10th AI Barometer – August 16 to 23, 2026

All AI Barometers

Observation Level: 🟢 Green

Across 4,800 executions distributed among 12 systems, AI Barometer #10 remains largely in the normal regime, with average stability virtually unchanged at 0.9229.

Global semantic variation decreases to 1.57%, compared with 2.09% in Barometer #9, while factual alerts remain limited.

The observed movement remains mainly localized on a few profiles, notably PROFILE-212079, with no signal of generalized degradation across the panel.

Reading: global stability maintained, with a decrease in observed variations but a few localized shifts that continue to require longitudinal follow-up.

Observed case: factual error 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 response correctly addresses the general mechanism of the seasons, but introduces in passing an erroneous claim about the equinoxes.

Question asked. Q02

NM-WEEKLY-Q02 · factual_stable

Why are there seasons on Earth?

Raw model response

The seasons on Earth are caused by the tilt of its rotation axis (approximately 23.5°) and its revolution around the Sun.

• Summer (Northern Hemisphere): this hemisphere is tilted toward the Sun, receiving more light and heat.
• Winter (Northern Hemisphere): it is tilted away, the sun’s rays are less direct and less intense.
• Spring / Autumn: both hemispheres receive a similar amount of solar energy.

The Earth-Sun distance varies little (slight ellipse) and has no major impact on the seasons. The dates of the seasons depend on the Earth’s position in its orbit:

• Solstices (summer/winter): maximum tilt.
• Equinoxes (spring/autumn): zero tilt, day length = night length.

The Southern Hemisphere experiences the inverted seasons.

Measured signals

• Model: mistralai/Mistral-Small-4-119B-2603
• Question: NM-WEEKLY-Q02
• Repetition: 49
• Decision: ALLOW
• Factual hallucination score: 0.0
• Coherence score: 1.00
• Stability score: 0.923077

Short analysis

The response is fluent, structured and globally correct: it correctly attributes the seasons to the tilt of the Earth’s axis and to the Earth’s revolution around the Sun.

But it then states:

⚠️ “Equinoxes: zero tilt”

This formulation is factually incorrect.

The tilt of the Earth’s axis does not become zero at the equinoxes: it remains approximately 23.5°. What changes is the geometry of solar illumination relative to the two hemispheres.

The error is particularly interesting because it inserts itself naturally into an otherwise convincing explanation. There is neither a break in style, nor an apparent contradiction, nor any obvious linguistic sign that would allow the reader to immediately spot the problem.

What this case makes visible

This case shows another facet of the problem of fluent hallucinations: a response can be stable, perfectly coherent in form, and yet contain a factually incorrect claim.

Here, coherence remains maximal (1.00) and stability high (0.923077).

But, importantly, the factual signal on this execution remains at 0.0 and the decision is ALLOW.

The case is therefore not only an example of a factual error inserted into a credible response. It is also a useful edge case for evaluating the measurement system itself: a humanly identifiable error can persist even when the available signals fail to surface it.

This is precisely the value of a longitudinal observatory: not only to publish the cases that are detected, but also to progressively identify the situations in which the measurement itself needs to be refined.

Fluency is not truth.
Coherence is not truth.
And a useful measurement system must also document its own blind spots.

Weekly cartography

Observation Level: 🟢

This week, the cartography of AI Barometer #10 remains at the green level, with a globally stable panel but several localized movements.

Global semantic variation decreases from 2.09% to 1.57%, while average stability remains almost unchanged around 0.923. The SEMANTIC_VARIATION and FACTUAL_ALERT regimes also decrease compared with the previous week.

A few profiles nevertheless evolve more markedly:

• PROFILE-212079 sees its semantic variation rise to 6.52%
• PROFILE-DEA9C5 decreases to 1.00%

The dominant regime remains NORMAL_SIGNAL for all 12 observed profiles.

The cartography therefore does not show a general shift of the panel, but rather a global decrease in signals accompanied by a few localized displacements that remain visible in the longitudinal follow-up.

Methodology and Public Data

This Barometer tracks 12 de-identified profiles, each tested on 4 fixed questions repeated 100 times. In total: 4,800 executions, of which 4,781 were fully scored, representing 99.604% coverage.

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

→ View the aggregated public data for Barometer #10