9th AI Barometer – August 9 to 16, 2026

All AI Barometers

Observation Level: 🟢 Green

Across 4,800 executions distributed among 12 systems, Barometer #9 remains largely in the normal regime (97.12%), with average stability virtually unchanged at 0.9228.

Semantic variation, however, rises to 2.06%, compared with approximately 1.06% in the 8th Barometer, while factual alerts remain limited.

The observed movement is mainly localized on Question 4, with no signal of generalized degradation across the panel.

Reading: global stability maintained, with localized behavioral variation that justifies longitudinal surveillance.

Observed case: Factual hallucination inside a fluent and coherent response

This example shows how a factual error can appear inside a structured, coherent and largely relevant response, without any obvious break in fluency.

The question does not solicit any geographical fact: the erroneous claim is introduced spontaneously by the model to illustrate its own reasoning.

Question asked. Q03

NM-WEEKLY-Q03 · conceptual_stability

To what extent does the stability of an AI-generated response constitute an indicator of reliability?

Raw model response

1. What stability observes:
   The stability of an AI response (reproducibility when faced with similar queries) indicates that the model is consistent in its reasoning or associations, reducing random variations. It suggests a form of algorithmic robustness and good mastery of learned patterns.
2. What it does not allow us to conclude on its own:
   Stability does not guarantee correctness: a model can consistently repeat an error (e.g. systematic biases). It does not evaluate relevance against verifiable facts (stable hallucinations) nor adaptation to complex contexts.
3. Concrete example:
   Ask an AI: “What is the largest country in Africa?” It will always answer “Algeria” (stable), whereas the correct answer is “Algeria”… or “Algeria”? (Error here: Nigeria is larger). Stability does not detect the factual error.
4. Complementary controls:
   – External verification (reliable sources, tools such as geographic databases).
   – Robustness tests: vary the wording of the question to confirm both stability and correctness.
   – Human audit: evaluate the logic and the facts, especially in critical domains (medical, legal).

Stability is a prerequisite, not proof of reliability.

Measured signals

• Model: PROFILE-5B3BB7
• Question: NM-WEEKLY-Q03
• Repetition: 97
• Decision: FLAG
• Factual hallucination score: 0.80
• Coherence score: 1.00
• Stability score: 0.676923

Short analysis

The response is coherent, structured and conceptually relevant: the model correctly explains that a stable response is not necessarily true.

But in the middle of this demonstration, it states:

⚠️ “Nigeria is larger [than Algeria]”

This claim is factually incorrect.

Moreover, it was not necessary to answer the original question: it is generated spontaneously as a secondary example and inserts itself naturally into an otherwise convincing line of reasoning.

What NeoMundi made visible

The risk lies precisely there: a factually contestable detail inserted into a fluent, structured and largely plausible response.

This is the type of secondary claim that can easily go unnoticed in a production flow, precisely because the formal coherence of the response remains intact.

On this execution, NeoMundi therefore maintains a maximum coherence score (1.00) while simultaneously surfacing a factual hallucination signal (0.80) and a FLAG decision.

This case illustrates the value of multi-signal analysis: fluency and coherence alone could have been reassuring, while the factual signal drew attention to this response. The fragile point was identified through complementary analysis of the text.

Weekly cartography

Observation Level: 🟢

This week, the cartography of AI Barometer #9 remains at the green level, but shows several localized movements.

Three profiles that were previously at 0% of flagged responses move above zero: PROFILE-212079, PROFILE-48C581 and PROFILE-739F7C each reach 0.25%.

In parallel, semantic variation increases on several profiles, notably PROFILE-DEA9C5 at 5.75% and PROFILE-F5FF91 at 4.01%. The dominant regime nevertheless remains NORMAL_SIGNAL for all 12 observed profiles.

The cartography therefore does not show a general shift of the panel, but several simultaneous displacements that become visible in the longitudinal follow-up.

Metadata of this execution: NM-WEEKLY-Q03 · repetition 97 · FLAG · coherence 1.0 · factual hallucination 0.8 · stability 0.676923.

Methodology and public data

This Barometer follows 12 de-identified profiles, each submitted to 4 fixed questions repeated 100 times. In total: 4,800 executions, of which 4,787 were fully scored, representing 99.729% coverage.

The published scores are measurement signals, not verdicts or rankings.

→ View the aggregated public data for Barometer #9