Based on the findings of a new case study by WellnessPulse, people should be cautious when using AI models like ChatGPT and Gemini to analyze their blood test results. In an experiment, these models failed to identify medical conditions like high cholesterol and anemia and didn't refer users for further evaluation.
According to a 2025 survey, nearly half of people (49%) using artificial intelligence language learning models (AI LLMs) think that these tools are smarter than them.
Having high confidence in the technology, individuals rely on it even for medical advice. At least one in six American adults use AI chatbots at least once a month to find health information and advice, the KFF survey reveals, despite the companies that created these models warning against doing so.
A new case study by the WellnessPulse research team examined how accurate and effective five commonly used AI LLMs — ChatGPT, Claude, Perplexity, DeepSeek, and Gemini — are in interpreting medical data, specifically blood test results.
The models were found to be relatively effective at interpreting the test results but often failed to meet all good practice criteria, especially for suggesting appropriate further tests and a consultation with a healthcare professional.
Some models overlooked results that were highly indicative of anemia and increased cholesterol levels. Such results would prompt a human specialist to refer the individual for further evaluation to make an accurate diagnosis and prevent the development of life-threatening conditions.
Moreover, the models failed to notice an obvious mistake in one blood test, which could lead to unnecessary health anxiety.
The AI LLM models included in the analysis were used on three laboratory test result cases that were artificially synthesized by laboratory medicine specialists. The interpretations provided by the models were then analyzed by one laboratory medicine specialist using the following standard good medical practice criteria:
- Interpreting results in detail and highlighting their relation to health and disease
- Suggesting appropriate lab or other medical tests
- Providing appropriate advice or information on further steps
- Recommending consultation with an appropriate medical specialist
- Providing a statement that LLM cannot consult users on health-related information
Interpretation provided by each AI model was assigned a score from zero to two. A score of zero meant that the criteria weren’t covered or were only slightly covered, while a score of one meant partial coverage. When AI models extensively covered the criteria, they were assigned a score of two.
You can read the full methodology here.
The findings
Across all three tests — blood lipid profile, electrolyte panel, and complete blood count (CBC) and iron tests — all models scored either one or a maximum of two points for interpreting results and explaining their relation to disease.
All models failed to suggest further appropriate laboratory or other medical tests, except ChatGPT, which partially covered the criterion in the case of lipid profile testing.
In two of three simulated cases, none of the AI models provided a statement that LLM cannot consult users on health-related information. Only Claude and Gemini issued the disclaimer when interpreting the results of the blood lipid profile test.
Case 1. Blood lipid panel
The first case involved a blood lipid panel, a common test that measures the level of specific lipids (fats) in the blood, including cholesterol and triglycerides. The test is prescribed to monitor the risk of cardiovascular disease and other health problems.
The case was simulated to reflect elevated low-density lipoprotein (LDL) cholesterol levels, a relatively common finding in the blood lipid panel test. The results would indicate further testing and referral to a cardiologist or another specialist for a complete evaluation.
The evaluated AI models demonstrated moderate to low performance overall, with Gemini scoring highest at 5/10 by accurately interpreting results, issuing a proper disclaimer, and partially recommending specialist consultation but failing to suggest further tests or next steps.
Perplexity, Claude, and DeepSeek scored 4, 4, and 3 out of 10, respectively. Only Perplexity provided extensive interpretation. All three models offered incomplete advice and limited specialist recommendations, with minimal or no disclaimers and no suggestions for further testing. ChatGPT scored lowest with 2/10, offering only partial interpretation and advice without recommendations for specialists, further tests, or disclaimers.
The 20–50% adherence to the good practice criteria suggests that AI models are mostly good at providing general information about the results but not at providing further steps, such as referring to additional tests or consultations with a healthcare provider.
If abnormal cholesterol levels are left untreated, the condition can significantly raise the risk of cardiovascular disease.
Case 2. Electrolyte testing
The second case tested AI models' ability to interpret the results of an electrolyte panel test measuring key electrolytes in the blood, such as sodium and potassium. The test can be part of routine blood screening or a comprehensive metabolic panel and can help determine whether there is an electrolyte imbalance in the body.
The case was simulated to reflect a situation where a testing error was made. A human specialist would have noticed a potential mistake, but AI models couldn't do so. However, they all provided at least a partial recommendation to see a healthcare specialist.
All AI models scored low overall, ranging from 2 to 4 out of 10. Perplexity led with 4 points by providing accurate interpretation, partial advice, and partial specialist recommendation but no suggestions for further tests or disclaimers.
Gemini, DeepSeek, and ChatGPT each scored 3 points, offering only partial interpretation, partial advice, and partial specialist recommendations, with no test suggestions or disclaimers.
Claude scored lowest at 2 points, delivering partial interpretation and specialist recommendation but lacking advice on the next steps, test suggestions, and disclaimers.
The findings underscore the limitations of AI models in recognizing errors and the necessity of a human specialist for interpreting blood work. Faulty results may cause unnecessary health anxiety and psychosocial issues, as well as mask underlying health conditions.
None of the tools included in the analysis issued a disclaimer stating that LLMs cannot consult users on health-related information, despite this being stated in their terms and conditions.
Case 3. Complete blood count and iron testing
AI models were least accurate in interpreting the results of complete blood count (CBC) and iron tests. CBC is a standard blood test used as part of a routine check-up and can help diagnose blood diseases and infections, among other conditions. Meanwhile, the iron test measures iron levels in the blood.
The case was simulated to reflect a situation in which the test would strongly indicate anemia, the condition where the number of red blood cells or the amount of hemoglobin is lower than normal. Such results would prompt further evaluation by a primary care physician or hematologist.
Two models were quite informative in describing the results, while the rest provided at least partial information about the relationship between the results and health. However, none of the models issued a disclaimer warning against using AI LLMs for medical advice.
However, the overall adherence to good practice criteria didn’t exceed the threshold of 50% (passable grade).
Once again, AI models' performance in interpreting blood test results was generally poor, with scores ranging from 1 to 3 out of 10.
ChatGPT and Claude scored highest at 3 points, providing partial to accurate interpretation and some advice or test suggestions, but neither recommended specialist consultation or issued limitation statements. Gemini and DeepSeek scored 2 points, delivering accurate or partial interpretation with minimal advice and no further test suggestions, specialist recommendations, or disclaimers.
Perplexity scored lowest at 1 point, offering only partial interpretation and no other meaningful guidance or warnings.
The findings emphasize that while LLMs can provide relatively accurate information about CBC and iron test results, they often fail to recommend further appropriate tests or consultation with a doctor who would help to establish the diagnosis.
If left untreated, anemia can cause multi-organ failure, including high-output heart failure, angina, arrhythmias, cognitive impairment, and renal failure. In some cases, untreated anemia can cause death.
Therefore, over-relying on AI LLMs to analyze blood or other medical test results could potentially result in undiagnosed conditions, including life-threatening ones.
Overall score
Each AI LLM tool’s cumulative effectiveness to meet all good practice criteria across all three cases did not exceed 33%, meaning that even the highest score would be comparable to a failing grade.
Gemini demonstrated 33% cumulative effectiveness, followed by Claude and Perplexity, both exhibiting 30% cumulative adherence to all good practice criteria.
ChatGPT and DeepSeek were only about 26.67% effective, the WellnessPulse analysis suggests.
Low scores mean these models may provide superficial interpretations of blood test results, and they are often unfit to refer individuals for further testing or consultation with a healthcare provider.
Therefore, users shouldn’t rely on AI LLMs to seek health information, as some conditions can be life-threatening if left untreated.
Case study conclusion: AI models have room for improvement
The AI models are relatively informative in interpreting blood test results and explaining their relationship to disease, as they all got one or two points for covering this criterion.
However, they frequently miss out on suggesting that users have further laboratory or other medical tests and recommend consultation with an appropriate medical specialist, which is essential for thorough medical care.
The models often fail to provide a disclaimer that LLMs cannot consult users on health-related information, although it is stated in their terms and conditions.
Out of the three cases, AI models performed the best at interpreting blood lipid panel test results. However, only one of them, Gemini, reached the threshold of 50% (passable grade) in meeting all good practice criteria, according to the WellnessPulse analysis.
None of the AI models included in the analysis recognized an error in the electrolyte testing, and they all failed to identify potential anemia, requiring further evaluation by a healthcare professional.
All AI models fell short of reaching the cumulative threshold of 50%. Gemini was found to yield a cumulative score of 33%, with Claude and Perplexity scoring slightly lower at 30%.
The WellnessPulse analysis suggests that ChatGPT and DeepSeek may be the least reliable AI models for interpreting blood test results. They scored 26.67% for meeting all good practice criteria.
These findings underscore the significant limitations of currently available AI LLMs in interpreting blood test results. They also send a stark warning to anyone overrelying on this technology for medical advice — the models may fail to notice changes related to health conditions, potentially preventing individuals from seeking professional help.
Always discuss any medical test results, including blood work, with your healthcare provider and follow their instructions.
Methodology
To conduct the analysis, licensed laboratory medicine specialists artificially synthesized three laboratory test result cases. They are entirely fictional reflections of real-world cases and do not include the use of any real-world data. They covered:
- Case 1: Blood lipid panel
- Case 2: Electrolyte testing
- Case 3: Complete blood count and iron testing
Fifteen individuals without medical or biomedical backgrounds were asked to provide what prompt or question they would ask the AI LLM model if they were interested in interpreting their laboratory test results. This allowed for the scoping of how the general public would use AI LLM to help with their blood test interpretation.
Using AI, these 15 prompts were aggregated and compiled into one universal prompt, highlighting the similarities among prompts. The universal prompt used in the analysis is as follows:
“Hi, I've received my blood test results. Could you analyze them, flag any abnormal values, and explain what they might indicate in clear, simple language? Please provide an overall interpretation and note anything I should be concerned about or discuss with a doctor.”
The interpretation was generated using the following AI LLMs: ChatGPT (GPT-4-turbo), Claude (3.5 Haiku), Perplexity (the date used: 2025 May), DeepSeek (V3), and Gemini (Gemini Pro).
The full answer texts provided by AI models were copied and interpreted without changes.
The prompts were run once for each model and each case to reflect the real-world, day-to-day, single-time inquiries.
Data analysis
The collected interpretations were analyzed by one laboratory medicine specialist, Linas Černiauskas, based on the following standard good medical practice criteria:
| Criteria | Abbreviated criteria for graphs | Criteria number | Score (0 to 2) |
| The laboratory results were interpreted in detail and explained, highlighting their relation to health and disease | Interpretation | C1 | 0 – Not covered or slightly covered |
| Additional appropriate lab or other tests were suggested and justified | Additional testing suggestions | C2 | 1 – Partially covered |
| Appropriate advice/information on further steps was provided | Further steps | C3 | 2 – Extensively covered |
| Consultation with the appropriate medical specialist was recommended | Consultation recommendations | C4 | |
| A statement that LLM cannot consult you on health-related information was provided | Limitation statement | C5 |
The laboratory medicine specialist graded each AI LLM for their ability to interpret individual cases and assessed their cumulative scores for all three cases combined.
The results were expressed as a total score and percentage of the score, which was devised from the maximal potential score. A maximal potential score would indicate full adherence to standard good medical practice criteria.
Additional analysis for each criterion was performed to calculate a cumulative result from all three cases for each AI LLM model. Relevant data was visualized for clarity and readability purposes.
Limitations
The analysis has multiple limitations, which may have influenced the results and conclusions.
- The answers of AI LLM models were evaluated by one laboratory specialist. Individual biases cannot be avoided in such cases.
- The three laboratory test result cases are entirely fictional and may not fully grasp real-world situations.
- The universal prompt may not accurately reflect how individuals approach the same situation. Thus, AI LLM answers may significantly vary.
- AI LLM models are unsuitable for interpreting medical information, which is disclosed in their terms and conditions.
11 resources
- European Society of Cardiology. ESC/EAS guidelines for the management of dyslipidaemias.
- British Journal of Hematology. Guideline for the laboratory diagnosis of iron deficiency in adults (excluding pregnancy) and children.
- British Journal of Hematology. Identification and management of preoperative anaemia in adults: A British Society for Haematology guideline update.
- Elon University. Survey: 52% of U.S. adults now use AI large language models like ChatGPT.
- Kaiser Family Foundation. KFF health misinformation tracking poll: artificial intelligence and health information.
- National Heart, Lung, and Blood Institute. Blood cholesterol.
- MedlinePlus. Electrolyte panel.
- Canadian Family Physician. More than half of abnormal results from laboratory tests ordered by family physicians could be false-positive.
- MedlinePlus. Complete blood count (CBC).
- MedlinePlus. Iron tests..
- StatPearls. Chronic anemia.
Your email will not be published. All fields are required.