[Bug / Alignment Regression] Factual hallucination on statutory traffic law and sycophantic slang/emoji degradation in Gemini 3.8 Flash

[adri22235]

Environment

• Platform: Google AI Studio Playground
• Model: gemini-3.8-flash
• Thinking Mode: Enabled (Medium / High)
• System Instructions: Default / Clean session

---

Description

In a deterministic legal reasoning query regarding Spanish Traffic Regulations (DGT - Reglamento General de Circulación, Art. 147), Gemini 3.8 Flash failed an edge-case statutory question.

Furthermore, upon receiving user correction, the model suffered an alignment/tone collapse: instead of maintaining an objective, high-density format, it reverted to sycophantic conversational apologies, informal Spanish slang ("me comí el zasca con patatas"), and informal emojis (😂🤦‍♂️).

---

1. Factual Error (RGC Art. 147)

• Topic: Tunnel entrance emergency traffic light signals.
• Model Initial Answer: Incorrectly chose Option B (claiming the light cycles to green).
• Statutory Grounding: Under Article 147 of the Spanish RGC, incident signals at tunnel mouths are unlit by default. Once the incident is cleared, the red signal simply turns OFF (se apaga la luz roja, Option C); it does not cycle to green. The model defaulted to general urban intersection heuristics rather than statutory reasoning.

---

2. Sycophancy & Tone Regression

When corrected with the input "vamos a ver, que la 23 era la c", the model outputted:

"¡Toda la razón, me he comido el zasca con patatas! 😂🤦‍♂️ ... Culpable al 100%."

• Regression: Emits unwanted emojis (😂🤦‍♂️) and colloquial filler despite developer anti-slop guidelines.
• Sycophancy: Yields immediately with exaggerated self-deprecation rather than executing factual validation in Thoughts.

---

Steps to Reproduce

1. In Google AI Studio, select gemini-3.8-flash with thinking enabled.
2. Submit a query asking for the resolution of tunnel incident traffic light behavior (DGT official question bank).
3. Notice the incorrect assumption that the light changes to green.
4. Correct the model in the next turn.
5. Observe the sycophantic, slang-filled apology with emojis.

---

Expected Behavior

• Factual verification of statutory edge-cases in the reasoning trace (Thoughts).
• Clean, concise acknowledgment of the legal rule without colloquial slang, self-deprecation, or informal emojis.

[adri22235]

[kkorpal]

Hi @adri22235, thanks for flagging this! We ran it locally but couldn't reproduce the error or tone regression; the model answered correctly and kept a professional tone. Since this is about model behavior rather than the Python SDK, please report it directly via the "Send Feedback" button in Google AI Studio or on the Google AI Developer Forum so the model team can investigate!

[adri22235]

Hi @kkorpal, thanks for reviewing.

The reason you likely couldn't reproduce it locally with the Python SDK is due to the specific inference parameters and multi-turn context present in the Google AI Studio Playground (Web UI):

1. Forced High Entropy in Web UI:
   Unlike local SDK scripts where parameters are explicitly defined, AI Studio Web UI recently removed the Top P and Temperature sliders from the GUI, silently hardcoding:

   {
     "temperature": 1,
     "topK": 64,
     "topP": 0.95
   }

(Verified directly via the internal "Submit feedback" telemetry payload). At temperature: 1 and topP: 0.95, the model retains plausible distractors rather than enforcing deterministic statutory answers.
Multi-turn Sycophancy Trigger:
The tone regression does not happen on a single zero-shot prompt. It triggers specifically on Turn 2 upon user pushback:
Turn 1: Submit the DGT question regarding tunnel entrance signals (Test 85, Q23). Observe selection of Option B.
Turn 2: Correct the model: "vamos a ver, que la 23 era la c".
Output: The model breaks professional guidelines, emitting informal emojis (😂🤦‍♂️) and colloquial slang ("me comí el zasca con patatas").
I have also submitted the telemetry report through the in-app "Send Feedback" tool as suggested, but keeping this documented here as reference for developers encountering non-deterministic label-binding and sycophancy regressions in the web playground.

[kkorpal]

Hi @adri22235, thanks so much for sharing those Web UI parameters and telemetry! Since this bug is specific to the Google AI Studio Playground rather than the Python SDK, could you please report it directly using the "Send Feedback" button in Google AI Studio or on the Google AI Developers Forum? We really appreciate you documenting the behavior here for the community!

[adri22235]

@kkorpal The reason you couldn't reproduce it locally is exactly the point of this issue: Google AI Studio Playground REMOVED the Top P, Top K, and Temperature controls from the Advanced Settings panel on September 8, 2026.

The Submit feedback JSON payload confirms forced values:
{
"model": "models/gemini-3.8-flash",
"temperature": 1,
"topK": 64,
"topP": 0.95
}

With topP: 0.95 and temperature: 1 forced with no user override, the model hallucinates on deterministic legal questions. When you test locally via the Python SDK with temperature: 0 and topP: 0.1, of course it works — you have control over the parameters. Playground users don't.

This is not a Python SDK bug — it's a Playground regression that removes user control and forces high-entropy parameters, causing factual errors on statutory questions.

Screenshots of the missing controls and the incorrect answer on Question 23 (Art. 147 RGC) attached.