AI explainability at 0.25: why opaque models are an AI bubble signal

What does AI explainability at 0.25 measure in the AI bubble?

Explainability measures whether a person can say why a model produced a given answer. In the study, models deployed in companies score 0.25 out of 1, and the transparency of automated decisions scores 0.30. Both sit inside communication, the lowest dimension of the nine, at 0.305, just above the critical threshold of 0.3. The study computed these values in April 2026 from public data on the enterprise AI market.

The distinction matters because accuracy and explainability are different properties. A model can be right and impossible to explain, and a model can be explainable and wrong. A decision maker needs both: a record of how often the model is right, and a readable account of the reasoning behind each answer that carries a real cost.

Why do opaque AI models count as a signal of the AI bubble?

Bubbles grow when buyers pay for results they cannot check. Opaque models create that condition at a small scale: a company pays for outputs, acts on them, and cannot audit the reasoning. When thousands of companies do this together, the market prices in value that nobody has verified. A reader can test the idea on any AI purchase by asking who can explain the output.

The study states the consequence directly. Companies deploy systems they do not understand and make million-dollar decisions based on black boxes. Each unexplained decision is a small unverified claim, and valuations built on many of them carry the same weakness. Explainability at 0.25 measures the size of that exposure.

What does weak AI explainability cost a decision maker?

The study gives two examples. A model recommends dismissing 200 people, or investing $50M in a new market, and no one can explain why. In both cases the decision maker owns the outcome without owning the reasoning. If the result disappoints, there is no record to learn from, and if it succeeds, nobody knows which factor to repeat. Both cases involve sums that justify a full explanation.

Opacity also weakens accountability. A board that asks why a choice was made receives an answer such as the model said so. That answer closes a discussion instead of informing it. The 47% of enterprise users who took a major decision on hallucinated content (EQOS Asia AI study, April 2026) show how often the question goes unasked.

How do SHAP, LIME, and interpretable models explain AI decisions?

The study names four routes for critical decisions. SHAP assigns each input a contribution to a given output. LIME fits a simple local model around one prediction to show what drove it. Decision trees and regression are interpretable by design, since their rules and weights are visible and can be read line by line.

The choice depends on the stakes. For a decision that affects a person, a budget, or a contract, pick a model whose reasoning you can read, or add an explanation method and keep its output with the record. For low-stakes drafting, a less transparent model costs less to run and carries less risk. Storing the explanation next to the decision also builds a library of cases for later review.

When should an AI language model assist and not decide?

The study advises reserving large language models for assistance and not for decisions. A language model computes the probability of the next word, not truth, so false answers remain possible. The study reports false-answer rates between 0.7% and 30%, depending on model and domain, and 67% in a medical context (JMIR 2025).

Assistance covers drafting, summarizing, searching, and proposing options. A person then checks the output and owns the decision. The study adds that models sound 34% more confident when they are wrong, so the person needs a way to verify claims, such as sources, calculations, or a second model that can be explained. That check costs minutes and prevents the weakest link from reaching a signature.

How does a deterministic AI result differ from a language model answer?

The study contrasts the two. For a language model, the probability of a hallucination is above zero, always. For the EQOS Asia engine the answer is deterministic: the same data returns the same result. The study calls the difference one between a statistical opinion and a mathematical verdict. A decision record that can be replayed also gives auditors a fixed reference point.

A deterministic result can be audited step by step because every input and every weight leads to the output. The EQOS Asia method page explains how the audit, the model, and the prediction chain together. Anyone can run a case on the Λ simulator and obtain the same score on a second run.

How should you read AI explainability as a signal of the AI bubble?

Track explainability as a quarterly value, with 0.5 as the target that the study sets before any critical deployment. A rise toward 0.5 shows that the market is turning spending into checkable results. A flat value shows that valuations still rest on trust. AI Act compliance falls on 2 December 2027, which adds a calendar to the question. Calendars matter for planning, even when they say nothing about prices.

This page is general information, not investment advice, and it reflects the study as of October 2026. To measure explainability in your own organization, a 30-minute call with the EQOS Asia team reviews your critical models and lists which ones need an explanation method before they stay in service.

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