# Smart Lies and Sharp Eyes: Pragmatic Artificial Intelligence for Cancer Pathology: Promise, Pitfalls, and Access Pathways

## TL;DR

This paper discusses how AI can help in cancer pathology by improving diagnoses and identifying biomarkers, but also highlights risks like automation bias and how to safely implement AI in both high and low-resource settings.

## Contribution

The paper provides a pragmatic framework for deploying AI in cancer pathology, emphasizing validation, equity, and risk mitigation in diverse settings.

## Key findings

- AI can reliably detect lesions, quantify biomarkers, and assist in triage in cancer pathology.
- Common AI failure modes include data leakage, stain drift, and automation bias, which can undermine trust and accuracy.
- Practical adoption pathways for low- and middle-income countries include prioritizing high-impact use cases and shared validation resources.

## Abstract

Artificial intelligence (AI) is rapidly entering cancer pathology, promising faster, more reproducible diagnoses and new biomarkers that are invisible to the naked eye. Yet the same systems can mislead through shortcut learning, batch and stain drift, data leakage, or over-confident user interfaces that nudge clinicians into “automation bias.” Drawing on our International Conference on Contemporary Oncology 2025 presentation and a targeted review, we summarize what AI can reliably do today (detection, quantification, triage, and selected biomarker predictions), where it fails, and how to buy and validate tools safely. We also focus on practical pathways for low- and middle-income countries: choosing high-value use cases, minimizing compute and storage needs, building governance and training, and sharing validation assets across sites. Our goal is pragmatic: help pathology teams deploy AI that reduces variance and frees expert time, while guarding against the “smart lies” that can erode trust, equity, and patient outcomes.

Background: Whole-slide imaging and algorithmic advances have moved computational pathology from research to routine consideration. Despite notable successes, real-world deployment remains limited by generalization, validation gaps, and human-factor risks, which can be amplified in resource-constrained settings. Content/Scope: This narrative review and implementation perspective summarizes clinically proximate AI capabilities in cancer pathology, including lesion detection, metastasis triage, mitosis counting, immunomarker quantification, and prediction of selected molecular alterations from routine histology. We also summarize recurring failure modes, dataset leakage, stain/batch/site shifts, misleading explanation overlays, calibration errors, and automation bias, and distinguish applications supported by external retrospective validation, prospective reader-assistance or real-world studies, and regulatory-cleared use. We translate these evidence patterns into a practical checklist covering dataset design, external and temporal validation, robustness testing, calibration and uncertainty handling, explainability sanity checks, and workflow-safety design. Equity Focus: We propose a stepwise adoption pathway for low- and middle-income countries: prioritize narrow, high-impact use cases; match compute and storage requirements to local infrastructure; standardize pre-analytics; pool validation cohorts; and embed quality management, privacy protections, and audit trails. Conclusions: AI can already serve as a reliable second reader for selected tasks, reducing variance and freeing expert time. Safe, equitable deployment requires disciplined validation, calibrated uncertainty, and guardrails against human-factor failure. With pragmatic scoping and shared infrastructure, pathology programs can realize benefits while preserving trust and accountability.