Is your radiology AI sycophantic?

The word sycophant is Greek for “showing figs”—a gesture of calculated pleasing. In the world of AI, a sycophantic model is one that is frightened of silence. It forces an answer for every image, even when the data is inadequate, just to “please” the radiologist.

But in medicine, a guess is dangerous.

Large Language Models like Gemini 3.0 are finally learning the power of saying “I don’t know.” It’s time Radiology AI adopts the same epistemic humility.

Read more about why silence is a safety feature.

In the rapidly evolving world of large language models (LLMs), the biggest challenge hasn’t just been making them smarter; it’s been making them honest.

Early generations of generative AI were infamous for being “sycophantic.” Like a people-pleasing assistant terrified of disappointing its boss, if you asked an LLM a question based on a false premise, or one outside its training data, it wouldn’t admit ignorance. Instead, it would confidently hallucinate a plausible-sounding answer just to be “helpful.”

This changed significantly with the release of models like Gemini 3.0 from Google/DeepMind. A key technological breakthrough in Gemini 3.0 is what researchers call “epistemic tuning.” (Epistemology is a branch of philosophy concerned with “how we know what we know”; the Greek epistēmē literally means “to stand over”)

Put simply, DeepMind specifically trained the model to recognise the boundaries of its own knowledge. They utilised massive datasets designed to trap the AI with unanswerable questions or false premises, rewarding the model only when it refused to answer or admitted it didn’t know. They taught the AI intellectual humility. The result is a tool that is far more trustworthy, because when it speaks, you know it’s based on high-confidence data, not a probabilistic guess meant to appease you.

The Need for “Epistemic Humility” in Radiology

In the high-stakes environment of radiology, “hallucination” isn’t just an annoying quirk of a chatbot; it’s a potential patient safety issue. An AI that feels compelled to provide a diagnosis on a suboptimal image, or on pathology it wasn’t trained for, is a dangerous tool.

Radiology AI needs the same ability to say “I don’t know.”

While epistemic tuning in LLMs is the current hot topic, this concept of built-in uncertainty management has been pioneered in medical imaging by established tools like BoneXpert, an AI for automated bone age assessment.

BoneXpert achieves the same goal as Gemini’s epistemic tuning—knowing what it doesn’t know—though it arrives there through different, domain-specific means.

The BoneXpert Approach: Redundancy as Validation

Bone age assessment provides a unique opportunity for internal validation because the human hand offers biological redundancy. There isn’t just one indicator of maturity; there are many.

BoneXpert analyses the 21 tubular bones in the hand. Each of these bones can yield an independent assessment of biological maturity. The AI attempts to locate and interpret every single one of them.

Here is the crucial anti-sycophantic mechanism: BoneXpert will only report a bone age value to the user if it can reliably locate at least 8 of these bones and find consistency in their developmental stages.

If the input image makes this impossible, the AI rejects it. Common reasons for rejection include:

• Very abnormal bone morphology (such as achondroplasia), where the standard developmental rules don’t apply.
• Wrong pose, such as an oblique view instead of the standard flat hand, distorting the bone shapes.
• Wrong body part, where the image fed to the algorithm isn’t even a hand.

A sycophantic AI might look at an oblique view of a hand and force a reading based on distorted measurements because it was programmed to output a number. BoneXpert, however, recognises that its internal consistency checks have failed. It knows that it is outside its safe operational boundaries and refuses to guess.

Pioneering Trust in Medical AI

BoneXpert pioneered this feature in the radiology AI space long before generative models made “epistemic tuning” a headline. It recognised that an AI’s silence is sometimes its most valuable output.

A rejection from BoneXpert tells the radiologist: “The prerequisites for a reliable automated analysis are not met here; a human expert needs to take over.”

This functionality must become an industry standard for radiology AI solutions for triage and detection. For an AI to be truly clinical-grade, it cannot be a sycophant. Ultimate user trust and patient safety depend not only on how accurate the AI is when it works, but also on how it behaves when no answer should be given.

Read how BoneXpert avoids sycophancy here