The phrase hangs in the air of the executive conference room—casual, optimistic, forward-looking. "Let's try AI and see what happens." Around the table, heads nod. The vendor's demonstration was impressive. The technology promises efficiency gains, better diagnoses, streamlined workflows. The decision feels like progress, like innovation, like keeping pace with the future of healthcare.

What the moment lacks is the weight it deserves.

Short on time? Read the TLDR version.

In a hospital, "let's try it and see" is a phrase reserved for low-stakes experiments. New coffee in the cafeteria. Adjusted visiting hours on one unit. Small pilots where failure means inconvenience, not harm. It is not—it cannot be—the framework for deploying tools that shape diagnoses, influence triage decisions, and guide treatment in real time.

Yet artificial intelligence, especially the large language models now entering clinical and operational workflows with remarkable speed, is being adopted precisely this way. With limited real-world validation. With unclear accountability. With insufficient monitoring of how these systems perform when confronted with the messy, high-pressure reality of patient care.

Unlike the drugs and devices that move through rigorous trials and regulatory pathways before touching patients, many AI tools slip into practice through side doors. Productivity pilots. Innovation initiatives. Technology refresh cycles. The safeguards that healthcare has built over decades—the demand for evidence, the insistence on transparency, the requirement that someone be accountable when something goes wrong—somehow do not apply with the same force.

This creates risks that are predictable, preventable, and increasingly consequential. Patient safety risks when outputs sound authoritative but contain fabrications. Equity risks when models trained on incomplete data systematically fail certain populations. Accountability risks when no one can explain how a recommendation was generated or who bears responsibility when it proves wrong. Trust risks when clinicians and patients discover that the impressive demonstrations did not translate into reliable performance in their specific context.

Treating AI like a casual experiment in a clinical environment is dangerous. Without evidence, governance, bias mitigation, transparency, and continuous monitoring, hospitals risk preventable harm, widened disparities, and loss of the trust that makes healthcare possible.

The Evidence Gap: Why Patient Safety Cannot Accept "Let's See What Happens"

The appeal of large language models lies partly in their apparent intelligence. They generate responses that sound knowledgeable, structured, complete. They adopt the tone of medical literature. They present differential diagnoses in numbered lists. They summarize complex cases with apparent confidence.

This fluency creates danger precisely because it increases perceived credibility while masking uncertainty. In fast-paced clinical settings—emergency departments at three in the morning, intensive care units managing multiple crashing patients, primary care providers racing through packed schedules—the ability to detect subtle inaccuracies collapses. An incomplete drug interaction warning. A fabricated citation to a non-existent study. A contraindication buried in confident-sounding prose that fails to mention it. These errors prove harder to catch when the output appears authoritative and time is short.

Clinical environments amplify rather than absorb such risks. Small errors cascade. A delayed diagnosis becomes delayed treatment. An incorrect medication order propagates through documentation systems. A missed escalation criterion leaves a deteriorating patient in an inappropriate care setting. The stakes are not abstract.

Traditional medical interventions—new drugs, new devices, new procedures—typically move through defined pathways before widespread adoption. Clinical trials establish safety and efficacy profiles. Regulatory reviews assess evidence. Implementation follows protocols designed to catch problems early and limit exposure until performance is proven. The process is slow, expensive, and often frustrating to innovators eager to help patients. It exists because medicine learned, through painful experience, that good intentions and impressive mechanisms do not guarantee safe outcomes.

Many AI tools bypass this discipline. Hospitals inherit vendor claims about accuracy without confirming local performance. Retrospective testing on clean datasets substitutes for prospective validation in messy real-world workflows. Lab benchmarks replace clinical outcome measurement. The pressure to innovate, to appear modern, to solve operational problems with technological solutions, pushes organizations toward deployment before evidence supports it.

What should evidence look like for clinical AI? It begins with clarity about intended use. Not vague promises of "productivity" or "efficiency," but specific answers: What decision does this tool support? For which patients? In what clinical setting? What must it not be used for? These constraints are not limitations but safeguards—they define the boundaries within which performance has been validated and outside which harm becomes likely.

Clinical validation must happen in real workflows, not controlled environments. The model that performs well on historical data may fail when integrated into active documentation systems, when confronted with incomplete information during triage, when used by night shift providers with different training backgrounds. Performance must be benchmarked against current standard of care and measured through downstream outcomes. Does diagnostic accuracy actually improve? Does time-to-treatment decrease? Do adverse events decline? Or do the metrics that matter to patients and clinicians remain unchanged while new administrative burdens accumulate?

Before any patient-impacting use, hospitals need actionable checkpoints. Define the clinical problem clearly—no vague innovation mandates. Pre-specify success metrics and safety thresholds. Document known limitations and failure modes with the same rigor applied to drug side effects. Require independent review and approval from clinical safety leadership, not just IT or innovation teams. Most critically, define the plan for stopping or rolling back if harm signals emerge. Without this last element, the "pilot" becomes permanent by inertia rather than evidence.

Evidence is necessary. But organizational culture and incentives often conspire to bypass it.

The Solutionism Trap: How Hype Accelerates Unsafe Adoption

Technological solutionism takes many forms in healthcare, but its core belief remains consistent: complex organizational problems can be solved with the right technology. AI becomes the "magic bullet" that will bypass staffing shortages, fix broken workflows, eliminate operational inefficiencies, and somehow accomplish what years of process improvement, training, and resource allocation have not.

This belief transforms AI adoption from targeted intervention into innovation mandate. The question shifts from "Is AI the right tool for this specific problem?" to "How quickly can we implement AI?" Organizations frame deployment as strategic imperative rather than clinical decision, and strategy rarely demands the same evidence standards as clinical practice.

The assumption that implementation is primarily technical proves particularly corrosive. Purchase the software. Integrate with existing systems. Train users on the interface. Go live. This framing ignores that the hardest parts of healthcare AI adoption are not technical but human: clinical workflow redesign, accountability clarification, change management, trust building. When organizations treat these as afterthoughts rather than prerequisites, pilots launch in live clinical workflows without clear escalation paths, without adequate training, without monitoring systems to detect problems before they become patterns.

Leadership pressure to "innovate" creates shortcuts. Governance reviews get compressed. Validation studies get deferred to "phase two." Training becomes optional webinars rather than competency requirements. Monitoring gets planned for "after we see how it goes." These shortcuts feel pragmatic in the moment—faster time to value, reduced friction, fewer barriers to adoption. They prove dangerous when the tool reaches the bedside.

Teams confuse early demonstrations with evidence. The vendor's benchmark results become proof of capability. The positive pilot feedback from three enthusiastic early adopters becomes validation of readiness for enterprise deployment. The absence of reported incidents during the first month becomes confirmation of safety. None of these constitute the evidence that patient safety demands, but organizational eagerness to believe can reframe them as sufficient.

What AI cannot fix becomes visible only after deployment. Poor data quality remains poor—now automated at scale. Broken workflows accelerate—now with algorithmic assistance that creates new handoff failures. Understaffing persists—now with additional review burdens as clinicians verify AI outputs. Inequitable access to care continues—now potentially amplified by models trained on data from better-resourced populations. Unclear clinical accountability becomes more unclear—now diffused across human judgment and algorithmic recommendation with no clear line of responsibility.

AI can shift work rather than reduce it. Someone must review the AI-generated summaries for accuracy. Someone must investigate when outputs seem questionable. Someone must handle the incidents that emerge from inappropriate reliance or misunderstood limitations. Someone must manage the vendor relationship, monitor performance metrics, coordinate updates, and maintain the governance oversight that prevents gradual drift from safe practice. These someones require time, training, and organizational support that the business case often fails to account for.

A more rigorous approach starts with root cause analysis rather than solution selection. Map the clinical and operational workflow end-to-end. Identify the specific bottlenecks, failure modes, and accountability gaps that create the problem you want to solve. Ask whether simpler interventions would work: process redesign that eliminates unnecessary steps, decision support rules that codify established protocols, staffing changes that address capacity constraints, training that builds capabilities currently lacking.

Use AI only when it genuinely is the best-fit tool for the problem—when the decision requires predictive capability beyond what rules can capture, when the information synthesis demands exceed human processing capacity under time constraints, when personalization at scale would otherwise be impossible. And use it only when the organization is prepared to govern and monitor it with the rigor that clinical safety requires.

Even well-scoped, properly governed AI can cause harm if it performs unevenly across the populations it serves.

The Bias Reality: Systematic Harm to Vulnerable Populations

Training data creates blind spots with mechanical precision. Models learn patterns from the examples they see. When those examples underrepresent patients by race, ethnicity, sex, age, language, disability status, or socioeconomic circumstance, the resulting models systematically underperform for the missing populations.

This is not an edge case. It is a predictable consequence of how machine learning works.

Consider skin cancer detection. Models trained predominantly on images of lighter skin tones learn to identify malignancies as they appear on that skin. When those same models encounter lesions on darker skin, where presentation differs, they fail more often. Miss the diagnosis. Delay the referral. Allow the cancer to progress while the algorithm's confidence score suggests everything is fine. The performance gap may appear modest in aggregate statistics—a few percentage points of reduced accuracy. But aggregated across thousands of clinical encounters, that gap becomes systematic disparity: some patients get timely diagnosis and treatment, others do not, and the difference tracks along racial lines.

Adult-trained models deployed in pediatric contexts carry similar risks. Physiologic differences between children and adults—in metabolism, in normal vital sign ranges, in disease presentation—mean that algorithms validated on adult populations may generate unsafe recommendations when applied to children. Without population-specific validation, the hospital cannot know whether the tool that works well in adult medicine will help or harm in the pediatric ward.

Local population characteristics matter profoundly. "Validated elsewhere" does not guarantee safe performance at your institution. Patient demographics differ. Disease prevalence varies. Documentation practices evolve differently across sites. Clinical protocols reflect local expertise and resource availability. A model trained at academic medical centers may fail in community hospitals. A model validated in one geographic region may underperform in another where the patient population presents with different characteristics.

These failures constitute more than statistical artifacts. They are ethical failures—violations of fundamental fairness and equal treatment obligations. They are legal exposures—potential discrimination claims when outcomes systematically disadvantage protected groups. They are operational risks—damage to institutional reputation, loss of community trust, erosion of clinician confidence in systems that don't work for all their patients equally.

They are, most fundamentally, patient safety failures. When bias causes models to systematically miss diagnoses in certain populations, to under-predict risk for some patient groups, to recommend inappropriate treatments based on incomplete training data, people are harmed. The harm is not random. It is patterned. It falls disproportionately on populations already facing healthcare disparities. And it persists silently unless organizations specifically look for it.

Mitigation begins with transparency demands. Require vendors and internal development teams to report performance by relevant subgroups—not just overall accuracy, but accuracy stratified by race, ethnicity, age, sex, language, and other characteristics that might predict differential performance. Define acceptable variance thresholds: how much performance difference is tolerable before the model becomes unsafe for deployment?

Build representative data strategies that address underrepresentation proactively. This means more than simply collecting more data—it requires intentional effort to include adequate representation of minority and underserved populations, to validate that the data quality is comparable across groups, to confirm that the clinical concepts and labels mean the same thing across different contexts.

Conduct bias audits before deployment and continuously after. Test whether recommendations differ systematically across patient groups in ways that cannot be justified by actual clinical differences. Involve patients and community stakeholders from affected populations in the design and validation process. Their perspectives often surface issues that technical teams miss.

Define clear escalation and rollback plans. What signals would indicate that the model is creating disparate outcomes? Who monitors for those signals? How quickly can the tool be withdrawn if disparity appears? These questions must be answered before deployment, not discovered through incident reports after harm has occurred.

Even unbiased models remain unsafe if no one can explain, audit, or own the clinical decisions they influence.

The Accountability Void: When No One Owns the Outcome

The transparency gap in "let's try AI" deployments begins with fundamental unknowns. Teams implementing the tool often cannot articulate how outputs are generated. What data was used? What assumptions are embedded in the model? What statistical thresholds trigger different recommendations? When is the model likely to fail? These questions go unanswered not because the information is deliberately hidden, but because the questions were never asked before go-live.

Without this understanding, clinicians cannot appropriately calibrate trust. They cannot distinguish between contexts where the tool is reliable and contexts where it should not be used. They cannot identify when an output falls outside the model's validated scope. The tool becomes a black box offering recommendations that must either be accepted on faith or ignored entirely—neither of which serves patient care well.

Operationally, lack of transparency undermines incident investigation and quality improvement. When something goes wrong—a missed diagnosis, an inappropriate treatment, a delayed escalation—the investigation cannot trace the reasoning that led to the error. Was it a model failure? A misunderstanding of the output? An integration problem? Inappropriate use outside validated contexts? Without traceability, organizations cannot learn from failures and therefore cannot prevent recurrence.

Responsibility and liability must be defined before deployment, not discovered through litigation after harm. Who owns model performance? Not the vendor alone—hospitals deploying the tool bear responsibility for validating its performance in their environment and monitoring it over time. Who signs off on clinical use? Not innovation teams or IT departments, but clinical leadership with accountability for patient safety. Who handles adverse events related to AI outputs? Not a diffuse network of people who assume someone else is responsible, but named individuals with clear authority and defined processes.

Most critically, how can patients and clinicians challenge or appeal AI-influenced decisions? The model's recommendation cannot be the final word when it affects patient care. There must be pathways for human review, for escalation when outputs seem wrong, for documentation of disagreement when clinical judgment diverges from algorithmic suggestion. Without these pathways, accountability disappears into "the model said so"—a diffusion of responsibility that leaves everyone and no one answerable.

Governance essentials emerge from these needs. Establish a documented oversight pathway—a clinical safety committee or AI governance board with defined membership, decision rights, and meeting cadence. This body reviews proposed use cases before deployment, approves changes to validated tools, investigates incidents, and maintains authority to withdraw tools that prove unsafe.

Implement change control for model updates. When the vendor releases a new version, when retraining occurs, when integration points change, these modifications require review and approval before reaching production systems. The clinical validation does not automatically transfer to updated models.

Create incident reporting processes specific to AI. Staff need clear channels to report when outputs seem wrong, when workflow integration creates safety risks, when training proves inadequate, when monitoring reveals concerning patterns. These reports must reach decision-makers with authority to act.

Define clinician-in-the-loop requirements based on risk. For high-stakes decisions—triage in emergency departments, treatment recommendations in critical care, diagnostic conclusions in pediatrics—require explicit human review and documentation before action. The model can inform but cannot decide alone.

Document everything with the rigor that medical-legal review will eventually demand. Create model cards that function like drug labels: intended use clearly stated, contraindications listed, limitations acknowledged, known failure modes documented, required human review steps specified. Maintain audit logs that capture who used the tool, what inputs were provided, what outputs were generated, what action was taken. Implement versioning that enables investigation to determine exactly which model version was active when a particular decision was made.

Establish retention and review processes that support clinical governance, regulatory compliance, and root-cause analysis. The documentation created during deployment is not administrative overhead—it is the evidence trail that enables accountability when questions arise.

Even with governance and accountability structures in place, human overreliance on AI remains a persistent danger.

The Overreliance Trap: When Fluency Masks Fallibility

Large language models present unique risks in clinical settings because they generate text that sounds more authoritative than their actual knowledge justifies. The output arrives in complete sentences, organized paragraphs, professional tone. It cites studies—sometimes real, sometimes fabricated. It presents differential diagnoses with apparent logical structure. It summarizes complex cases with confident assertions about what matters most.

This fluency is dangerous. It persuades through form rather than substance. Users see polished prose and infer careful reasoning. They encounter consistent formatting across queries and assume systematic reliability. They read confident conclusions and trust that uncertainty has been properly accounted for.

The reality is different. Models hallucinate facts without obvious signals. They fabricate citations to studies that do not exist, presenting them with correct formatting and plausible titles. They omit critical contraindications while including lesser considerations, creating summaries that look comprehensive but lack essential safety information. They sound consistent across repeated queries even when giving contradictory answers, because consistency of tone differs from consistency of content.

Clinicians face particular vulnerability to these failures under time pressure. In the emergency department managing multiple acute patients, the temptation to accept an AI-generated summary without verification becomes strong. During overnight shifts with reduced staffing, the confident-sounding recommendation from the model may go unchallenged. In primary care with back-to-back appointments running behind schedule, the apparently thorough differential diagnosis may be trusted without confirmatory testing.

Common overreliance scenarios hospitals must anticipate include clinicians deferring judgment to AI suggestions when outputs appear confident, even in contexts where clinical examination or additional testing would normally be required. Staff accepting AI-generated summaries or treatment recommendations without verifying the underlying source data, leading to propagation of errors through documentation systems and care transitions. Patients treating AI-generated health information as clinical guidance without appropriate context, potentially delaying care or making inappropriate self-management decisions.

These scenarios are not hypothetical. They represent predictable failure modes that emerge when fluent but fallible systems meet high-pressure clinical environments.

Safeguards must reduce automation bias—the tendency to favor algorithmic recommendations over contradictory human judgment. Force uncertainty displays that communicate when models are confident versus uncertain, when recommendations fall near decision boundaries, when inputs fall outside training distributions. These signals help users calibrate appropriate skepticism.

Require citation and grounding where possible. Link recommendations to specific source data within the patient's record. Avoid unsupported claims that cannot be traced to documented information. When models generate text without clear grounding, flag it explicitly as synthesized content requiring independent verification.

Mandate double-check workflows for high-risk decisions. Medication orders, diagnostic conclusions, treatment plans, discharge decisions—these require human review against source documentation, not just acceptance of AI-generated summaries. Set clear "do not use" rules for contexts where the risk of error is unacceptable: critical triage decisions, pediatric dosing calculations, end-of-life care recommendations.

Training must address safe use, not just technical operation. Teach staff what the model can and cannot do. Explain how to recognize hallucinations—watching for unsourced claims, checking citations, comparing outputs to known facts. Demonstrate common failure modes specific to the tool being deployed. Show how outputs degrade when inputs fall outside validated parameters.

Standardize documentation of AI involvement in clinical decision-making. When a model influences diagnosis, treatment, or triage, that influence must be recorded in ways that support accountability and learning. Not "AI recommended X," but documentation that captures what information the model processed, what recommendation it generated, what clinical reasoning led to acceptance or rejection of that recommendation, and what outcome resulted.

Use scenario-based training that mirrors real workflow pressures. Simulations of handoffs where AI summaries might be trusted without verification. Emergency department surge situations where time pressure increases overreliance risk. Night shift coverage scenarios where reduced supervision makes independent verification more important. Training in controlled environments prepares staff for the pressures they will face in practice.

Safeguards and training only work when paired with structured deployment and continuous monitoring. Hospitals need repeatable operating models for AI, not one-off experiments.

The Deployment Framework: From Pilots to Safe Production

Readiness components before any patient-impacting AI goes live include technical validation proving the model performs as claimed, privacy and regulatory review confirming compliance with HIPAA and applicable FDA pathways, workflow integration planning that maps where outputs appear and how they influence decisions, cybersecurity assessment addressing new attack surfaces and data protection requirements, and data governance establishing ownership, permitted uses, and audit capabilities.

User training and competency checks must align to actual clinical touchpoints. Train for the specific context: how the tool appears in order entry screens, how outputs integrate into documentation workflows, how recommendations should influence triage decisions, how information flows through discharge planning. Generic training on "how AI works" proves insufficient when users encounter the system in their specific high-pressure contexts.

Clear operating procedures for downtime, escalation, and incident response prevent crisis improvisation. What happens when the system is unavailable? How do workflows revert to manual processes? Who decides when performance degradation requires tool withdrawal? Who investigates when outputs appear wrong? These procedures must exist before problems arise, not be invented during emergencies.

Context-specific validation is non-negotiable. Models must prove performance for the intended population and setting. Adult medicine validation does not transfer to pediatrics. Inpatient validation does not transfer to emergency departments. Academic medical center validation does not transfer to community hospitals. Each context brings different patient characteristics, different clinical practices, different documentation patterns, different workflow pressures.

Validate against local data distributions. The model that performs well on vendor test sets may fail when encountering your institution's specific mix of patient demographics, disease prevalence, and care patterns. Test with actual data from your environment, not just published benchmarks.

Validate against local documentation habits and clinical practice patterns. Models trained on complete, structured documentation may fail when encountering the abbreviated, free-text notes common in busy clinical practice. Models trained on one institution's protocols may generate inappropriate recommendations when deployed where different evidence-based guidelines are followed.

Test integration points where errors most commonly occur: handoffs between providers, medication reconciliation at transitions of care, high-acuity pathways where multiple decisions compress into short timeframes. These stress points reveal problems that never appear in controlled testing environments.

Framework-driven approaches like BRIDGE provide structure for defining gates and responsibilities. Establish safety and trust criteria upfront: What performance thresholds must be met? What equity standards must be satisfied? What clinician-in-the-loop requirements apply? These criteria enable objective go/no-go decisions rather than political negotiations during deployment.

Set validation gates that must be passed before advancing: initial feasibility assessment, workflow fit verification, technical performance validation, bias audit completion, training readiness confirmation, monitoring infrastructure deployment. Each gate has specific artifacts that must be produced and specific approvals that must be obtained.

Define monitoring metrics and reporting responsibilities before scaling. Who tracks model performance? How often? What thresholds trigger investigation? Who receives monitoring reports? What authority do they have to act on concerning findings? These operational details prevent the diffusion of responsibility that allows problems to persist unaddressed.

Align stakeholders across clinical safety, IT operations, compliance, risk management, quality improvement, and frontline users. Distribute ownership appropriately: clinical leadership owns validation standards and safety thresholds, IT owns technical integration and monitoring infrastructure, compliance owns regulatory adherence, risk management owns incident investigation, frontline users own workflow integration and training adequacy.

Post-deployment monitoring must treat AI like other safety-critical systems. Track accuracy and prediction performance against established baselines. Monitor error rates—both false positives and false negatives—because each creates different failure modes and patient impacts. Watch subgroup performance to detect disparities that emerge or worsen over time. Record near-misses and close calls that reveal problems before they cause harm. Track clinician override rates as signals of trust calibration and model appropriateness. Measure patient outcomes—the ultimate test of whether the tool improves care.

Create predefined thresholds that trigger action without requiring debate. If accuracy drops below a specific level, retraining is triggered. If disparities exceed acceptable variance, bias audit is initiated. If override rates suggest users don't trust the tool, workflow investigation begins. If patient outcomes don't improve despite model deployment, use case is reassessed. These triggers enable quick response when monitoring reveals problems.

Continuously iterate with governance oversight. Review monitoring data regularly. Learn from incidents and near-misses. Update documentation and training based on what real-world use reveals. Reassess whether the tool remains suitable for its intended use as clinical practices evolve, as patient populations shift, as workflows change.

The Standard Healthcare Needs

The evidence is clear. The risks are predictable. The safeguards are known.

Hospitals cannot treat AI as a casual experiment because patient safety requires evidence of performance in real clinical contexts, validation that extends beyond vendor claims to local populations and workflows, and clear criteria for when tools should be stopped or rolled back. Hype-driven solutionism encourages shortcuts—skipped governance, compressed validation, deferred monitoring—that AI cannot compensate for and that multiply risk when tools reach the bedside.

Bias and inequity are not edge cases but systematic risks that can predictably harm underserved populations when training data underrepresents them or when models validated in one context fail in another. Black-box decisions without clear accountability and documentation undermine the trust necessary for safe incident response and continuous improvement. Overreliance—especially with large language models whose fluency masks fallibility—can delay interventions and increase clinical errors as users trust confident-sounding outputs more than their accuracy justifies.

Safe adoption requires structured readiness, context-specific validation, governance with defined decision rights and responsibilities, and continuous monitoring that treats AI like the safety-critical system it is.

Replace "let's try AI" with a safety-first deployment standard. Define the clinical problem clearly and specifically. Require intended-use clarity that establishes boundaries for validated performance. Demand real-world validation in your institution's workflows and patient populations. Implement governance structures with documented oversight, change control, and incident response. Create documentation standards—model cards that function like drug labels, audit logs that enable investigation, versioning that supports accountability. Mandate equity reporting that tracks subgroup performance and defines acceptable variance. Train users not just for operation but for recognizing failure modes and appropriate reliance. Establish continuous monitoring with predefined thresholds that trigger retraining, investigation, or withdrawal.

Get a comprehensive AI safety readiness assessment to evaluate your organization's preparedness for safe AI deployment.

AI can improve care—better diagnoses, earlier interventions, reduced clinician burden, more personalized treatment. But only when hospitals treat it like any other high-risk clinical intervention: measured carefully, validated thoroughly, monitored continuously, and held accountable rigorously. The stakes are lives, trust, and the equity that should define healthcare. "Let's try it and see" is not a standard those stakes can accept.

The conference room conversation needs different words. Not "let's try AI," but "let's prove AI is safe." Not "see what happens," but "define what success requires and measure whether we achieve it." Not innovation as mandate but intervention as responsibility.

The difference between those framings is the difference between technology theater and patient safety. Healthcare has learned this lesson before, in other contexts, through painful experience. The question is whether that learning will be applied to AI before experience makes it painful once again.