

It's the revolution of "incredibly useful tools that will confidently tell you false information."

Know the difference.

Your career depends on it.