47% of AI decisions rest on hallucinations: a hidden AI bubble risk

What does the 47% hallucination figure mean for AI decisions?

Read the figure exactly. The study reports that 47% of enterprise users took at least one major decision based on hallucinated content. It does not say that 47% of all AI outputs are false. It says that nearly half of the people using these tools in companies have, at least once, acted on an answer that was wrong.

One bad decision is enough to matter. A single wrong market size, legal reading, or customer analysis can steer a budget for months. The figure counts people who acted, which makes it a measure of exposure and a stronger signal than error rates alone. It shows how far false output has reached into real choices, and it grows with every team that adopts a new assistant.

How often do AI models produce a hallucination?

The study cites a range: language models such as GPT, Claude, and Gemini produce factually false answers between 0.7% and 30% of the time, depending on the model and the domain (Vectara Hallucination Leaderboard 2025). In a medical context the rate rises to 67% (JMIR 2025). The spread is wide, so a rate quoted without its domain says little.

The same sources add a second concern: models are 34% more confident when they are wrong. A false answer therefore arrives in the same fluent tone as a correct one, and the user receives no signal to double-check. Confidence cannot serve as your test, and a polished style tells you nothing about accuracy. Verification has to come from outside the model.

Why are AI hallucinations a hidden AI bubble risk?

A bubble rests on the belief that AI creates value. Part of that value is claimed through decisions: faster analysis, cheaper research, better targeting. If some of those decisions rest on false output, the claimed value is overstated, and the cost appears later as a failed launch, a lost contract, or a correction. The error stays out of sight until then.

That delay is what makes the risk hidden. Costs of a hallucination do not show in the AI budget; they show in other budgets, months later, under other names. Spending on AI can therefore look productive while the results quietly erode, and the link between cause and effect fades with each month. A market full of such delayed costs holds more risk than its reported returns suggest.

Which business decisions are most exposed to AI hallucinations?

Use 4 criteria. First, the stakes: large amounts or hard-to-reverse choices. Second, verifiability: answers that nobody can check quickly. Third, domain: areas such as medical or legal content, where the study cites higher rates. Fourth, the human check: decisions where nobody reviews the output before action. A decision that meets all 4 criteria deserves the strictest control.

Examples help. When a model recommends cutting 200 jobs or investing $50 million in a new market, the study notes that no one can explain why. That is the profile of an exposed decision: high stakes, no explanation, no review. Low-stakes drafting with a human editor sits at the other end of the scale, where a wrong sentence costs minutes and gets caught.

What do explainability and transparency scores say about AI decisions?

The study scores the explainability of AI models deployed in companies at 0.25 out of 1, and the transparency of automated decisions at 0.30. Communication, the dimension where the study reports both, scores 0.305, only 0.005 above the stop threshold of 0.3. In plain terms, companies deploy systems they cannot explain and decide on the results.

The study also separates 2 kinds of output. A language model computes the probability of the next word, so it can always hallucinate. A deterministic engine returns the same result for the same data every time. The first gives a statistical opinion, and the second gives a verdict you can reproduce and audit. Both have uses, in different places, and the skill lies in placing each one where its strengths count.

How can a business verify AI-assisted decisions before acting?

The study recommends that a company require explainability above 0.5 before deploying a model. For critical decisions it suggests interpretable methods such as SHAP, LIME, decision trees, or regression, and it reserves language models for assistance, not decision. Add a rule that every number or citation in an AI answer is checked against a primary source before it enters a decision.

Log the decision, the source, and the person who checked it. After a quarter, review which AI answers changed a choice and which checks caught errors. The study expects that requiring explainability lifts Trust from 0.450 to 0.60 and Clarity from 0.487 to 0.55. Better verification therefore raises the organization’s score and lowers its hidden exposure.

How can you measure your own AI hallucination exposure?

List the decisions in your organization that use AI output. Mark each one against the 4 exposure criteria, and count how many lack a documented check. That count is your hidden risk, and it can be reduced by one process at a time. Repeat the count each quarter and keep the criteria fixed so the trend is readable. A falling count shows controls working, while a rising count shows adoption outrunning checks.

The EQOS Asia method explains how the study turns such inputs into a score, and the Λ simulator estimates your φ(S) in 10 minutes. A 30-minute session on the contact page applies the scoring to your own AI-assisted decisions and returns a ranked list of controls to add first.

Frequently asked questions

About the author

Stéphane Jambu

President, EQOS Forecast Asia

Stéphane Jambu leads EQOS Forecast Asia and works as a consultant in e-commerce, search visibility, and AI. LinkedIn profile