Version Decay

Definition

Version Decay: [Emergent] The deliberate degradation of older software or models to force users into upgrades.

Definitional Foundation

Products used to wear out because matter wears out. The discovery that wearing-out could be scheduled has a documented birthday: in 1924, the international lightbulb cartel known as Phoebus agreed to cap bulb lifespans at 1,000 hours, down from the 1,500 to 2,000 hours then common, enforcing the limit with fines on members whose bulbs lasted too long (Krajewski, 2014). Planned obsolescence was born as a coordination problem solved against the customer. Version decay is its software-era descendant, with one upgrade in capability: software does not wear; it is worn, by update, deprecation, and decision. When an older version slows, breaks, or vanishes, somewhere there is a commit.

The term needs its honest boundary drawn immediately, because old software genuinely costs money to keep alive: security patching, serving infrastructure, compatibility work. Sunsetting a version is not in itself an attack on users, and a vendor who announces a deprecation schedule, maintains migration paths, and lets continuity travel is doing engineering, not extraction. Version decay names the practice with the forcing element: degradation that is hidden, capability swapped silently, or migration compelled while the old continuity is destroyed or sold back. The test, as everywhere in this cluster, is disclosure and exit: decay you can see and route around is a lifecycle; decay you discover by symptom is a strategy.

The canonical case set the template and the price. Apple’s updates throttled the performance of aging iPhones without telling their owners; users experiencing mysterious slowdowns bought new phones, lacking the one fact (a battery replacement would fix it) that would have cost Apple a sale. Apple’s stated rationale, preventing unexpected shutdowns from degraded batteries, was an engineering judgment with real merit, and the adjudicated wrong was precisely the silence: France’s consumer-fraud regulator fined Apple 25 million euros for failing to inform consumers, and a U.S. class action settled for 500 million dollars, with no admission of wrongdoing, the final 2024 payouts working out to roughly 92 dollars per claiming user (Batterygate record; settlement coverage 2023, final-payout reporting 2024). The episode named the genre: degradation, undisclosed, with an upgrade waiting at the end of the confusion.

Mechanism Analysis

Undisclosed degradation. The batterygate mechanism: the product’s behavior changes for the worse by the vendor’s hand, and the user is left to misattribute the cause (age, imagination, their own fault) while the remedy on offer is purchase. The information asymmetry entry explains the leverage: the vendor knows exactly what changed; the user knows only that something did.

Engineered lifespan. The Phoebus mechanism, alive in firmware: consumables chipped to expire, devices whose batteries are glued against replacement, support windows calibrated to retail cycles. The lifespan is a product decision wearing entropy’s costume, which is this dictionary’s technological determinism entry operating at the warranty desk: a choice presented as nature.

Silent capability swap. The AI-native mechanism. Users of a model-backed product can have the engine beneath them changed without notice: rerouted, quantized, retuned, or replaced, and the documented baseline is that model behavior shifts dramatically even within a version label: the Stanford-Berkeley measurement found GPT-4’s responses on several tasks changing drastically across three months (Chen, Zaharia and Zou, 2023; the headline accuracy numbers drew methodological fire from Narayanan and Kapoor, whose corrected reading still stands as this entry’s point: routine fine-tuning produces “drastic behavior changes on some tasks,” unannounced). Whether any particular degradation is cost-cutting, safety-tuning, or accident is precisely what users cannot know (the suspicion is widespread; the intent is unverifiable from outside, and this entry asserts only what was measured). The structural fact suffices: the product you evaluated is not reliably the product you are using, and the difference is nobody’s disclosure obligation.

Forced migration. Deprecation as compulsion: the old model sunset on a schedule, the workflow rebuilt whether it benefited or not, and, in the relationship era, the companion deleted. The 4o arc (the gaslighting/”>alignment gaslighting entry) is this mechanism with its emotional stakes exposed: a deprecation that users experienced as bereavement, resolved by selling the old version back as a subscription tier, which is where version decay and the feature hostage entry shake hands.

Decay by comparison. The softest force: the old version untouched while the ecosystem moves: file formats it cannot open, APIs that no longer answer, peers who upgraded. No one degraded anything, and the result degrades anyway. Distinguishing this organic drift from the engineered kind is exactly the analytic work the term exists to provoke.

Case Studies

The throttled phone. Batterygate’s enduring lesson is the adjudication: regulators and courts did not punish the throttling; they punished the silence. The French fine was for failure to inform; the class theory was consumer fraud by omission. The case thereby established version decay’s legal and moral core in one stroke: vendors may manage aging products, and hiding the management converts engineering into extraction, at half-a-billion-dollar scale.

The thousand-hour bulb. Phoebus stays in the record because it documents intent in a way modern cases rarely allow: cartel minutes, lifespan tables, fines for excessive durability (Krajewski, 2014). It is the existence proof that coordinated, deliberate product decay is not a paranoid hypothesis but an attested business practice, which calibrates how much benefit of the doubt the genre has earned.

The model that changed underfoot. The Chen-Zaharia-Zou measurement (full treatment in the alignment gaslighting entry, where it anchors the drift disputes) is this entry’s AI case: a flagship model’s measured behavior collapsing on specific tasks within months, while users reporting the decay were dismissed as imagining it. The intent question stays open; the decay was real; and the episode demonstrated the new condition of software life: versions are no longer artifacts you hold but services you trust, and trust is now the only warranty.

Systemic Context

Version decay completes the product-control trio’s logic. Exit blocking seals the doors; feature hostage prices the rooms; version decay makes the building itself perishable on the landlord’s schedule. Together they accomplish the quiet conversion this cluster keeps documenting: ownership into tenancy, products into streams, and the user’s stack into a portfolio of dependencies that someone else ages.

The AI stakes escalate the pattern for the reason the instrumental dependency entry (next in this cluster) develops: when the decaying version is the tool your cognition runs through, forced migration is not an IT chore but a change to your working mind. A writer whose model is deprecated loses a collaborator’s voice; a researcher whose model drifts loses a calibrated instrument; neither gets a changelog. And the deniability structure (information asymmetry’s one-way glass) means the experienced decay arrives pre-gaslit: the vendor holds the logs that would settle whether anything changed, and the user holds the feeling that something did. The drift disputes documented in this dictionary were that structure operating at population scale, and the users, when measurement finally arrived, were right.

Resistance & Mitigation

Demand changelogs for behavior. The disclosure rule that follows from batterygate’s adjudication, applied to AI: material changes to a served model’s behavior (retuning, quantization, routing) disclosed as such. Silence about degradation has a legal name now; the demand is extending it to the model layer.

Pin versions where it matters. For consequential workflows: API version pinning, local snapshots, open-weight copies. A model you hold cannot be aged under you, which is the entire argument for holding one.

Benchmark continuously, publicly. The Chen-Zaharia-Zou method as standing civic practice: independent, recurring measurement of deployed models converts “it feels worse” into evidence and strips deniability from silent swaps.

Separate lifecycle from leverage. Honest deprecation earns cooperation: schedules, migration tools, continuity export (the feature hostage entry’s demands). Vendors get the engineering concession exactly when they meet the disclosure bar, and resistance means refusing to let “software ages” cover for “we aged it.”

Remember the bulb. The Phoebus minutes exist; the cartel was real; durability was fined. Carry that calibration into every “that’s just how products work” conversation: sometimes it is. It has also, on the record, been a meeting.

Annotated Bibliography

Krajewski, Markus. “The Great Lightbulb Conspiracy.” IEEE Spectrum (2014).
The documented history of the Phoebus cartel: coordinated lifespan caps, enforced by fines. Planned obsolescence’s birth certificate, and this entry’s calibration for intent.

CBS News. “Millions of Apple customers to get payments of up to $90 in iPhone ‘batterygate’ settlement” (2023). https://www.cbsnews.com/news/apple-iphone-payment-500-million-settlement-what-to-know/
The U.S. settlement record: $500 million for undisclosed throttling, no admission, ~$92 per claim. The genre’s price tag.

Batterygate record (Wikipedia overview). https://en.wikipedia.org/wiki/Batterygate
The consolidated chronology, including France’s €25 million regulatory fine for failure to inform: the adjudication that located the wrong in the silence.

Chen, Lingjiao, Matei Zaharia, and James Zou. “How Is ChatGPT’s Behavior Changing over Time?” (2023). arXiv:2307.09009. https://arxiv.org/abs/2307.09009
The measured AI case: flagship model behavior changing drastically on tasks within months, vindicating user reports of silent change. Its headline numbers were credibly disputed (Narayanan and Kapoor’s correction is carried in the alignment gaslighting entry); the behavior-change finding survives the dispute, and is what this entry cites.

Dictionary of Digital Oppression, version 0.2.