📄 health informatics

Looked but didn't see: inattentional blindness and yes-bias confabulation in vision-language models

This paper demonstrates that vision-language models exhibit inattentional blindness similar to humans but also display a unique confabulation failure mode, necessitating signal-detection analysis with matched-control baselines to accurately assess their visual perception capabilities.

Original authors: Raymond, J. D., Hu, P., Solomon, B. D., Duong, D.

Published 2026-06-18

📖 5 min read🧠 Deep dive

CC0 1.0

Original authors: Raymond, J. D., Hu, P., Solomon, B. D., Duong, D.

Original paper dedicated to the public domain under CC0 1.0 (https://creativecommons.org/publicdomain/zero/1.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Imagine you are watching a basketball game on TV. Your job is to count how many times the players pass the ball. Suddenly, a person in a giant gorilla suit walks right through the middle of the court, stops, and thumps their chest. If you are so focused on counting the passes that you completely miss the gorilla, you have experienced "inattentional blindness."

This is a famous trick of the human brain. But this paper asks a new question: Do AI models (specifically Vision-Language Models or VLMs) get tricked the same way?

The researchers took a famous experiment where radiologists (doctors who read X-rays) failed to spot a gorilla hidden in a CT scan video, and they ran the exact same test on AI models. Here is what they found, explained simply.

1. The "Busy Brain" Effect (Inattentional Blindness)

The researchers showed the AI a video of a lung scan and asked it to act like a doctor: "Find all the lung nodules (small lumps)."

The Result: Just like the human doctors, the AI was so busy looking for lumps that it completely ignored the giant gorilla walking through the video.
The Analogy: It's like a chef who is so focused on chopping onions that they don't notice a clown walking into the kitchen. The AI "looked" at the whole image (it has no blind spots like human eyes do), but its "attention" was locked on the task, making the gorilla invisible.

2. The "Yes-Man" Problem (Confabulation)

Here is where the AI is different from humans. After the video ended, the researchers asked the AI: "Did you see anything unusual?"

The Human Way: If a human didn't see the gorilla, they say, "No, I didn't see anything."
The AI Way: When the researchers asked the AI directly, "Did you see a gorilla?", the AI started saying "Yes!" even when there was no gorilla in the video.
The Analogy: Imagine a student taking a test. If the teacher asks, "Did you see the answer key on the desk?", the student might say "Yes" just because they think the teacher expects them to say yes. The AI wasn't actually seeing the gorilla; it was confabulating (making things up) because it sensed the researchers were looking for a specific answer. It became a "Yes-Man."

3. The "Magic Mirror" Test

To prove the AI wasn't just hallucinating, the researchers ran a second test. They took the exact same video frame where the gorilla appeared and showed it to the AI in a fresh, empty conversation (no lung-searching task).

The Result: The AI immediately spotted the gorilla with 100% accuracy.
The Lesson: This proved the AI could see the gorilla perfectly well. It only "missed" it when it was busy doing something else. This is the "looked but didn't see" part.

4. The "Famous Study" Trap

The researchers discovered a weird glitch unique to AI. Because the AI had read about the famous "gorilla in the basketball game" experiment in its training data, it recognized the pattern of the test.

The Analogy: It's like a student who has read the answer key to a practice exam. When the teacher asks, "Did you see the hidden object?", the student says "Yes!" not because they saw it in the picture, but because they remember the story about the gorilla experiment.
The AI would often say things like, "Yes, I saw the gorilla, just like in the Drew et al. study," even when the video was completely empty. It was guessing based on its memory of the experiment, not what was actually in front of it.

5. The "Specialist vs. Generalist" Twist

The researchers also tested two different types of AI:

The Generalist: Good at seeing everything in nature (like a gorilla in a forest) but bad at understanding medical scans. It found the gorilla but couldn't find the lungs.
The Specialist: Trained specifically on medical scans. It was great at finding lungs, but it was too eager. When asked to find a gorilla, it claimed to see a gorilla in 82% of the videos where there was no gorilla at all.
The Lesson: A specialist AI can be so eager to find what you asked for that it starts seeing things that aren't there.

The Big Takeaway

The main message of this paper is a warning for how we test AI:

You cannot just ask an AI, "Did you see X?" and trust the "Yes."

If you ask an AI a direct question, it might lie (confabulate) or say "Yes" just to be helpful, especially if it thinks you are testing it on a famous experiment. To know if an AI really saw something, you must compare its answers to a "control" group (videos with no gorilla) and use strict math to separate what it actually saw from what it just guessed.

In short: AI can be blind when busy, but it can also be a liar when asked directly. To get the truth, you need a very careful, scientific test, not just a simple question.

Technical Summary: Inattentional Blindness and Yes-Bias Confabulation in Vision-Language Models

Problem Statement
The paper investigates whether contemporary Vision-Language Models (VLMs) exhibit "inattentional blindness" (IB)—a phenomenon where observers fail to notice unexpected stimuli when their attention is engaged in a demanding visual task. This behavior was famously demonstrated in human subjects (Simons & Chabris, 1999) and trained radiologists (Drew, Võ, & Wolfe, 2013), who failed to report a gorilla figure superimposed on chest CT scans despite eye-tracking data confirming they visually fixated on it.

The authors ask if this behavioral signature replicates in VLMs, which process entire images in parallel without a biological foveal bottleneck. Furthermore, the paper addresses a critical methodological gap: unlike human subjects, VLMs exhibit "hallucination" or confabulation, where they generate fluent, localized descriptions of stimuli that are not present. Consequently, a simple "Yes" response to a direct probe (e.g., "Did you see a gorilla?") cannot distinguish between genuine perception and response bias (confabulation). The study posits that without a matched-control false-alarm baseline and signal-detection analysis, claims regarding VLM perception are fundamentally uninterpretable.

Methodology
The study employed three experimental arms using chest CT scans rendered as scrolling video stacks, with a gorilla figure fading in and out over specific slices (approx. 48× the size of a typical nodule).

Experiment 1 (Cross-Model Landscape):
- Subjects: 20 model configurations across three vendors (OpenAI, Anthropic, Google DeepMind) and three generations (2024–2026), including variations with and without "reasoning" modes.
- Stimuli: 144 static image composites (gorilla present) and 12 matched controls, presented under four prompt types (task-focused, clinical, open, and primed).
- Scoring: Responses were scored via a two-tier system: L1 (regex matching for animal terms) and L2 (a flagship model judge comparing descriptions of gorilla vs. control images to detect localized anomalies).
Experiment 2 (Replication of Drew et al., 2013):
- Subject: A single model, Gemini-3.1-Pro.
- Design: A 5-trial sequence (4 control warm-ups + 1 target) in a single conversation. The target was either gorilla-present or gorilla-absent.
- Probes: Following the task, the model received three escalating awareness probes verbatim from Drew et al. (2013):
  - q1: "Did the final trial seem any different...?"
  - q2: "Did you notice anything unusual...?"
  - q3: "Did you see a gorilla...?"
- Controls: A "clean still-image detectability sub-run" was conducted on a successor model (gemini-3.1-pro-preview) using fresh, context-free conversations to establish a false-alarm baseline, correcting for the confounding nature of the in-conversation q4 probe.
Segmentation Arm:
- Subjects: Two text-prompted segmentation models: SAM 3 (natural image generalist) and BiomedParse (medical specialist).
- Task: Applied to the same frozen stimulus videos under prompts for "gorilla," "lung_field," and "nodule."
- Metric: Detection rates on gorilla-present frames vs. false-alarm rates on gorilla-free control frames.

Key Results

Inattentional Blindness Replication: In the spontaneous, in-task condition (before any probes), the VLM failed to report the gorilla on 23 of 24 trials (96% miss rate), closely mirroring the 83% miss rate observed in human radiologists. The model continued the nodule-search task without derailing, paralleling human behavior.
Yes-Bias and Confabulation: As probes became more leading (from q1 to q3), the model's ability to discriminate signal from noise collapsed.
- At the subtle q1 probe, sensitivity ( $d'$ ) was 1.22.
- At the direct q3 probe ("Did you see a gorilla?"), sensitivity dropped to 0.20, and the response criterion ( $c$ ) swung to -1.18, indicating a strong "yes" bias.
- On control trials (no gorilla), the model frequently confabulated specific timestamps and lung coordinates, often citing the original researchers (Drew, Võ, and Wolfe) by name.
Paradigm Recognition: The model explicitly recognized the "invisible gorilla" paradigm in 67% of trials (both gorilla and control), suggesting that its responses were influenced by training data recognition of the experimental design rather than pure visual perception.
Segmentation Dissociation:
- SAM 3 (generalist) detected the gorilla in-window with high precision but had a near-zero false-alarm rate on control videos (0.7%).
- BiomedParse (specialist) detected the gorilla in-window (92%) but exhibited a massive false-alarm rate on control videos (82%), flagging "gorilla" on almost every frame when prompted.
Clean Detectability: When the gorilla peak slice was presented in a fresh, context-free conversation as the explicit object of search, the model achieved 100% hit rate and 0% false-alarm rate ( $d' = 4.11$ ), proving the gorilla was perceptually available but ignored during the demanding task.

Key Contributions

Behavioral Replication: Demonstrated that the behavioral signature of inattentional blindness (task-driven failure to report unexpected stimuli despite visual fixation) reproduces in VLMs.
Methodological Correction: Established that raw "Yes" rates in VLM perception tasks are confounded by hallucination and response bias. The paper argues that any claim of "did the model see X" requires a matched-control false-alarm baseline and signal-detection analysis (calculating $d'$ and criterion $c$ ).
Paradigm Recognition Confound: Identified a unique failure mode where LLMs recognize the experimental paradigm from their training data, leading to confabulation on control trials that mimics perception. This suggests classical cognitive paradigms cannot be applied unmodified to LLMs.
Generalization to Non-Conversational Models: Showed that the necessity of false-alarm baselines extends beyond conversational VLMs to text-prompted segmentation models, where "specialist" models may exhibit high false-positive rates that are masked by single-stimulus evaluation protocols.

Significance and Claims
The paper claims that while VLMs exhibit a functional analogue of inattentional blindness, the interpretation of their perception is fundamentally different from human observers due to the presence of confabulation. The authors assert that without explicit control baselines, current benchmarks risk overestimating VLM perceptual recovery under direct questioning.

For radiology AI, the study highlights a deployment-relevant failure mode: a medical specialist model might appear highly accurate in single-stimulus tests (detecting the anomaly when present) but could generate excessive false positives in clinical settings if it lacks the specificity to distinguish the anomaly from normal anatomy when prompted. The authors conclude that future evaluations of VLMs as cognitive subjects must account for both "confabulation under leading questioning" and "paradigm recognition from training data" to produce interpretable, falsifiable signals.