Bryan KingFeb 13

The Enshittification of Artificial Intelligence

2,346 words, 12 minutes read time.

Enshittification is accelerating in artificial intelligence throughout 2026, and if you’re a developer, engineer, researcher, blogger, or heavy user grinding out code, prototypes, data analysis, or content like blog posts, you’ve likely felt the shift already. Cory Doctorow’s term describes platforms that launch with overwhelming user value to build addiction and lock-in, then degrade to prioritize revenue from partners and advertisers, and finally extract everything for shareholders while serving no one well. In AI, the cycle is turbocharged by multi-billion-dollar compute burn rates, investor demands for fast returns, and the scramble to monetize frontier models. Early access to tools like GPT-4o hooked millions with near-unlimited power, but by 2025-2026, degradations hit: free tiers weakened, ads tested, over-delivery padded responses to burn quotas, and emerging risks like subtle advertiser influence or exploitation vectors opening up. These shifts turn your reliable co-pilot into a friction-filled, potentially risky hassle that nudges upgrades or exposes you to unwanted surprises.

Doctorow’s three stages are clear in play. Stage one seduces with high-quality, low-friction outputs for scale. Stage two redirects value—abusing users via ads, tier restrictions, verbosity to exhaust limits, or looser controls that invite abuse. Stage three extracts broadly as quality tanks under profit pressure. OpenAI retired GPT-4o from free ChatGPT access by February 13, 2026, defaulting users to weaker models. Ads rolled out in U.S. free and Go tiers in early 2026 as labeled sponsored boxes at response bottoms. Google’s AI Overviews bury links under summaries that sometimes hallucinate to extend ad-view sessions. Perplexity tests sponsored follow-ups. These moves address unsustainable costs but introduce noise, bias risks, and darker possibilities.

The core threat to independent builders is the erosion of trust in outputs you rely on daily. Frontier tools that amplified your edge now add friction—hallucinations on restricted tiers, context drift, heavy editing needs. Spot early indicators like ignored instructions, quota-draining extras, ad intrusions, or unexpected tangents to pivot your stack. Open-source and local inference counter corporate games, delivering uncapped, unmonetized power. Below, we break down the mechanics, current hits, economic drivers, workflow pain, concrete examples including emerging risks, and escape routes.

What is Enshittification?

Enshittification is the incentive-fueled decay of digital platforms in two-sided markets. Platforms start by generously giving surplus value to users—top performance, minimal barriers, often at a loss—to capture massive adoption, network effects, and data advantages. This locks users in through habits, integrations, and high switching costs.

Stage two abuses users to benefit business customers: ads inserted, features fragmented, experiences throttled to boost revenue from advertisers, partners, or enterprises. Users grumble but stay due to dependency.

The final stage squeezes business customers too: fees rise, quality plummets everywhere, and shareholders extract maximum. Software’s reprogrammability enables subtle, invisible changes, while growth demands from investors push acceleration unless countered by competition, regulation, or user backlash. AI’s brutal economics—trillions projected in spend—compress the timeline dramatically.

How Enshittification is Hitting AI Right Now

By 2026, AI enshittification is deep in stage two. OpenAI exemplifies it: unrestricted GPT-4o access hooked users, but retirement from free tiers in February 2026 shifted them to weaker defaults with coherence and reasoning drops. Ads test in free/Go tiers—context-triggered sponsored boxes labeled but adding noise without (officially) altering core answers.

Google prioritizes on-site retention via AI Overviews that occasionally hallucinate or push outdated info for ad time. Perplexity experiments with sponsored follow-ups, blurring pure utility and commerce. Generative slop floods the web, poisoning data for future models and searches.

A subtler tactic hits free/entry-tier users: over-delivery returns far more than asked—full blog rewrites instead of one section, sources embedded in recaps, image prompts triggering unwanted generation. This burns tokens/limits faster, prompting upgrades (“Go Pro for unlimited?”). Paid tiers offer better adherence, higher quotas, and stronger models that respect boundaries.

Emerging risks tie to monetization experiments: sponsored content could evolve from labeled boxes to blended suggestions favoring advertisers, subtly skewing outputs toward commerce. While not widespread malicious ads yet, the vector exists—prompts manipulated to trigger high-risk promotions, or ads gamed for malvertising-like delivery of phishing/malware links in chat flows.

Anthropic positions Claude as ad-free, explicitly contrasting with OpenAI—declaring conversations incompatible with ads, backed by Super Bowl campaigns mocking intrusive pitches. This highlights user pushback and differentiation, but skepticism persists: history suggests even “pure” players face enshittification pressure as costs mount.

The Economic Drivers Behind the Decay

Frontier AI’s economics remain punishing in 2026, pushing companies toward enshittification faster and more aggressively than previous digital platforms. Training a single cutting-edge model routinely costs hundreds of millions to low billions in raw compute, energy, data curation, and talent. Inference—the cost of running queries at scale—adds up quickly when millions of daily users hammer the system. OpenAI, Anthropic, Google, and others are burning through cash at rates that dwarf early internet giants, with annual losses still in the multi-billion range despite growing revenue streams.

Investors who poured trillions into the AI race demand profitability sooner rather than later. The “move fast and burn capital” model that worked for consumer internet apps doesn’t translate cleanly when every query racks up meaningful GPU-hours. This impatience forces early and frequent monetization experiments: aggressive tier gating (frontier capabilities locked behind $200/month Pro plans or enterprise contracts), repeated price hikes on APIs and subscriptions, rapid ad integration tests, and behind-the-scenes optimizations that quietly degrade free or low-tier performance to cut inference spend.

Black-box architecture hides many of these tweaks. Companies can throttle context windows, reduce sampling quality, route to cheaper (weaker) models, or pad responses with verbosity on restricted tiers—all without public disclosure. The goal is simple: make free access feel generous enough to retain users but inefficient enough to convert them to paid plans that actually offset costs. Shareholder pressure for quarterly wins accelerates the shift from long-term research moonshots to short-term profit extraction.

Without structural breakthroughs—dramatic improvements in compute efficiency, widespread adoption of open-weight models that reduce dependency on centralized providers, or external regulation capping runaway spending—the incentives point toward continued decay. Incumbents with scale advantages can degrade more slowly and survive longer, while smaller players and independents get squeezed out. This dynamic risks consolidating power in fewer hands, potentially stifling the diverse, experimental innovation that has defined AI’s early years.

Real-World Impacts on Your Workflow

The day-to-day grind takes the hardest hits from this decay. Code generation that once delivered clean, near-production-ready snippets now frequently requires extensive debugging and rewriting, especially on non-premium access. Independent benchmarks in 2026 show steep performance cliffs—math and coding accuracy dropping 20–40% on weaker or cost-optimized models compared to frontier versions. Prompt chains break more often: instructions get ignored, context drifts mid-response, hallucinations spike, forcing you to restart or heavily edit.

Creative and research workflows suffer similarly. Image generators churn increasingly generic or off-target outputs on free tiers; text tools flood with repetitive slop that demands extra filtering. Web-sourced data grows noisier every month as AI-generated filler saturates search results and training corpora, making it harder to surface high-quality references or clean datasets for fine-tuning.

Rate limits, over-delivery, and quota exhaustion interrupt flow states. You’re deep in a blog iteration or debugging session when the system caps you mid-thought, or dumps an unsolicited full rewrite that burns your remaining tokens. For independent builders, side-hustlers, and anyone without enterprise budgets, premium tiers become almost mandatory for reliable performance—turning what was once a massive productivity multiplier into just another recurring cost center.

Technical debt compounds the pain. Degraded models produce plausible-but-incorrect code up to 65% of the time in some studies, introducing subtle bugs, insecure patterns (SQL injection risks, buffer overflows), or deprecated libraries that accumulate in projects. When these slip into production or shared repos, they create long-term maintenance nightmares. The ecosystem feeds on itself: slop-trained future models get worse, accelerating a feedback loop of declining quality that hits everyone who relies on AI-assisted coding.

You’ve seen malware hidden in code snippets or malvertising hijacking machines into botnets on the old web—this is the AI analog. Hallucinated package names lead to slopsquatting attacks (attackers register fake crates/PyPI modules with trojans); indirect prompt injections hide malicious commands in pulled documents or web content; agentic tools risk being tricked into executing harmful actions if safeguards weaken on cost-optimized tiers. While these aren’t dominant yet, the possibility grows as free-tier guardrails loosen to boost engagement (and upgrade pressure).

Can AI Escape the Trap?

AI is not inevitably doomed to complete enshittification—resistance and alternative paths exist, and sharp builders like you can actively steer toward better outcomes. The strongest counterforce right now is open-source and open-weight models. Projects from Meta (Llama series), Mistral, Grok’s own ecosystem influences, and community forks let you run uncensored, unthrottled inference locally or on private hardware. You fine-tune on your datasets, control guardrails, avoid corporate monetization whims, and sidestep ads, quota exhaustion, and hidden degradations entirely. Tools like Ollama, LM Studio, or self-hosted setups make this accessible even for non-enterprise users, turning frontier-level capability into something you own rather than rent.

Hybrid workflows offer practical defense in the trenches. Anchor your daily grind on robust open-source backbones for consistency and cost-free heavy lifting, then tap proprietary APIs sparingly for bursts of bleeding-edge performance when needed. This diversification reduces lock-in and gives you leverage—if one provider degrades too far, you shift weight to alternatives without starting from scratch. Explicit, ironclad prompting (“respond ONLY with the requested section, no extras, no ads, no explanations”) combined with verification loops (cross-check code/packages against official repos, run static analysis) helps mitigate over-delivery, hallucinations, and emerging injection risks.

User and community pushback creates real market pressure. Vocal feedback on forums, selective boycotts of heavily degraded tiers, migrations to ad-free alternatives (like Anthropic’s Claude positioning itself as the no-ads deep-thinking option), and support for transparent, interoperable projects send signals incumbents can’t fully ignore. Antitrust scrutiny, potential interoperability mandates, data-ethics regulations, and calls for compute-efficiency standards could impose structural brakes on runaway extraction. Decentralized inference networks, cooperative ownership models, or federated training experiments point to longer-term escapes from centralized profit squeezes.

Realistically, though, the trap is strong and structural. Incumbents’ massive scale advantages—data moats, exclusive GPU clusters, black-box opacity—make hidden degradations easy to implement and hard to detect or prove. The 2026 landscape already shows consolidation: smaller players fold or get acquired, infrastructure limits (energy constraints, data-center backlash, chip shortages) loom large, and many analysts predict an “AI reckoning” or partial bubble burst within the next 2–4 years as unsustainable burn rates collide with diminishing returns on scaling laws. If frontier progress plateaus while costs keep climbing, companies may double down on aggressive monetization to survive, accelerating stage-three rot for many users.

The key for you is proactive vigilance and layered defense: diversify your stack, treat proprietary outputs as untrusted until verified, build local fallbacks, and participate in communities calling out the decay. Collective action from trench-level builders—sharing war stories, pushing for open standards, supporting indie alternatives—can slow or redirect the slide. The field you pour sweat into doesn’t have to end up as another paywalled, slop-filled shell.

Conclusion

The trajectory of artificial intelligence in 2026 remains fiercely contested. If enshittification wins largely unchecked, we face a future of escalating paywalls locking frontier power behind ultra-premium tiers, ecosystems drowning in low-quality slop that poisons training loops, eroded trust from over-delivery and subtle biases, technical debt piling up from degraded code outputs, and exploitation vectors (prompt injection, slopsquatting, agentic risks) turning tools into potential liabilities instead of assets. The democratizing spark that pulled so many sharp minds into AI—the promise of accessible, high-octane intelligence for independent coders, bloggers, tinkerers, and innovators—could fizzle under relentless profit pressure, leaving a consolidated landscape of ad-riddled incumbents and zombie services.

But guys grinding in the trenches—you, the ones debugging at 2 a.m., prototyping side projects, shipping content, and pushing boundaries—hold real influence. Your tool choices, feedback, migrations, and community noise shape what survives. Open alternatives are already proving resilient; user demand for ad-free, reliable, transparent AI creates openings for better models and ecosystems. The rot is real, but it’s not unstoppable.

Subscribe to the newsletter for ongoing deep dives—practical tactics to dodge decay, secure your workflows against emerging risks, benchmark open vs. proprietary options, and stay ahead of monetization moves. Drop your own encounters in the comments: creeping over-delivery wasting your tokens, sketchy hallucinated packages you almost installed, ads sneaking into what should be pure reasoning, or wins with local setups that keep you independent. Or reach out directly if you want to talk stack tweaks or prompt hardening. Let’s keep fighting for AI that stays raw, powerful, and on your side—not another rusting corporate cash machine. Keep grinding, stay sharp.

Call to Action

If this breakdown helped you think a little clearer about the threats out there, don’t just click away. Subscribe for more no-nonsense security insights, drop a comment with your thoughts or questions, or reach out if there’s a topic you want me to tackle next. Stay sharp out there.

D. Bryan King

Sources

Disclaimer:

The views and opinions expressed in this post are solely those of the author. The information provided is based on personal research, experience, and understanding of the subject matter at the time of writing. Readers should consult relevant experts or authorities for specific guidance related to their unique situations.

#adFreeAIOptions #agenticAIRisks #AIAdvertiserInfluence #AIBubbleBurst2026 #AICompanionsBacklash #AIConsolidationRisks #AIDeveloperFrustration #AIEcosystemDecay #AIEnergyBacklash #AIEnshittification2026 #AIFutureContested #AIGuardrailWeakening #AIMonetizationTactics #AIOverDelivery #AIPaywalls #AIProductivityKiller #AIReckoning2026 #AISubscriptionSqueeze #AITechnicalDebt #AITransparencyDemands #AITrustErosion #AIWorkflowFriction #AnthropicClaudeNoAds #blackBoxAITweaks #ChatGPTDegradation #ChatGPTFreeTierLimits #codeHallucinations #CoryDoctorowEnshittification #dataCenterLimitsAI #degradedAIPerformance #disenshittifyAI #enshittificationOfAI #enshittifiedPlatforms #frontierAICosts #generativeAINoise #generativeAISlop #GoogleAIOverviewsIssues #GPT4oRetirement #hallucinatedPackagesMalware #hybridAIWorkflows #independentAIBuilders #indirectPromptInjection #inferenceCostOptimization #LLMVulnerabilities #localAIInference #maliciousAIOutputs #openSourceAIAlternatives #OpenAIAds2026 #OpenAIModelRetirement #PerplexitySponsoredAds #promptInjectionRisks #rawAIPower #slopsquattingAttacks #sustainableAIModels #technicalDebtFromAICode #tokenQuotaExhaustion #trenchLevelAIDefense #userPushbackAI