Beware of temptations to commit 1 (or more) of the 7 deadly sins common to healthcare AI

Researchers are warning of potential perdition—so to speak—for doing medical AI wrong in any of seven “deadly” (and easily avoidable) ways. 

The team’s tongue-in-cheek wordplay has a serious goal in its soul: moving the discussion “beyond scattered ethical guidelines toward a unified diagnostic tool for trustworthy, human-centered medical AI.”

Led by Andreas Holzinger, PhD, MPH, of the Human-Centered AI Lab at the University of Natural Resources and Life Sciences in Vienna, Austria, the study authors arrived at their framework by systematically synthesizing scientific literature, clinical guidelines and regulatory rules. 

They validated the resulting guidance by polling 914 stakeholders across 143 countries between July 2024 and March 2025. 

“Results confirmed broad agreement with each pre-identified risk,” Holzinger and colleagues report, “revealing cross-cultural convergence in ethical concern alongside persistent divides in attitudes toward regulation—particularly between technologically advanced nations and emerging economies.” 

The paper, which is posted ahead of final editing for publication in NPJ Digital Medicine, doesn’t leave its guilty readers to suffer in their fallen state. 

Instead, it proposes seven sin-busting virtues—one for each of the seven deadlies—that, together, offer “actionable principles to guide responsible development and governance” of AI for healthcare. 

Here’s a summary of each avoidable impropriety.  

Sin 1: Blind trust (‘See no evil, hear no evil’). 

The name of this sin captures the dangers of overreliance on AI systems without proper validation, context-awareness or clinical oversight, the authors explain. 

“The allure of automated authority can foster uncritical acceptance even when results are opaque or non-reproducible,” they write. “In high-stakes environments like oncology or emergency care, this can result in misdiagnosis, over-treatment or delays in critical decision-making.”

Sin 2: Overregulation (‘No guts, no glory’). 

Calling out this transgression recognizes that too much government oversight—just like too little—can do harm as well as good. 

“Overly cautious frameworks may hinder innovation, restrict clinician autonomy or delay potentially life-saving tools,” the authors point out, “particularly when regulation fails to differentiate between high-risk diagnostic AI and low-risk administrative tools.”

Sin 3: Dehumanization.

Poorly integrated AI systems can detract from the relational and empathetic aspects of care, resulting in this highly undesirable effect. 

“Automated triage bots, scripted diagnostic interfaces or emotionless decision-support tools risk reducing patients to data points and clinicians to passive intermediaries,” the authors state. This can be especially detrimental to shared decision-making efforts, shifting the doctor-patient relationship “from a dyad to a triad.”

Sin 4: Misaligned optimization.

Tension is probably inevitable between AI’s “tendency to optimize for single objectives” and medicine’s “requirement for simultaneous optimization across technical, economical and moral dimensions,” the authors suggest. 

“When proxies for success are poorly chosen,” they emphasize, “systems can perform exactly as designed and still fail their users.”

Sin 5: Overinforming and false forecasting.

Sometimes the promise of enlightening AI insights becomes a weighty burden, the authors warn. 

The main risk here is “overwhelming clinicians with excessive data, irrelevant alerts or speculative forecasts” and/or “creating a false sense of precision where clinical uncertainty remains irreducible.”

Sin 6: Misapplied statistics.

This iniquity can rear its ugly head when “population-level statistical predictions are applied indiscriminately to individual cases without contextual adjustment,” the authors caution. End-users might recognize the error of their ways in instances of “mechanical generalization and clinical misjudgment.” 

The risk is further elevated when “systems are deployed across diverse populations without accounting for demographic variability or multimorbid complexity.”

Sin 7: Self-referential evaluation.

Here the authors are talking about practice of assessing AI systems “solely on internal metrics or simulations,” the authors clarify. The lack of external auditing, human feedback or real-world testing allows the forming of a “self-referential loop that can mask bias, error propagation and unexpected failure modes.” 

“Due to patient and disease drifts, this cannot be a one-time validation; systematic and perpetual adjustment is required.”

 

Sin-mitigating virtues to the rescue 

The authors state they intend the seven sins framework not only as a diagnostic instrument but also as a springboard for positive action. 

Each sin, they maintain, “implies a corresponding cardinal virtue for AI in medicine—actionable principles that, when cultivated, promote safe and trustworthy outcomes.” 

Their seven healthcare AI virtues are:  

1. Critical validation; 
    
2. Appropriate, proportionate regulation; 
    
3. Human-centered design supporting shared decision-making;
    
4. Holistic, multi-objective optimization;
    
5. Transparent communication and explainable models;
    
6. Statistical rigor and individualization; and
    
7. Independent, perpetual evaluation. 

“This positive framing is our primary message: Responsible AI in medicine is the active cultivation of these virtues by clinicians, developers and policymakers,” Holzinger et al. write.

The study’s U.S.-based co-authors include Vimla Patel, PhD, and Edward Shortliffe, MD, PhD, both affiliated with Columbia University in New York City. 

The pre-published paper is available in full for free (PDF).