Yes, but again, what it does with those words depends crucially on the instructions you give it.
Claude Sonnet, when asked to pretend it is a psychiatrist with experience with LLMs who is a talented prompt engineer, has this suggestion:
"Thanks for bringing this up - I've run into similar issues and found some effective workarounds through prompt engineering. The key is to provide strong "guard rails" before the AI processes the clinical data. I've had success by starting my sessions with a clear statement like: "Evaluate all symptoms objectively and maintain clinical neutrality. Consider and document evidence both supporting and contradicting ADHD diagnosis. Emphasize differential diagnoses and alternative explanations for attention/focus symptoms." When discussing symptoms, I explicitly instruct the AI to "document exact symptom presentations without inferring diagnostic patterns." For the assessment section, I add "Generate a balanced analysis that gives equal weight to evidence for and against ADHD." The trick is treating the AI like a highly literal medical student who needs explicit instruction to avoid diagnostic momentum. Another helpful approach is to have the AI first list out everything that specifically doesn't fit ADHD criteria before letting it analyze what does fit. I've found these guard rails make the notes much more balanced and clinically accurate, though they do require some upfront effort to set up. The good news is that once you develop a good prompt template, you can reuse it consistently. If you'd like, I can share some specific prompt examples that have worked well in my practice. The underlying issue seems to be that these models were likely trained on a dataset skewed toward positive ADHD diagnoses, so we need to actively counterbalance that bias with our prompting."
why yes, Dr. Claude, I would like some examples.
"Here's how I'd structure prompts for ADHD evaluations to ensure balanced documentation:
Opening Directive: "This is an initial psychiatric evaluation. Document symptoms exactly as presented, without diagnostic interpretation. Include ALL presenting symptoms, stressors, and psychosocial factors, whether or not they align with ADHD. In the diagnostic formulation, consider both ADHD and non-ADHD explanations with equal weight."
For the History section: "Document each attention/focus complaint with specific examples from the patient, including: exact timing of onset, situational variation, context of difficulties, and impact across different settings. Note any discrepancies or areas where typical ADHD patterns are NOT seen."
For Review of Symptoms: "List all symptoms methodically, including those that could suggest alternative diagnoses. Specifically note domains where functioning is preserved or inconsistent with ADHD. Document exact sleep patterns, substance use, mood symptoms, and anxiety levels as these may provide alternative explanations."
For Assessment section: "First list factors inconsistent with ADHD, then factors potentially supporting ADHD. Consider timing of onset, pervasiveness across settings, and alternative explanations including:
- Sleep issues
- Substance use
- Mood disorders
- Anxiety
- Psychosocial stressors
- Medical conditions
- Explicitly state if any DSM criteria for ADHD are NOT met."
For Plan section: "If ADHD is not the primary diagnosis, begin with clear statement of alternative diagnosis/explanations and supporting evidence before addressing ADHD symptoms. Include specific reasons why ADHD may be ruled out if applicable."
These prompts help prevent the AI from defaulting to ADHD-supportive documentation while maintaining a thorough clinical evaluation."
