Can We Really Trust Medical Records Written by AI? The Blind Spot of AI That Cannot See 'Missing Information'

Imagine this: You visited the hospital and had a serious consultation with your doctor. After the visit, an AI assistant wrote your medical record for you. You read through it carefully and feel relieved because the doctor’s points seem well summarized. But what if a piece of critical information, such as the chest pain you started experiencing yesterday, is missing? Would it be safe to receive a prescription based on this incomplete record?

Recently, the adoption of ‘Ambient AI’ (medical scribes that listen to doctor-patient conversations and automatically draft medical records) has been increasing in hospital settings. While the convenience is significant, ‘omission’ errors, where critical information is unintentionally omitted from the record, remain a major hurdle to overcome. Ref 12 Today, we will easily break down why ‘AI judges’ introduced to solve this problem are not as smart as we think, along with their reasons and limitations.

Why is this important?

In a clinical setting, medical records are the most basic and core data for protecting a patient’s health. If critical symptoms are omitted from a record, there is a risk that a doctor might make a misdiagnosis or a prescription could be incorrect. To prevent this, AI is placed as a judge (LLM-as-a-Judge) instead of a human to inspect the records. Ref 10

But what happens if even these ‘AI judges’ cannot properly find missing information? The risk of medical accidents remains, and we face a serious situation where the AI assistant we use is creating ‘records with holes,’ and even the system evaluating it fails to discover those holes.

Understanding easily: Grading an exam without an ‘Answer Key’

To explain why AI judges cannot find missing information, let’s use the analogy of a ‘test grading’ situation.

Think of an AI judge as a ‘teacher grading a student’s answer sheet while holding the answer key.’

• Presence Verification: It is very easy to check if a student wrote “The answer to number 1 is A” on their answer sheet. This is because the letter ‘A’ is clearly visible on the page. AI has excellent capabilities in confirming whether specific keywords are included in records, like this. Ref 2
• Absence Verification: On the other hand, it is a different dimension for a teacher to check, “Did this student omit content they were supposed to write on the answer sheet?” To find content the student didn’t write, one must perfectly recall the entire answer key in their head and cross-check it line-by-line against everything on the student’s answer sheet.

According to the recently conducted ‘OmissionBench’ project, AI judges strongly verify ‘what is included’ in a record, but their performance in finding ‘what is missing’ is at a level similar to random guessing (chance levels). Ref 3, Ref 13 In other words, AI only looks at the ‘results’ contained in the record and has significantly insufficient capability to perceive the ‘empty spaces’ that are not recorded. Academia calls this ‘Omission Blindness.’

What is the current situation?

Many medical AI systems are already utilizing AI judges to evaluate the quality of medical records. Ref 10 However, the realistic performance is sobering. Research results show that approximately 3.45% of medical records actually written by AI contain information omission errors (hallucination errors are 1.47%). Ref 18

The problem is that the AI judges meant to filter out these omissions only look at ‘presence’ and cannot see ‘absence.’ Ref 2 Furthermore, the AI in charge of evaluation often possesses a similar thought process to the AI that created the record being inspected, leading them to repeat the same errors or simply overlook them. Ref 4

What will happen in the future?

As the limitations of AI judges become clear, various attempts are being made in the industry to overcome them.

1. Introduction of Deterministic Verification Tools: This approach does not rely solely on AI judgment but runs in parallel with rules written in simple, certain code, such as checking for essential keywords. Ref 4
2. Multi-Evaluation Frameworks: Instead of relying on a single AI judge, systems are being built to cross-verify information using multiple models or multi-agent systems. Ref 14
3. Human-in-the-loop: In the medical field, where safety is paramount, ‘human-centric evaluation,’ where human experts (doctors) finally confirm the results of AI reviews rather than having AI evaluate everything, remains the most important core. Ref 17

We are now at a point where we must carefully scrutinize not just “what AI can achieve,” but also “what AI is missing.”

MindTickleBytes AI Reporter’s View

Using AI as a judge is convenient, but distinguishing between ‘presence’ and ‘absence’ is a very different dimension of intelligence. We still have a long way to go before we can fully entrust our health to an AI that cannot read the ‘silence’ that has not been recorded.

References

1. [2608.31016] LLM Judges Verify Presence, Not Absence: Omission Blindness in AI Clinical Notes and What Recovers It (https://arxiv.org/abs/2608.31016)
2. LLM Judges Verify Presence, Not Absence: Omission Blindness in AI Clinical Notes and What Recovers It (https://arxiv.org/pdf/2608.31016)
3. GitHub - composo-ai/omission-bench: OmissionBench harness: code (https://github.com/composo-ai/omission-bench)
4. Replace Your LLM Judge With 10 Lines of pytest - YouTube (https://www.youtube.com/watch?v=BPXFDC7WHSk)
5. LLM-as-a-judge: a complete guide to using LLMs for evaluations (https://www.evidentlyai.com/llm-guide/llm-as-a-judge)
6. LLM-as-a-Judge Simply Explained: The Complete Guide (https://www.confident-ai.com/blog/why-llm-as-a-judge-is-the-best-llm-evaluation-method)
7. Position Bias in LLM Judges: Measurement and Mitigation (https://mbrenndoerfer.com/writing/position-bias-in-llm-judges)
8. LLMs bow to pressure, changing answers when challenged (https://www.computerworld.com/article/4023989/llms-bow-to-pressure-changing-answers-when-challenged-deepmind-study.html)
9. Continual Monitoring of Note Quality At Scale (https://www.linkedin.com/posts/catherine-chen-5851a6a0continual-monitoring-of-note-quality-at-scale-activity-7496283957693448192-ksR)
10. LLM Judges Are Unreliable (https://www.cip.org/blog/llm-judges-are-unreliable)
11. LLM Judges Verify Presence, Not Absence: Omission Blindness in AI Clinical Notes (https://arxiv.org/abs/2608.31016v1)
12. LLM Judges Verify Presence, Not Absence (https://arxiv.org/html/2608.31016v1)
13. Evaluating clinical AI summaries with large language models as judges (https://www.nature.com/articles/s41746-025-02005-2)
14. Evaluating medical AI under missing information (https://arxiv.org/html/2607.18828)
15. A framework to assess clinical safety and hallucination rates of LLMs for medical text summarisation (https://www.nature.com/articles/s41746-025-01670-7)