Correction scenario for the AI bubble: the inputs a decision engine needs

Which inputs set the correction score in the AI bubble case?

Four groups of inputs set the score. Margins say how far each dimension sits from the line. Shocks say what could push it. Dependencies say which dimension drags another one along. Exposure says how hard a reset would hit your own position. The engine reads the 4 groups together and returns one score per correction.

The engine is deterministic, so identical inputs return an identical score. That rule makes the sheet a working tool: change one input, rerun, and the difference in the score belongs to that input. Keep the sheet short enough to update in one sitting, with one line per input and one owner per line, so that a quarterly review takes under an hour.

How do you measure each correction input’s margin to the 0.3 threshold?

A margin is the reading of a dimension minus 0.3. In the study, Communication reads 0.305 and has a margin of 0.005. Resonance reads 0.325 and has 0.025. Care reads 0.338 and has 0.038. Stability reads 0.412 and has 0.112. The smaller the margin, the less shock it takes to cross the line.

Measure the same margins on your own system with the 9 ratings of the simulator, then list the three smallest. Those three decide the depth of your correction. A margin below 0.05 deserves a monthly check, and a margin above 0.1 deserves a quarterly one.

Which shock inputs does a decision engine test in a correction?

The study names 3 shocks that can push several dimensions below threshold: a regulation, an incident, and a recession. Incidents include a bias discovered, a public hallucination, and a data leak. Each shock needs a size, expressed as the drop in points that it would cause on a named dimension.

Fix the sizes before you run the engine and keep them identical across quarters. A shock of 0.02 on Communication tells you little on its own, and the same shock run every quarter tells you whether your margin is growing or shrinking. Sizes are working assumptions, and the trend is the output.

How do hallucination rates enter the inputs of a correction?

The study reports that language models give false answers between 0.7% and 30% of the time, depending on the model and the domain, and that the rate reaches 67% in a medical context (JMIR 2025, quoted by the study). It adds that models sound 34% more confident when they are wrong, and that 47% of enterprise users took a major decision on hallucinated content.

Your input is the false-answer rate in your own use cases. Check a sample of outputs against a source, count the errors, and record the share with its date. A rate that rises feeds the Communication reading downward, which narrows the smallest margin and raises the correction score.

Which dependency inputs link Trust to Communication in a correction?

Dependencies show how a fall in one dimension spreads. The study states that Trust, at 0.450, falls along with Communication, and that requiring explainability above 0.5 lifts Trust to 0.60 and Clarity from 0.487 to 0.55. These links are inputs: each one says how much of a change in one dimension passes to another.

The Λ simulator shows the links as an influence cascade on a 9 by 9 matrix, with first-order effects that are direct and second-order effects that pass through an intermediate dimension (EQOS Asia simulator page, October 2026). Read the cascade for your weakest dimension and record the two dimensions it moves most.

Which exposure inputs show how hard a correction would hit you?

Exposure inputs turn the market score into a personal one. Record the share of your portfolio or business plan tied to AI, the concentration in a few names or suppliers, and the horizon of each commitment. A concentrated position takes the full reset, and a spread one takes a fraction.

Add your runway. The method page defines runway as financial leeway including cash, debt, and operational resilience (EQOS Asia method page, October 2026). A long runway lets you hold through a reset, and a short one forces decisions at the lowest point. Enter the months of cash you hold at current spending.

Which dates belong in the inputs for an AI bubble correction?

Dates give inputs their meaning. The study lists AI Act compliance on 2 December 2027, about 13 months after this page. Add your own review dates, the horizons of the prediction pillar at 30, 60, and 90 days, and the reading dates of every input. A reading without a date cannot enter the sheet.

Build the sheet in order: margins, shocks, hallucination rate, dependencies, exposure, dates. Run it once, save the score, and run it again at the next review. After 3 runs you hold a trend, and the trend shows whether the correction case is gaining or losing weight.

Where can you test correction inputs on your own AI bubble case?

The Λ simulator turns your 9 ratings into margins, a cascade, and projections in about 10 minutes, free and anonymous (EQOS Asia simulator page, October 2026). It is the quickest way to fill the margin and dependency inputs for your own system.

A free 30-minute pre-diagnostic with EQOS Asia reviews your sheet and names the inputs that move the correction score most. This page is general information, not investment advice, as of October 2026.

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