Ethics in IT — Week 3 · Ch. 8
IT Impact on Society
"Technology creates and destroys jobs. Who decides which ones?"
13 slides ~16 minutes ETH-10 · The Factionless
By the end of this module, you will explain the productivity paradox, map AI-driven displacement risk, evaluate healthcare IT's twin promise and hazards, articulate the digital divide, and analyze the IBM Watson for Oncology failure.
Slide 2 of 13
The Productivity Paradox
Robert Solow, 1987: "You can see the computer age everywhere except in the productivity statistics."
The paradox (1970s-80s)
Massive IT investment. Negligible measurable productivity gains. The computer age was visible everywhere except in the data that was supposed to show it.
The resolution (late 1990s)
Productivity gains finally appeared — but only when paired with business process redesign, network effects, and supply chain optimization. IT alone does not create value. Organizational change does.
The ethical implication: organizations justify significant disruption — layoffs, restructuring, displacement — on the promise of productivity gains that may not materialize without the complementary organizational investment. The workers displaced carry the cost of a bet they did not make.
"The quarterly earnings report does not capture the career of a 47-year-old data entry clerk whose job was automated. The productivity gain is real. The cost is externalized."
Slide 3 of 13
Automation Risk Tiers
Earlier waves replaced manual labor. AI/ML targets cognitive work that previously felt automation-resistant.
"The engineer who builds an automated legal document review system is not responsible for law firm layoffs. But they are part of a chain of decisions. Pretending the chain does not exist is its own kind of choice."
Slide 4 of 13
What Tech Owes the Displaced
Active debate. The minimum positions are well-established; the stronger ones are contested.
Minimum positions
Adequate notice. Retraining investment. Transition support. Severance proportionate to tenure.
Where most large employers operate when the law requires it.
Middle positions
Industry-funded reskilling pools. Worker representation on automation decisions. Phased rollout that internalizes the human cost.
Some unions, some EU jurisdictions.
Stronger positions
Automation tax to fund universal reskilling. Job guarantee programs. UBI. Mandatory retention during transition periods.
Active policy debate, varying enactment.
The IT professional position: the engineer building the automation is not the policymaker. But they are the visible face of the technology to the displaced workforce. Refusing to engage with the consequence is a choice; the chain still exists.
Slide 5 of 13
Healthcare IT: Promise & Hazards
Real benefits. Real harms. The ethical complexity is in the distribution of both.
EHR Benefits
Reduced duplicate testing. Fewer medication errors. Better care coordination. Patient outcomes improve in mature EHR organizations.
EHR Risks
High-value ransomware targets. Breaches expose the most sensitive PII that exists. Interop failures fragment care.
Algorithmic Bias
ML diagnostics trained on non-representative data perform worse on underrepresented populations — including dark-skinned patients.
Data Monetization
De-identified patient data is commercially valuable. Patients rarely see how their data generates profit downstream.
"The same dataset that trains a life-saving diagnostic AI may also be sold to an insurance company to identify high-risk policyholders. The patient signed a consent form that did not describe either use."
Slide 6 of 13
Algorithmic Bias in Healthcare
A diagnostic AI that works for one population and fails for another isn't a technical glitch. It's the training data made visible.
The pattern
ML diagnostic systems trained primarily on light-skinned patient data have been documented to perform significantly worse on dark-skinned patients.
Same model. Same inputs. Different accuracy depending on the patient.
The mechanism
The training set is the model's view of the world. If the training set underrepresents a population, the model will be less accurate for that population.
The bias is not in the algorithm — it's in the data the algorithm was given.
The professional response: dataset audits before training. Performance metrics broken out by demographic group. Documentation of known limitations in the deployment package. "We didn't intend the bias" doesn't help the patient who got the wrong diagnosis.
Slide 7 of 13
The Digital Divide
When essential services move online, the divide stops being inconvenience and becomes structural exclusion.
Economic
Cost of devices, broadband, mobile data. Households below the poverty line have lower connectivity rates.
Geographic
Rural and tribal lands have lower broadband availability. Even when available, speeds and reliability lag urban areas.
Generational
Older adults face steeper learning curves and less institutional support for digital literacy. Compounds with limited mobility.
Educational
Digital literacy correlates with formal education. Pandemic-era remote learning revealed how unevenly this is distributed.
What this becomes when government services move online: applying for unemployment, scheduling vaccines, accessing benefits. Citizens without broadband, devices, or literacy aren't merely inconvenienced — they're structurally excluded from civic participation.
Slide 8 of 13
Inclusive Design Is a Professional Obligation
It's not charity. ACM and IEEE codes both require contribution to the public good.
The exclusion default
App requires latest smartphone. Assumes constant high-bandwidth connectivity. Designed for mouse + 27" monitor. Accessibility added "if there's time."
The design declares: we serve some users; the rest are out of scope.
The inclusive default
Works on a 5-year-old phone. Degrades gracefully on slow networks. Keyboard-navigable. Screen-reader compatible. Localizable.
The design declares: who you are, what you have, doesn't determine whether you're served.
"The app that works beautifully on a 1Gbps fiber connection and a current-generation smartphone is not a neutral product. It is a product that serves some people and excludes others. That choice was made by the people who designed it."
Slide 9 of 13
Case: IBM Watson for Oncology
Marketed as evidence-based cancer treatment AI. Trained on hypothetical patients. Withdrawn 2022.
What was claimed: Watson for Oncology analyzed patient data against medical literature to recommend cancer treatments concordant with expert oncologist guidelines. Sold to dozens of hospitals globally, including major cancer centers, by 2017.
What was discovered: the system was not trained on real patient data at scale. It was trained primarily on a small set of hypothetical patient scenarios developed by physicians at Memorial Sloan Kettering Cancer Center. Recommendations reflected MSK practice patterns, not broader clinical evidence.
What happened: internal documents leaked to Stat News in 2018 showed Watson recommended "unsafe and incorrect" cancer treatments. Hospital customers discontinued. IBM quietly withdrew the product from oncology markets by 2022.
Slide 10 of 13
The Watson Failure Pattern
Each step defensible in isolation. The aggregate sold a product that influenced patient care without disclosing what it actually was.
Marketing claim
"Evidence-based cancer treatment recommendations." Technically true — the training was supervised by oncologists. Materially misleading — the cases weren't real.
Methodology omission
IBM disclosed Watson was trained by oncologists. Did not prominently disclose the training data was primarily hypothetical, not real patient outcomes.
Hospital procurement
CIOs and procurement boards purchased based on the marketing. The technical detail of "synthetic vs. real training data" was below their evaluation depth.
Physician deference
Physicians used Watson believing the methodology was robust. Patients received recommendations from a system whose actual basis they were never informed about.
"Marketing an AI as 'evidence-based' when it was trained on hypothetical scenarios is not a technical failure. It is a communication decision made by people who knew what the training data actually was."
Slide 11 of 13
Who Bears Responsibility?
Distributed-fault systems are the hardest. Apply the four frameworks.
Stockholder
IBM had a fiduciary duty to disclose material risks to its product. The training methodology gap is material. Failure to disclose = breach of duty.
Stakeholder
Patients, physicians, hospitals, IBM, MSK all affected. Patients had least power and least information — they bore the harm without consenting to the risk.
Utilitarian
Net welfare: hypothetical benefit (good treatments) vs. concrete harm (unsafe recommendations) + market damage to AI-in-medicine credibility. Net negative.
Deontological
Marketing a product as evidence-based when it isn't = deception. Deception of a hospital that becomes deception of a patient. Duty violation regardless of outcome.
The professional position: the engineers who knew what Watson was trained on had a professional obligation to ensure the marketing matched the methodology. The product managers who shaped the marketing had a parallel obligation. Both layers failed.
Slide 12 of 13
What Builders Owe
The recurring lesson of Week 3: the IT system reaches farther than its builders' field of view.
Honest communication
Marketing must match methodology. Limitations must be visible at decision time, not buried in technical notes. The hospital CIO can't audit the training set.
Distributed harm visibility
The system displaces, excludes, or harms people who never agreed to interact with it. The builder's job is to make that visibility part of the design conversation.
Refusal of complicity
When the product manager wants to ship a system whose actual capabilities are below its marketing — the engineer's professional code requires refusal. Not subtle pushback. Refusal.
Pattern across Week 3 cases: Therac-25, Uber ATG, Watson. Each is a different industry. Each is the same mechanism: an engineering decision made under cost/schedule pressure that produced foreseeable harm to people who were not in the room.
Slide 13 of 13
Module 10 Summary
IT Impact on Society — key takeaways.
1The productivity paradox: IT investment alone doesn't create gains. Productivity follows organizational change, not the technology in isolation.
2Externalized cost: the productivity gain is real; the cost to displaced workers is real. Both are part of the ethical calculus, even if only one is in the quarterly report.
3Automation risk tiers: routine manual (done) → routine cognitive (now) → non-routine cognitive (augmentation) → social/relational (low). AI changed which tiers are at risk.
4Healthcare IT trade: EHR benefits are real; ransomware risk and algorithmic bias are real; data monetization is rarely disclosed to patients.
5Algorithmic bias is a training-data problem. Documented worse performance on underrepresented populations. Dataset audits + per-group metrics are the professional response.
6Digital divide is structural exclusion when government and essential services move online. Inclusive design isn't charity — it's a professional obligation.
7Watson for Oncology: marketed as evidence-based; trained on hypothetical patients. Recommended unsafe treatments. Withdrawn 2022. The communication decision was made by people who knew.
8What builders owe: honest communication, distributed-harm visibility, refusal of complicity. The system reaches farther than its builders' field of view.
Next up: Week 3 review (ETH-r2) and Week 3 Checkpoint (ETH-11). The patterns of IT's impact on individuals and society set up Week 4's synthesis on social media and IT organizations.