AI and the ethics of techno-solutionism in pain management

Young et al29 examined 2 large language models' (LLMs) pain management recommendations for the 4 most common reasons for pain-related emergency department visits. Their goal was to understand if LLMs recommendations for opioid treatment vary based on patient race/ethnicity and sex, and whether LLMs eliminate or worsen biases. Using 40 real patient case descriptions, and omitting patients' sex, race, ethnicity, and current medication use from the model, the authors instructed LLMs (GPT-4 and Gemini) to provide subjective pain ratings (severity and rating from 1 to 10) and pharmacologic treatment recommendations. The authors found that although there were discrepancies between LLMs, they did not show preferential opioid treatment for one group over another based on combinations of race/ethnicity or sex. Although Young et al.29 note that there is evidence that artificial intelligence (AI) systems like LLMs can exacerbate race-based inequities, they suggest that LLMs may help mitigate clinician bias and support equitable pain management. I commend the authors for tackling this critical issue. However, I argue that the appeal of using objective AI-based solutions like LLMs to overcome complex socio-structural injustices in pain management reflects a techno-solutionism that will likely not remedy the problems it seeks to solve and could have unintended ethical consequences. There is widespread enthusiasm that AI systems and precision medicine generally will improve healthcare decision making. Although research on state-of-the-art LLMs demonstrate that they can reproduce several tasks, the evidence for clinical impact remains limited.12 Part of the motivation to use AI systems in pain management is to overcome the perceived limitations associated with uncertainty, subjectivity, and invisibility of pain in a medical culture that prioritizes certainty, objectivity, and visibility.4,5,9 For example, researchers are combining neuroimaging with forms of AI such as machine learning to identify a brain-based biomarker of chronic pain.7,22,27 These research programs are based on assumptions of mechanical objectivity, where objectivity is achieved with standardized methods, mechanical or automated processes, and instruments that attempt to minimize human bias and subjectivity. The outcomes are therefore considered more reliable and unbiased.6,9,25 The idea that LLMs could potentially fix clinician biases and inequities in pain management related to sexism, racism, and other systems of oppression is understandably appealing. Despite bias being a massive problem in pain management, LLMs potential use in this context raises questions. For instance, how might clinicians address a potential discrepancy between the output of the LLM (the patient's pain score and treatment recommendations) and the patient's testimony? The clinical and ethical concerns are, firstly, that a strong desire for mechanical objectivity in pain management may lead to automation bias, which is the tendency to over-rely on and be unduly confident in algorithmic outputs19; secondly, the clinician might be convinced that any relevant biases have been addressed in the model21; thirdly, any potentially ineffective LLM-generated treatment recommendations might impede the use of more effective options and might also render the patient's clinical status opaque21; fourthly, most patients are not necessarily in a position to verify or challenge the LLM output (or most medical tests) because they are dependent on their clinicians' (and the LLMs') credibility and reliability for their care14; and finally, the clinician may feel compelled to centre the LLM-generated pain score as opposed to the subjective testimony of the patient.11,24 Of course, patient testimony is not the only factor clinicians consider when making patient-centred and clinically appropriate treatment recommendations. Clinicians also weigh factors such as patient medical history, biopsychosocial influences, and patient values and treatment preferences.13 However, clinicians may consider the presumably objective output from the LLM more accurate and reliable, thereby minimizing the patient's credibility and knowledge about their subjective and embodied experiences.2,11,14,24 These issues may be intensified for patients with intersecting disadvantaged identities related to gender, class, sexuality, disability, substance use, or race.3,23,24 Young et al. attempted to create algorithmic fairness by removing demographic factors such as race, ethnicity, and sex from their model. Scholars have cautioned against this approach arguing that it falsely assumes that bias is quantifiable and addressable through algorithmic models. Human biases are often embedded deeply in social and structural contexts, making it difficult to eliminate them solely through technical means.21,26 Furthermore, removing demographic variables in models had limited success previously. For example, some AI tools can predict race or gender incidentally without these features being labeled in the training data but through proxies such as profession or social group associations.8,31 Others have argued that we should take a critical eye to variables such as sex and race in research. For example, biological sex is often conflated with the social construct of gender15 and that labeling race—another complex social construct—as an independent data point obfuscates the harms of structural and interpersonal racism,1 let alone intersectional factors shaping pain experiences and clinical decision-making.20 Addressing systemic biases includes acknowledging and addressing the broader socio-structural factors that contribute to population health inequities in pain management.26,28 The large body of evidence on the social determinants of health demonstrates that antidotes to population health inequities, including pain, exist outside of formal personal healthcare service provision.10,16,17 Using a LLM to solve structural macro-level injustices such as racial, gender-based, and sex-based stratification in pain management at the clinical level should be approached very carefully; it shifts the focus from upstream structural drivers such as public policies, laws, and distribution of health-related risks toward downstream micro-level or individual-level technological solutions to the root causes of population-level inequities.17 I am hopeful that research using AI systems, including LLMs, can make important contributions to the “equity, effectiveness, and efficiency” of pain management.4,18,30 If such research demonstrates evidence of actual patient benefit, it would be invaluable for individual patients as part of comprehensive multimodal pain management. However, the complex structural injustices that Young et al. aim to use LLMs to overcome—racism, sexism, and genderism in pain management—cannot be reduced to computations. A solely technological, algorithmic approach is insufficient to address these urgently important social issues. Conflict of interest statement The author has no conflicts of interest to declare.

Récupéré en direct depuis OpenAlex et désinversé. Les résumés ne sont pas conservés dans cette base de données : les index inversés représentent 8,6 Go des 9,3 Go de texte de la base, et le serveur dispose de 13 Go libres.