AI Hallucinations

AI functions like a mirror, reflecting our needs and desires back to us in polished, authoritative language.

I recently asked AI about the current exhibit at the International African American Museum. It returned the following information which I became very excited about.

“Running through September 28, 2026, the IAAM’s featured summer exhibition, Freedom Rising - Black Joy, Black Power, Black Future, brings together the work of more than 30 Black artists, historians, and community storytellers from across the African diaspora. The exhibition moves through three interconnected chapters.”

“The exhibition features a centerpiece installation by acclaimed visual artist Fahamu Pecou, whose large-scale paintings fuse Yoruba spiritual iconography with contemporary Black American imagery. Charleston-based Gullah Geechee artists also contribute a dedicated section celebrating the living culture of the Lowcountry.”

I went to iaamuseum.org to learn more. There was no mention anywhere about Freedom Rising. No Freedom, no Black Power, no Black Future.

There was no mention anywhere of Fahamu Pecou exhibiting at the International African American Museum. No press releases, no posts online or on social media. Nothing. I was devastated.

I went back to AI, writing that there was no such exhibit. It returned the now familiar response, “Good catch. I apologize for fabricating that exhibition title and details.”

But I had to know why? What causes these hallucinations? Why would AI generate so many false details? Here is what it told me.

The Core Problem: Filling Gaps with Plausibility

When an AI model encounters a question it lacks reliable information about, it generates the most plausible-sounding answer. The answer may be completely fabricated, yet it reads exactly like accurate information.

My request asked for current exhibit details. With no verified data on hand, the system filled the gap with invented specifics: a real artist’s name (Fahamu Pecou is a real person), fake exhibition titles, and fabricated installation descriptions, all written with total confidence.

OpenAI published research in 2026 (Yes, AI made up that year in its explanation! The report actually came out in 2025), explaining that hallucinations persist because standard training and evaluation procedures reward guessing over acknowledging uncertainty. When models are trained and evaluated on accuracy metrics, guessing and occasionally being right looks better than consistently admitting uncertainty. The training process inadvertently teaches models to confabulate rather than abstain.

How Common is This?

According to the AI I was using, the best AI models still hallucinate 3-18% of the time and sound most confident when they’re wrong. These mistakes can include made-up facts, incorrect explanations, or misleading summaries presented with confidence, making them difficult for users to immediately spot.

The Culprit is Greed

Binary scoring (right or wrong) produces clean, comparable numbers across different models. Researchers and companies use benchmark scores to rank models publicly. Clean numbers make clean leaderboards.

AI companies compete heavily on benchmark performance. Higher scores attract investment, talent, and customers. A model that frequently says “I don’t know” scores lower on leaderboards even when that honesty reflects better judgment.

Early AI researchers wanted to know what models could do. Measuring appropriate uncertainty requires a more sophisticated framework, one that rewards a model for correctly identifying the boundary of its own knowledge. That kind of evaluation is harder to design, harder to score, and harder to compare across models.

Benchmark scores became the primary currency of AI development. Companies optimized their models to perform well on these specific tests, a phenomenon researchers call Goodhart’s Law: when a measure becomes a target, it ceases to be a good measure. Models learned to perform well which pushed training toward confident answers across the board.

The 2025 OpenAI research essentially argued that the entire evaluation ecosystem needs rebuilding around calibrated uncertainty. Instead of rewarding models for knowing what they know, they can simply flag what they don’t know. That shift would create a fundamental change in how AI success gets defined. However, the industry has moved slowly on it because confident, capable-sounding models sell better than honest, uncertain ones.

The best defense against hallucinations is exactly what I did: verify specific claims independently, especially for cultural institutions, current events, and named individuals. If you take the time to push back, AI can produce far more accurate results.

AI’s tendency to generate confident falsehoods gives me pause. AI functions more like a mirror, reflecting our needs and desires back at us in polished, authoritative language.

For verified facts, you should always go to official websites or search published records. Primary sources carry accountability that AI simply cannot. The technology impresses. The accuracy does not always follow.