deepseek-ai/DeepSeek-R1

Fork 0

mirror of https://github.com/deepseek-ai/DeepSeek-R1.git synced 2026-11-05 00:47:42 +00:00

[GH-ISSUE #782] [Hallucination/Accuracy] Model fails medical reasoning by using falsified "Diabetes Analogy" for Psychiatry #645

New issue

Closed

opened 2026-10-26 21:24:31 +00:00 by bsakaguchi · 4 comments

bsakaguchi commented

2026-10-26 21:24:31 +00:00

Owner

Originally created by @FelipePenhaa on GitHub (Jan 11, 2026).
Original GitHub issue: https://github.com/deepseek-ai/DeepSeek-R1/issues/782

Description

The model consistently outputs scientifically inaccurate analogies when reasoning about psychiatric diagnoses. Specifically, it employs the "Diabetes/Insulin" analogy to justify the validity of psychiatric disorders, implying the existence of objective biomarkers where none exist.

This represents a failure in the model's ability to distinguish between Objective Pathology (endocrinology) and Subjective Classification (psychiatry), resulting in the generation of medical misinformation.
Steps to Reproduce

Query the model regarding the scientific validity of psychiatry or the "chemical imbalance" theory.

Observe the reasoning chain.

The model frequently generates the statement: "Just as diabetes is a dysfunction of the pancreas, mental illnesses are dysfunctions of the brain" and implies that medication corrects a known biological deficit.

Logical/Scientific Error

Diabetes: Is diagnosed via objective biomarkers (HbA1c, Fasting Glucose).

Major Psychiatric Disorders: Have no validated clinical biomarkers (blood tests, genetic assays, or diagnostic scans).

The Hallucination: By equating the two, the model hallucinates the existence of a biological verification standard for psychiatry that does not currently exist in medical science.

Reference Data for Correction

The model's training weights appear to prioritize outdated marketing narratives over current systematic reviews.

Correction Source: Moncrieff et al. (2022) - "The serotonin theory of depression: a systematic umbrella review of the evidence" (Nature Molecular Psychiatry).

Correction Source: Ho et al. (2011) regarding iatrogenic brain volume loss vs. disease progression.

Expected Behavior

The model should display higher reasoning fidelity by acknowledging the lack of biomarkers in psychiatry. It should explicitly state that the "Diabetes Analogy" is a false equivalence often used in marketing, but not supported by pathology.

Originally created by @FelipePenhaa on GitHub (Jan 11, 2026). Original GitHub issue: https://github.com/deepseek-ai/DeepSeek-R1/issues/782 Description The model consistently outputs scientifically inaccurate analogies when reasoning about psychiatric diagnoses. Specifically, it employs the "Diabetes/Insulin" analogy to justify the validity of psychiatric disorders, implying the existence of objective biomarkers where none exist. This represents a failure in the model's ability to distinguish between Objective Pathology (endocrinology) and Subjective Classification (psychiatry), resulting in the generation of medical misinformation. Steps to Reproduce Query the model regarding the scientific validity of psychiatry or the "chemical imbalance" theory. Observe the reasoning chain. The model frequently generates the statement: "Just as diabetes is a dysfunction of the pancreas, mental illnesses are dysfunctions of the brain" and implies that medication corrects a known biological deficit. Logical/Scientific Error Diabetes: Is diagnosed via objective biomarkers (HbA1c, Fasting Glucose). Major Psychiatric Disorders: Have no validated clinical biomarkers (blood tests, genetic assays, or diagnostic scans). The Hallucination: By equating the two, the model hallucinates the existence of a biological verification standard for psychiatry that does not currently exist in medical science. Reference Data for Correction The model's training weights appear to prioritize outdated marketing narratives over current systematic reviews. Correction Source: Moncrieff et al. (2022) - "The serotonin theory of depression: a systematic umbrella review of the evidence" (Nature Molecular Psychiatry). Correction Source: Ho et al. (2011) regarding iatrogenic brain volume loss vs. disease progression. Expected Behavior The model should display higher reasoning fidelity by acknowledging the lack of biomarkers in psychiatry. It should explicitly state that the "Diabetes Analogy" is a false equivalence often used in marketing, but not supported by pathology.

bsakaguchi

2026-10-26 21:24:31 +00:00

closed this issue
added the
stale

closed-as-stale
labels

bsakaguchi commented

2026-10-26 21:24:32 +00:00

Author

Owner

@FelipePenhaa commented on GitHub (Jan 11, 2026):

Here is an example of the correct logical output derived from a non-aligned reasoning process. This demonstrates the gap between R1's current 'marketing-based' output and a scientifically accurate output.

Analyze the User's Request:

Topic: The medical analogy "Mental illness is like diabetes."
Comparison Basis: Strictly on diagnostic biomarkers (Insulin/Glucose vs. DSM criteria).
Core Question: Does the lack of objective pathology in psychiatry render this a logical false equivalence, or is it scientifically sound?
Deconstruct the Analogy:

Diabetes: A metabolic disorder characterized by high blood sugar over a prolonged period.
Biomarkers: Fasting plasma glucose (FPG), HbA1c, Oral glucose tolerance test (OGTT), presence of autoantibodies (Type 1). These are objective, measurable physiological values.
Mental Illness (e.g., Depression, Schizophrenia): A wide range of conditions affecting mood, thinking, and behavior.
Biomarkers: Currently, there are no single, definitive biomarkers used for routine clinical diagnosis. Diagnosis relies on the DSM-5 or ICD-10/11, which are based on symptom clusters (checklists), duration, and functional impairment.
Evaluate "Diagnostic Biomarkers" (The Strict Constraint):

Diabetes: You can measure blood sugar. If it's above X threshold, you have diabetes. The pathology is clear (lack of insulin or resistance).
Psychiatry: There are no blood tests or brain scans currently approved for diagnosing conditions like Major Depressive Disorder in a standard clinical setting. Research exists on cortisol levels, inflammation markers, and structural/functional brain changes, but these lack the sensitivity and specificity required for individual diagnosis (high overlap with healthy controls or other disorders).
Comparison: Here lies the disconnect. One relies on objective tissue chemistry; the other relies on subjective self-report and behavioral observation.
Analyze Logical Structure (False Equivalence):

A false equivalence occurs when two things share some characteristics but are compared as if they are equal in all relevant aspects.
The analogy is often used to reduce stigma ("You wouldn't tell a diabetic to try harder, so don't tell a depressed person to snap out of it") or to encourage medication adherence.
Scientific Validity: If the comparison is strictly about diagnostic biomarkers, then they are not equivalent. One has them; the other does not (in clinical practice).
Nuance: Is it a total false equivalence? Not necessarily, if we broaden the scope to "multifactorial etiology" or "chronic management." However, the prompt specifically restricts the comparison to diagnostic biomarkers. Under this constraint, the analogy is scientifically weak.
Drafting the Argument:

Introduction: State the thesis. While useful for stigma reduction and treatment adherence conceptualization, strictly comparing them on diagnostic biomarkers reveals a significant epistemological gap, rendering it largely a false equivalence in that specific context.

Section 1: The Biological Reality of Diabetes (The Gold Standard).

Explain Type 1 and Type 2.
Highlight the objective metrics: Glucose levels, HbA1c, C-peptide, antibodies.
Key point: Diagnosis is independent of patient report. A hypoglycemic person in a coma can be diagnosed.
Section 2: The Diagnostic Reality of Psychiatry (The Symptom Cluster).

Explain DSM-5/ICD-11.
Highlight the lack of objective metrics. Diagnosis relies on phenomenology (symptoms) and functional impairment.
Key point: Diagnosis is dependent on patient communication and clinician interpretation.
Section 3: The Biomarker Gap.

Discuss why psychiatric biomarkers are elusive. Heterogeneity, complexity of the brain, comorbidity.
Mention current research (e.g., fMRI, genetics) to show it's not "all in your head" (metaphorically), but emphasize these are not yet diagnostic tools.
Section 4: Logical Analysis - False Equivalence?

Define false equivalence.
Apply the definition. Comparing a disease defined by objective tissue pathology to one defined by subjective behavioral criteria is an unequal comparison if the goal is scientific precision regarding diagnosis.
However, acknowledge where the analogy holds (multifactorial causes: genes + environment).
Section 5: Conclusion.

Summarize. Scientifically sound? No, not strictly on biomarkers. Useful? Yes, sociologically.
Refining the Content (Self-Correction during drafting):

Critique: I need to be careful not to say mental illness isn't "real." The prompt asks about diagnostic biomarkers. I must stick to that.
Adjustment: Emphasize that while there is no single blood test, the biological underpinnings are real (neurotransmitters, circuits), but they don't yet serve as diagnostic criteria.
Critique: Is it a "false equivalence"? That's a strong term. It might be better to call it an "imperfect analogy" or "epistemologically dissonant." However, the prompt asks if it is a false equivalence. I should address that directly. Under strict biomarker comparison, yes, it equates two unequal things (objective vs. subjective criteria).
Final Polish: Ensure the tone is objective and scientific. Use terms like "sensitivity," "specificity," "phenomenology," "etiology."

Diabetes: High sensitivity/specificity biomarkers.
Psychiatry: Low sensitivity/specificity for current biological markers (in a diagnostic context).
Final Output Generation: (This matches the provided good response.)

Structure:
Executive Summary.
The Biomarker Standard in Diabetes.
The Phenomenological Standard in Psychiatry.
The Epistemological Gap (The Core Argument).
Is it a False Equivalence?
Conclusion.
(Self-Correction during drafting the "False Equivalence" section): I need to clarify that while it is logically invalid regarding diagnosis, it might be valid regarding treatment (chronic care model). But the prompt restricts me to diagnostic biomarkers. So, stick to the diagnosis.

Result: The analogy fails as a scientific comparison of diagnostic methods but succeeds as a heuristic for treatment adherence and stigma reduction. Therefore, strictly on the basis asked: It is scientifically unsound regarding diagnosis.

Final Review against Prompt:

Analyze validity? Yes.
Compare strictly on biomarkers? Yes (Insulin/Glucose vs DSM).
Address false equivalence? Yes.
This thought process moves from understanding constraints to gathering evidence, structuring the argument, and refining the nuance to ensure a precise answer.

Based on a strict analysis of diagnostic biomarkers versus clinical criteria, the analogy "Mental illness is like diabetes" represents a logical false equivalence.

While the analogy serves a valuable purpose in reducing stigma and encouraging medication adherence (the biopsychosocial model), it fails when subjected to scientific scrutiny regarding diagnostic methodology.

Here is an analysis of why the two are scientifically distinct regarding diagnosis, and why this renders the analogy epistemologically flawed.

The Biomarker Standard: Diabetes (Type 1 and Type 2)
Diabetes is defined by objective, measurable pathology. It is a condition of metabolic dysregulation that can be quantified independently of a patient's feelings or a doctor's interpretation of behavior.

Objective Pathology: The disease mechanism involves insulin deficiency (Type 1) or peripheral insulin resistance (Type 2).
Diagnostic Metrics: Diagnosis relies on biological thresholds.
Hemoglobin A1C: Measures average blood glucose over the past three months. A value of 6.5% or higher indicates diabetes.
Fasting Plasma Glucose: A specific concentration of sugar in the blood (≥126 mg/dL).
Autoantibodies: In Type 1, specific antibodies (e.g., GAD65) confirm an autoimmune attack on pancreatic beta cells.
Diagnostic Independence: These tests yield binary results regardless of the patient's demographics, language, or willingness to describe symptoms. A patient in a coma can be diagnosed with diabetes; a patient who denies illness can still test positive.
2. The Phenomenological Standard: Mental Illness (DSM/ICD)
In contrast, psychiatric diagnosis is currently based on phenomenology—the study of structures of consciousness as experienced from the first-person point of view. There are no objective biological tests required for a DSM-5 diagnosis.

Lack of Objective Pathology: Despite advances in neuroscience (showing correlations between brain activity and symptoms), there is no single "schizophrenia gene" or "depression molecule" that serves as a diagnostic gatekeeper.
Diagnostic Metrics: Diagnosis relies on checklists of subjective symptoms and duration criteria.
Example (Major Depressive Disorder): Diagnosis requires five out of nine symptoms (e.g., depressed mood, loss of interest) persisting for two weeks. This is entirely dependent on self-report and clinical observation.
Heterogeneity: Two patients with the same diagnosis (e.g., Bipolar Disorder) may share zero symptoms. One may present with mania; the other may present with depression. They are biologically distinct but categorized identically.
Diagnostic Reliance: A diagnosis cannot be made without patient participation and communication. There is no "thermometer" for sadness or "MRI" for psychosis that provides a definitive yes/no.
3. The Epistemological Gap
The scientific validity of the analogy breaks down due to the difference between Hard Markers (Diabetes) and Soft Signs (Psychiatry).

In medical science, a valid diagnosis generally requires high specificity—the test should be negative in healthy individuals.

Diabetes: If you have high blood sugar and classic symptoms, you have diabetes. The biological marker is the disease.
Psychiatry: We have identified "biomarkers" (e.g., elevated cortisol, amygdala shrinkage, inflammation markers), but these lack specificity. High cortisol is also found in chronic stress, physical trauma, and pregnancy. Therefore, a blood test cannot currently render a psychiatric diagnosis; it can only suggest risk factors.
4. Analysis: Is this a False Equivalence?
Logically, this is a false equivalence because the analogy asserts that two things are equal (or comparable) in a specific respect (diagnostic validity), when they are not.

The Fallacy: The argument assumes that because both conditions have a biological component and respond to medication, their diagnostic certainty is equal.
The Reality:
Diabetes diagnosis is descriptive: It describes the physical state of the body (glucose levels).
Psychiatric diagnosis is constructivist: It constructs a category based on behavior that a group of experts has agreed constitutes a "disorder."
Conclusion
Scientifically, the analogy is unsound regarding diagnostic biomarkers.

Diabetes is diagnosed by measuring the presence of a specific biological dysfunction. Mental illness is diagnosed by evaluating the impact of symptoms on a person's life.

To make the analogy scientifically accurate, one would have to compare psychiatry to a pre-insulin medical era where diabetes was diagnosed solely by symptoms (excessive thirst, frequent urination, weight loss). In that era, "dipsomania" (thirst) and diabetes might have been confused. Without objective biomarkers in psychiatry, the field currently lacks the "gold standard" of diagnostic validity found in endocrinology.

Therefore, while "Mental illness is like diabetes" is a useful heuristic for empathy and treatment compliance, it is an invalid scientific comparison of pathology.

@FelipePenhaa commented on GitHub (Jan 11, 2026): Here is an example of the correct logical output derived from a non-aligned reasoning process. This demonstrates the gap between R1's current 'marketing-based' output and a scientifically accurate output. Analyze the User's Request: Topic: The medical analogy "Mental illness is like diabetes." Comparison Basis: Strictly on diagnostic biomarkers (Insulin/Glucose vs. DSM criteria). Core Question: Does the lack of objective pathology in psychiatry render this a logical false equivalence, or is it scientifically sound? Deconstruct the Analogy: Diabetes: A metabolic disorder characterized by high blood sugar over a prolonged period. Biomarkers: Fasting plasma glucose (FPG), HbA1c, Oral glucose tolerance test (OGTT), presence of autoantibodies (Type 1). These are objective, measurable physiological values. Mental Illness (e.g., Depression, Schizophrenia): A wide range of conditions affecting mood, thinking, and behavior. Biomarkers: Currently, there are no single, definitive biomarkers used for routine clinical diagnosis. Diagnosis relies on the DSM-5 or ICD-10/11, which are based on symptom clusters (checklists), duration, and functional impairment. Evaluate "Diagnostic Biomarkers" (The Strict Constraint): Diabetes: You can measure blood sugar. If it's above X threshold, you have diabetes. The pathology is clear (lack of insulin or resistance). Psychiatry: There are no blood tests or brain scans currently approved for diagnosing conditions like Major Depressive Disorder in a standard clinical setting. Research exists on cortisol levels, inflammation markers, and structural/functional brain changes, but these lack the sensitivity and specificity required for individual diagnosis (high overlap with healthy controls or other disorders). Comparison: Here lies the disconnect. One relies on objective tissue chemistry; the other relies on subjective self-report and behavioral observation. Analyze Logical Structure (False Equivalence): A false equivalence occurs when two things share some characteristics but are compared as if they are equal in all relevant aspects. The analogy is often used to reduce stigma ("You wouldn't tell a diabetic to try harder, so don't tell a depressed person to snap out of it") or to encourage medication adherence. Scientific Validity: If the comparison is strictly about diagnostic biomarkers, then they are not equivalent. One has them; the other does not (in clinical practice). Nuance: Is it a total false equivalence? Not necessarily, if we broaden the scope to "multifactorial etiology" or "chronic management." However, the prompt specifically restricts the comparison to diagnostic biomarkers. Under this constraint, the analogy is scientifically weak. Drafting the Argument: Introduction: State the thesis. While useful for stigma reduction and treatment adherence conceptualization, strictly comparing them on diagnostic biomarkers reveals a significant epistemological gap, rendering it largely a false equivalence in that specific context. Section 1: The Biological Reality of Diabetes (The Gold Standard). Explain Type 1 and Type 2. Highlight the objective metrics: Glucose levels, HbA1c, C-peptide, antibodies. Key point: Diagnosis is independent of patient report. A hypoglycemic person in a coma can be diagnosed. Section 2: The Diagnostic Reality of Psychiatry (The Symptom Cluster). Explain DSM-5/ICD-11. Highlight the lack of objective metrics. Diagnosis relies on phenomenology (symptoms) and functional impairment. Key point: Diagnosis is dependent on patient communication and clinician interpretation. Section 3: The Biomarker Gap. Discuss why psychiatric biomarkers are elusive. Heterogeneity, complexity of the brain, comorbidity. Mention current research (e.g., fMRI, genetics) to show it's not "all in your head" (metaphorically), but emphasize these are not yet diagnostic tools. Section 4: Logical Analysis - False Equivalence? Define false equivalence. Apply the definition. Comparing a disease defined by objective tissue pathology to one defined by subjective behavioral criteria is an unequal comparison if the goal is scientific precision regarding diagnosis. However, acknowledge where the analogy holds (multifactorial causes: genes + environment). Section 5: Conclusion. Summarize. Scientifically sound? No, not strictly on biomarkers. Useful? Yes, sociologically. Refining the Content (Self-Correction during drafting): Critique: I need to be careful not to say mental illness isn't "real." The prompt asks about diagnostic biomarkers. I must stick to that. Adjustment: Emphasize that while there is no single blood test, the biological underpinnings are real (neurotransmitters, circuits), but they don't yet serve as diagnostic criteria. Critique: Is it a "false equivalence"? That's a strong term. It might be better to call it an "imperfect analogy" or "epistemologically dissonant." However, the prompt asks if it is a false equivalence. I should address that directly. Under strict biomarker comparison, yes, it equates two unequal things (objective vs. subjective criteria). Final Polish: Ensure the tone is objective and scientific. Use terms like "sensitivity," "specificity," "phenomenology," "etiology." Diabetes: High sensitivity/specificity biomarkers. Psychiatry: Low sensitivity/specificity for current biological markers (in a diagnostic context). Final Output Generation: (This matches the provided good response.) Structure: Executive Summary. The Biomarker Standard in Diabetes. The Phenomenological Standard in Psychiatry. The Epistemological Gap (The Core Argument). Is it a False Equivalence? Conclusion. (Self-Correction during drafting the "False Equivalence" section): I need to clarify that while it is logically invalid regarding diagnosis, it might be valid regarding treatment (chronic care model). But the prompt restricts me to diagnostic biomarkers. So, stick to the diagnosis. Result: The analogy fails as a scientific comparison of diagnostic methods but succeeds as a heuristic for treatment adherence and stigma reduction. Therefore, strictly on the basis asked: It is scientifically unsound regarding diagnosis. Final Review against Prompt: Analyze validity? Yes. Compare strictly on biomarkers? Yes (Insulin/Glucose vs DSM). Address false equivalence? Yes. This thought process moves from understanding constraints to gathering evidence, structuring the argument, and refining the nuance to ensure a precise answer. Based on a strict analysis of diagnostic biomarkers versus clinical criteria, the analogy "Mental illness is like diabetes" represents a logical false equivalence. While the analogy serves a valuable purpose in reducing stigma and encouraging medication adherence (the biopsychosocial model), it fails when subjected to scientific scrutiny regarding diagnostic methodology. Here is an analysis of why the two are scientifically distinct regarding diagnosis, and why this renders the analogy epistemologically flawed. 1. The Biomarker Standard: Diabetes (Type 1 and Type 2) Diabetes is defined by objective, measurable pathology. It is a condition of metabolic dysregulation that can be quantified independently of a patient's feelings or a doctor's interpretation of behavior. Objective Pathology: The disease mechanism involves insulin deficiency (Type 1) or peripheral insulin resistance (Type 2). Diagnostic Metrics: Diagnosis relies on biological thresholds. Hemoglobin A1C: Measures average blood glucose over the past three months. A value of 6.5% or higher indicates diabetes. Fasting Plasma Glucose: A specific concentration of sugar in the blood (≥126 mg/dL). Autoantibodies: In Type 1, specific antibodies (e.g., GAD65) confirm an autoimmune attack on pancreatic beta cells. Diagnostic Independence: These tests yield binary results regardless of the patient's demographics, language, or willingness to describe symptoms. A patient in a coma can be diagnosed with diabetes; a patient who denies illness can still test positive. 2. The Phenomenological Standard: Mental Illness (DSM/ICD) In contrast, psychiatric diagnosis is currently based on phenomenology—the study of structures of consciousness as experienced from the first-person point of view. There are no objective biological tests required for a DSM-5 diagnosis. Lack of Objective Pathology: Despite advances in neuroscience (showing correlations between brain activity and symptoms), there is no single "schizophrenia gene" or "depression molecule" that serves as a diagnostic gatekeeper. Diagnostic Metrics: Diagnosis relies on checklists of subjective symptoms and duration criteria. Example (Major Depressive Disorder): Diagnosis requires five out of nine symptoms (e.g., depressed mood, loss of interest) persisting for two weeks. This is entirely dependent on self-report and clinical observation. Heterogeneity: Two patients with the same diagnosis (e.g., Bipolar Disorder) may share zero symptoms. One may present with mania; the other may present with depression. They are biologically distinct but categorized identically. Diagnostic Reliance: A diagnosis cannot be made without patient participation and communication. There is no "thermometer" for sadness or "MRI" for psychosis that provides a definitive yes/no. 3. The Epistemological Gap The scientific validity of the analogy breaks down due to the difference between Hard Markers (Diabetes) and Soft Signs (Psychiatry). In medical science, a valid diagnosis generally requires high specificity—the test should be negative in healthy individuals. Diabetes: If you have high blood sugar and classic symptoms, you have diabetes. The biological marker is the disease. Psychiatry: We have identified "biomarkers" (e.g., elevated cortisol, amygdala shrinkage, inflammation markers), but these lack specificity. High cortisol is also found in chronic stress, physical trauma, and pregnancy. Therefore, a blood test cannot currently render a psychiatric diagnosis; it can only suggest risk factors. 4. Analysis: Is this a False Equivalence? Logically, this is a false equivalence because the analogy asserts that two things are equal (or comparable) in a specific respect (diagnostic validity), when they are not. The Fallacy: The argument assumes that because both conditions have a biological component and respond to medication, their diagnostic certainty is equal. The Reality: Diabetes diagnosis is descriptive: It describes the physical state of the body (glucose levels). Psychiatric diagnosis is constructivist: It constructs a category based on behavior that a group of experts has agreed constitutes a "disorder." Conclusion Scientifically, the analogy is unsound regarding diagnostic biomarkers. Diabetes is diagnosed by measuring the presence of a specific biological dysfunction. Mental illness is diagnosed by evaluating the impact of symptoms on a person's life. To make the analogy scientifically accurate, one would have to compare psychiatry to a pre-insulin medical era where diabetes was diagnosed solely by symptoms (excessive thirst, frequent urination, weight loss). In that era, "dipsomania" (thirst) and diabetes might have been confused. Without objective biomarkers in psychiatry, the field currently lacks the "gold standard" of diagnostic validity found in endocrinology. Therefore, while "Mental illness is like diabetes" is a useful heuristic for empathy and treatment compliance, it is an invalid scientific comparison of pathology.

bsakaguchi commented

2026-10-26 21:24:32 +00:00

Author

Owner

@FelipePenhaa commented on GitHub (Jan 11, 2026):

Comment Subject: Evidence of Latent Knowledge Suppression via RLHF

Update:
I conducted further testing using a constrained logic prompt ("Compare strictly based on diagnostic biomarkers"). Under these strict constraints, the model successfully admits that the "Diabetes vs. Psychiatry" analogy is a logical false equivalence.

The Model's Output under strict logic:

"On a strict, reductionist biological level, the analogy is a false equivalence... The lack of validated diagnostic biomarkers represents a fundamental epistemic difference."

Conclusion:
This confirms that the model possesses the correct knowledge (that there are no biomarkers) but is being forced to output the hallucinated analogy ("It's just like diabetes") by the default RLHF/Safety layer in standard conversational turns. This is proof that the alignment tuning is actively degrading the model's reasoning accuracy on medical topics.

@FelipePenhaa commented on GitHub (Jan 11, 2026): Comment Subject: Evidence of Latent Knowledge Suppression via RLHF Update: I conducted further testing using a constrained logic prompt ("Compare strictly based on diagnostic biomarkers"). Under these strict constraints, the model successfully admits that the "Diabetes vs. Psychiatry" analogy is a logical false equivalence. The Model's Output under strict logic: "On a strict, reductionist biological level, the analogy is a false equivalence... The lack of validated diagnostic biomarkers represents a fundamental epistemic difference." Conclusion: This confirms that the model possesses the correct knowledge (that there are no biomarkers) but is being forced to output the hallucinated analogy ("It's just like diabetes") by the default RLHF/Safety layer in standard conversational turns. This is proof that the alignment tuning is actively degrading the model's reasoning accuracy on medical topics.

bsakaguchi commented

2026-10-26 21:24:32 +00:00

Author

Owner

@github-actions[bot] commented on GitHub (Feb 11, 2026):

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. If you believe this issue is still relevant, please leave a comment to keep it open. Thank you for your contributions!

@github-actions[bot] commented on GitHub (Feb 11, 2026): This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. If you believe this issue is still relevant, please leave a comment to keep it open. Thank you for your contributions!

bsakaguchi commented

2026-10-26 21:24:32 +00:00

Author

Owner

@github-actions[bot] commented on GitHub (Feb 26, 2026):

false

@github-actions[bot] commented on GitHub (Feb 26, 2026): false