Near-Malpractice? Reid Hoffman Calls for Routine AI Second Opinions in Medicine
When Reid Hoffman, cofounder of LinkedIn and now a prominent investor and builder in AI-driven drug discovery, suggested that clinicians should routinely consult chatbots—calling it “near malpractice” not to seek an AI second opinion—he didn’t just toss a provocative sound bite into the void. He put a mirror up to modern medicine and to a technology community that has been racing from prototype to deployment with dizzying speed.
This is not a debate about whether AI can be useful in medicine; it’s a debate about what constitutes responsible clinical judgment, how trust is calibrated between humans and algorithms, and how an entire industry must adapt legal, regulatory, and operational norms when automated reasoning becomes both ubiquitous and fallible.
The framings: augmentation, not replacement
Hoffman’s formulation—treating AI as a second opinion—carries a deliberate framing. It is not a claim that a language model should displace clinician judgment. Rather, it asks clinicians to reconfigure workflows so that algorithmic perspectives are part of routine decision-making, like laboratory results or radiology reads. In this view, AI is another data stream: sometimes clarifying, sometimes contradictory, sometimes wrong.
Seen through that lens, calling omission to consult AI “near malpractice” is rhetorical urgency. Hoffman is staking a position that the marginal value of an algorithmic check, in certain high-stakes cases, has risen to a level where failing to use it could be negligent—especially when the tools are widely accessible and able to surface clinically relevant differentials, drug interactions, or up-to-date literature faster than a single human can.
Why this rattles the status quo
Medicine has long rested on a social compact: clinicians synthesize training, experience, and evidence to recommend care, and patients trust that judgement. Integrating AI as a routine, recommended step changes that compact in several ways.
- Authority and accountability. If a clinician defers to an AI recommendation that later causes harm, who is accountable? Conversely, if a clinician ignores an AI suggestion that would have prevented harm, how are standards of care updated?
- Workflow redesign. EHR fatigue is real. To make AI second opinions practical, interfaces, alerting thresholds, and information provenance must be thoughtfully integrated or clinicians will ignore yet another noisy notification.
- Trust and transparency. Clinicians and patients alike will demand to know why an AI suggested a particular path—yet many current models are inscrutable and prone to confident errors.
- Equity and bias. Models trained on skewed datasets may perform poorly in underrepresented populations, risking widening disparities if their guidance is treated as authoritative.
The promise: speed, scale, and rare-disease aid
Despite those challenges, the upside is compelling. AI can rapidly synthesize millions of papers, detect subtle associations across diverse datasets, and propose differential diagnoses for rare presentations that a single clinician might never encounter in a lifetime. For drug discovery, where Hoffman spends significant energy, algorithmic triage can surface plausible therapeutic pathways and accelerate translational research—work that will cascade back into clinical decision tools.
Imagine an emergency department where an AI flags a rare but treatable cause for a patient’s altered mental status within minutes, or a primary care clinic where an AI double-check identifies a medication interaction missed in a rushed visit. Those scenarios are why many are excited—this technology can augment vigilance and reduce diagnostic delay.
Real harms: hallucinations, data gaps, and overconfidence
But the technology is not magic. Language models can hallucinate, presenting fabricated but plausible-sounding claims. Clinical datasets are often noisy, biased, and siloed. Many deployed systems show robust performance on curated benchmarks but degrade in real-world settings. If clinicians are to rely on AI as a routine second opinion, the community must grapple with the reality that models can be confidently wrong.
There is also a psychological hazard: automation complacency. If a clinician grows used to deferring to a model, small degradation in model accuracy can have outsized clinical consequences. Conversely, alarm fatigue can make clinicians ignore valuable warnings. The human-machine interface, therefore, must be designed to preserve active, critical engagement.
Legal and regulatory tectonics
Hoffman’s “near malpractice” formulation lands squarely in the medicolegal domain. Courts, regulators, and hospitals will have to decide whether and when the use of AI is a standard of care. That hinges on evidence: randomized trials, post-deployment monitoring, and transparent reporting of performance across populations.
Regulators will need to consider new paradigms for software that learns after deployment. A static cleared device is different from a model that updates daily in response to new data streams. Accountability mechanisms—versioning, audit trails, and adverse event reporting—will become central to lawful, ethical use.
Practical guardrails for routine use
Accepting the premise that AI second opinions could be valuable, what would responsible implementation look like?
- Provenance and explainability: Every recommendation should include its rationale, data sources, and confidence estimates. Clinicians need a clear trail of why an AI offered a particular view.
- Validation and monitoring: Systems must be tested in the real world, with continuous performance monitoring and mechanisms to halt or revert updates that degrade outcomes.
- Human-in-the-loop governance: Tools should empower clinicians to accept, modify, or reject suggestions and to submit feedback that improves models.
- Interoperability and workflow fit: AI should appear where clinicians already work, with minimal disruption and tailored alerts to avoid cognitive overload.
- Equity audits: Regular assessment of performance across demographic groups to detect and correct biases before harm cascades.
- Patient transparency: Patients deserve to know when AI contributed to a recommendation and how their data were used, with options for consent where appropriate.
The cultural shift: humility and learning at scale
Perhaps the most profound change is cultural. Medicine has always balanced humility and confidence. AI introduces an external check at scale. That can be humbling: clinicians will be challenged by improbable differentials and by model suggestions that contradict experience. But it can also be liberating: decision fatigue could be eased, rare conditions could be recognized earlier, and continuous learning could be accelerated across health systems.
For the AI community, Hoffman’s statement is a call to build tools that earn clinician trust by demonstrating reliability, transparency, and measurably better outcomes. For health systems, it’s an imperative to rethink governance structures so that beneficial technology is deployed responsibly. And for society, it’s an invitation to update ethical and legal frameworks that currently assume a binary division between human judgement and automated tools.
A balanced path forward
Hoffman’s language—provocative by design—serves a useful purpose. It forces stakeholders to confront what responsible adoption looks like, rather than slipping into sensational optimism or reactionary fear. The practical path forward is neither uncritical embrace nor reflexive rejection. It is a careful, evidence-driven integration that recognizes AI as a partner in care, not a panacea.
Ultimately, whether AI second opinions become standard practice will depend on rigorous validation, thoughtful design, and adaptive regulation. The stakes are high: lives, trust, and equity hang in the balance. If Hoffman succeeds in making the conversation urgent, the AI news community must keep it nuanced, skeptical where warranted, and relentlessly focused on outcomes.
We stand at a crossroads where technology can amplify human care or magnify its flaws. The choice is not binary. It is procedural: build systems that respect clinical judgment, surface clear rationales, and improve measurable outcomes. Do that, and the second opinion—algorithmic and human—becomes a powerful safeguard. Fail to do it, and we risk enshrining new forms of error into routine practice.
Reid Hoffman’s challenge is uncomfortable—but that is precisely the point. The debate it provokes may well be the catalyst medicine needs to build AI that is safe, equitable, and genuinely augmentative. For a community that tracks every architectural advance and every policy tremor, this is the moment to move from headlines to hard questions, from hype to clinical evidence, and from rhetorical brinkmanship to durable standards.
Medicine has always adapted to new technologies. The question now is whether adaptation will be deliberate and accountable—or hasty and unforgiving. Hoffman’s words are a prod: the conversation must begin, and it must move quickly, thoughtfully, and with the courage to admit uncertainty while demanding proof.
