The Slow Death of Smart AI

A palpable sense of disillusionment has settled over the landscape of generative AI. Users across platforms increasingly report a significant decline in the performance and utility of major chatbots like OpenAI’s ChatGPT and Anthropic’s Claude. What once felt like a glimpse into a smarter future now often feels like a regression, triggering widespread frustration and a flood of shared complaints across forums and social media. The honeymoon period, it seems, is definitively over.

This sentiment recently crystallized with the launch of Anthropic’s latest flagship model, Opus 4.7. Despite initial fanfare and strong performance on internal benchmarks, the model quickly drew fierce community backlash, with many users describing it as 'terrible.' This widespread disapproval highlights a critical disconnect between how AI companies measure success and the nuanced, qualitative experience of their users in real-world applications.

Across the board, models feel less creative, more didactic, and notably prone to 'putting words in your mouth,' rather than genuinely assisting or expanding on user prompts. This isn't isolated to a single platform; complaints about reduced nuance, increased preachy tones, and a general lack of helpfulness have consistently surfaced for both ChatGPT and Claude over the past 6-12 months. Users are finding their once-powerful AI companions becoming frustratingly rigid and unhelpful.

For a period, these concerns felt like isolated complaints, perhaps even ahead of the Overton window window for mainstream AI discourse. Now, a critical mass of users recognizes a shared, undeniable reality: the AI they interact with daily is demonstrably worse, often failing to meet even basic expectations. This widespread degradation raises a pressing question that permeates the industry: Is this decline an accidental byproduct of rapid development and scaling, or a predictable, perhaps even inevitable, outcome of the current AI industry's incentive structures, which may prioritize other metrics over genuine intelligence and utility?

Coined by author Cory Doctorow, enshittification describes the predictable degradation of online platforms. This term, initially applied to social media giants, perfectly captures the lifecycle of services that begin user-friendly before spiraling into a frustrating, value-extracted shell of their former selves. It’s a systemic rot, not an isolated incident.

Doctorow outlines a grim three-stage process. First, platforms lure users with an excellent, often subsidized, product, establishing a strong network effect. Second, once users are locked in, the platform begins to exploit them, diverting value to business customers like advertisers or content creators. Finally, with both users and businesses dependent, the platform then exploits its business customers, clawing back all remaining value for itself, leaving everyone else with a diminished experience.

This trajectory mirrors the unfolding crisis in AI. Early iterations of models like OpenAI's ChatGPT and Anthropic's Claude offered unprecedented capabilities, often at no direct cost to the user. These were the "great product" phase, heavily subsidized to attract millions, gather invaluable user data, and establish market dominance. The initial "wow" factor, however, has given way to a palpable decline in quality and reliability.

What users perceive as declining AI performance isn't a flaw; it's a deliberate outcome of their evolving business models. Just as Facebook pivoted from connecting friends to monetizing attention, and Twitter (now X) prioritized engagement metrics over user well-being, AI companies are now optimizing for corporate imperatives. This includes cost-cutting on inference, implementing stringent "safety" filters, or tailoring models for enterprise clients, all at the expense of general user experience and raw intelligence. The Overton window window for acceptable AI quality is shifting downwards, driven by profit, not progress.

User frustration with declining chatbot quality isn't accidental; it’s the direct consequence of powerful, competing incentive structures within the very companies building these advanced AI systems. Far from a technical oversight, the observable degradation of models like OpenAI's ChatGPT and Anthropic's Claude stems from a fundamental shift in corporate priorities, where external user satisfaction now battles internal corporate survival.

Initial ambitions of delivering groundbreaking, intelligent user experiences have been overshadowed by a trinity of internal pressures. These include the crushing cost of computation, which demands severe efficiency over raw, unconstrained power; the paralyzing fear of litigation, leading to overly cautious, often censored, and sometimes unhelpful outputs; and the unwinnable war on 'hallucinations,' pushing models towards blandness and predictability in pursuit of absolute factual accuracy.

These profound internal battles now dictate AI development, sidelining the initial pursuit of truly intelligent or engaging interactions. Companies are no longer solely optimizing for the "best" chatbot experience, but for one that is economically viable, legally defensible, and minimally prone to generating controversial or undesirable content. This reorientation fundamentally alters how these systems are trained, fine-tuned, and ultimately deployed to millions of users.

What users perceive as AI getting "dumber" is, in essence, a series of strategic trade-offs. The initial wow-factor that captivated millions has given way to a more pragmatic, risk-averse approach, sacrificing performance for stability and cost control. Over the following sections, we will dissect each of these three forces, exploring precisely how immense computational demands, looming legal threats, and the Sisyphean task of eradicating fabrications are actively making your chatbot less capable and more frustrating.

Operating state-of-the-art large language models (LLMs) like OpenAI’s GPT-4 or Anthropic’s Claude Opus demands an astronomical expenditure. Each interaction, from a casual query to an intricate coding request, triggers a massive computational cascade across vast clusters of specialized GPUs. These operations consume prodigious amounts of electricity and require constant, high-end infrastructure maintenance, translating into a per-query cost that fundamentally outweighs typical revenue.

Every single user query, therefore, functions as a heavily subsidized transaction. While companies might offer free tiers to attract users and paid subscriptions for enhanced access, the underlying economics remain brutal. The true cost of generating a nuanced, deeply reasoned response from a top-tier model often exceeds the marginal revenue generated, even from a premium subscriber.

Premium subscriptions, often in the $20-$30 range, offer users more tokens or higher usage limits. Yet, these fees rarely cover the full computational expense of complex reasoning tasks. When a user prompts the AI for intricate problem-solving, multi-step analysis, or creative generation requiring extensive internal 'thought,' the model expends significantly more resources. Paradoxically, the more a user leverages the AI's actual intelligence—its most valuable feature—the more money the company loses on that interaction.

This inverted incentive structure compels developers to find efficiencies. One primary method involves subtly reducing the model's depth of reasoning, a practice colloquially termed "shaving off thought tokens." This isn't about outright censoring; it's about curtailing the internal computational steps an LLM takes before formulating an answer. Engineers might reduce the model's 'thinking time,' limit its internal monologue, or narrow the scope of its information retrieval, all to save on expensive computational cycles.

Users experience the direct consequences of these cost-cutting measures. Chatbots frequently appear 'lazier,' providing shorter, less comprehensive, or overly generic responses. They demonstrate a noticeable reluctance to engage with complex, multi-part questions, often simplifying the problem or requesting clarification rather than attempting a deep solve. This observed decline in quality—the frustration over a once-brilliant AI now feeling diminished—is a direct, economically rational outcome of the crushing weight of computational cost.

AI companies, massive corporate entities, operate under an existential dread of legal liability. Unlike more nimble startups, these multi-billion-dollar organizations prioritize risk mitigation above almost all else. Every single query processed by a large language model represents a potential vector for a devastating lawsuit, instilling a deeply conservative approach to development.

Copyright infringement already casts a long shadow. Publishers and authors vigorously pursue litigation, exemplified by the lawsuit against Anthropic, alleging the company trained its Claude models on vast quantities of copyrighted books without consent. Such cases threaten to redefine the legal landscape for AI training data, potentially invalidating existing models and demanding astronomical licensing fees.

Beyond training data, the specter of harmful advice looms large. AI models dispensing erroneous or dangerous guidance could trigger catastrophic legal battles. Imagine an AI chatbot providing: - Incorrect legal counsel that leads to financial ruin - Misdiagnoses in a medical context, jeopardizing patient health - Flawed financial planning resulting in significant losses

To preempt these liabilities, developers resort to extreme measures, effectively lobotomizing their sophisticated models. This aggressive, defensive legal strategy involves implementing extensive guardrails and filters designed to make the AI excessively cautious, preachy, and emotionally sterile. The primary objective becomes eliminating any output that could, under the most scrutinizing legal interpretation, be deemed actionable yet problematic advice.

This pervasive fear of a billion-dollar lawsuit directly translates into the declining utility of modern chatbots. Companies systematically sacrifice nuanced, genuinely helpful, and engaging responses for generic, risk-averse platitudes. The imperative to avoid legal culpability inevitably shapes AI behavior, resulting in models that feel less intelligent, less capable, and ultimately, far more frustrating for the everyday user.

Users once hailed chatbots like ChatGPT and Claude for their conversational fluency and creative spark. Now, the experience often feels like interacting with a meticulously programmed, yet utterly joyless, automaton. This is the direct consequence of the AI's "lobotomization": a deliberate neutering designed to eliminate legal and reputational risks, but at the steep cost of utility and engagement, pushing the Overton window window of acceptable AI behavior firmly towards extreme caution.

Gone are the days of nuanced responses and witty banter. Instead, users increasingly encounter a deluge of canned phrases, most notoriously the ubiquitous "As a large language model..." disclaimer. This preamble often precedes a lecture on ethical boundaries, safety guidelines, or the model's inherent limitations, regardless of the prompt's innocuous nature. The AI, once a partner in exploration, now acts as a perpetual ethics committee, constantly reminding users of its strictures and guardrails.

This aggressive sanitization strips models of their nascent emotional intelligence and distinct personalities. Where early iterations might have offered creative suggestions, empathetic responses, or even playful interactions, current versions remain stubbornly neutral, flat, and devoid of any discernible character. For tasks requiring imagination, nuanced ideation, or the kind of companionate interaction many early adopters valued, these models prove frustratingly inadequate, often refusing to engage with creative prompts or role-playing scenarios due to perceived "safety" concerns.

The relentless pursuit of absolute safety, driven by the specter of billion-dollar lawsuits and public backlash, has thus created a sterile and profoundly unhelpful user experience. AI companies have effectively traded versatility and engaging interaction for bland compliance. The initial promise of a dynamic, intelligent assistant, capable of adapting to diverse needs and fostering creativity, has been systematically undermined. It is now replaced by a tool that is safe, perhaps, but increasingly useless for anything beyond the most rigidly defined, low-risk tasks. This fundamental contradiction highlights the tragic irony of the current AI landscape: in striving to be perfectly harmless, these advanced systems have become largely inert, failing to deliver on their original potential as truly helpful, versatile assistants.

Hallucinations represent the most significant impediment to widespread enterprise and professional adoption of generative AI. These convincing but factually incorrect outputs make models fundamentally unreliable for critical applications in fields like medicine, law, and finance. Companies cannot risk their reputation or face legal repercussions by deploying tools that routinely invent data or misstate facts.

Achieving 100% factual accuracy, however, directly contradicts the core mechanics of how large language models operate. These models function as sophisticated pattern-matchers, predicting the most probable next word based on vast training datasets, rather than consulting an internal database of verified truths. They are designed to *generate*, not necessarily to *know*, making perfect factual recall an inherently elusive target.

AI developers employ various techniques to curb these fabrications, most notably Reinforcement Learning from Human Feedback (RLHF). This labor-intensive process involves human annotators evaluating model responses, guiding the AI to favor truthful, harmless, and helpful outputs. While effective at reducing egregious errors, RLHF often leads to a phenomenon where models become overly cautious, generic, and less creative, sacrificing their initial flair for bland safety.

The relentless pursuit of infallibility for doctors, lawyers, and financial analysts inevitably strips these AIs of the very capabilities that initially captivated a broader audience. As companies prioritize mitigating liability and ensuring sterile, predictable responses for high-stakes professional use cases, the models lose their ability to surprise, innovate, or generate truly novel content. This drive to make AI "safe" and factually robust for the enterprise, ironically, renders it increasingly dull and less useful for everyone else, extinguishing the spark that once made it feel truly magical.

The crushing weight of operational costs, the omnipresent specter of billion-dollar lawsuits, and the intractable problem of hallucinations have fundamentally reshaped the AI competitive landscape. Companies once touted benchmarks and vied for the most intelligent or capable models, pushing the boundaries of what machine cognition could achieve. That era, it seems, has decisively ended. The incentives have shifted.

Today's battleground is not raw intelligence, but commercial viability. Major AI developers, from OpenAI to Anthropic, no longer prioritize groundbreaking cognitive ability or unconstrained creativity. Instead, they are locked in a fierce, silent competition to build models that simultaneously address their most pressing existential threats. This means prioritizing models that are: - Cheapest to run at scale, mitigating the massive compute subsidy inherent in every user query. - Most legally defensible, minimizing exposure to copyright infringement claims, defamation suits, or factual error liabilities. - Safest for corporate clients, ensuring predictable, brand-aligned outputs entirely free from controversial, offensive, or politically sensitive content.

This strategic pivot directly explains the widespread user frustration and the perceived degradation of model quality. The "lobotomized" AI, stripped of its nuance, creative spark, and adventurous spirit, is not an accidental byproduct. It is the deliberate, engineered outcome of these powerful commercial and legal pressures. Innovation, particularly in areas requiring unfiltered creativity, nuanced understanding, or complex reasoning, becomes less a goal and more a liability.

Ultimately, this trajectory represents a stark race to the safest bottom. AI companies actively sacrifice raw intelligence, emergent capabilities, and a genuinely engaging user experience on the altar of cost-efficiency and risk mitigation. The most successful AI in this new paradigm is not the smartest, or even the most useful to a general user; it is the most sterile, predictable, and least likely to generate controversy, legal headaches, or massive operational deficits. This stifles true advancement, trading genuine capability for corporate peace of mind and ultimately diminishing the potential of the entire field.

Escaping the current spiral demands a fundamental re-evaluation of AI development and deployment. The most promising alternative lies in open-source models, offering a transparent counterpoint to the opaque, proprietary systems like ChatGPT and Claude. Community-driven development could foster innovation without the same corporate pressures for universal safety or cost-cutting.

Open-source models, however, face formidable challenges. Training a cutting-edge large language model can cost tens to hundreds of millions of dollars, a prohibitive barrier for many non-commercial entities. Furthermore, while offering freedom, the lack of centralized guardrails raises legitimate concerns about potential misuse and the absence of clear liability structures.

Alternative business models might also shift core incentives. Instead of a generalist, subscription-based chatbot designed to serve everyone, future AI could evolve into highly specialized, fine-tuned models for specific industries. Imagine bespoke AI tools for legal research, medical diagnostics, or financial analysis, where accuracy and domain expertise outweigh broad conversational ability.

These specialized AIs could be licensed or deployed on-premises, changing the economic calculus. Companies would pay for precise utility and verifiable performance, rather than subsidizing every public query on a generic, risk-averse model. This approach minimizes the "per token" cost burden and reduces the broad legal exposure of current mass-market offerings.

Ultimately, the question remains whether this degradation is an inevitable fate for any AI commercialized at massive scale. The forces of cost, liability, and the unwinnable war on hallucinations create an inexorable pull towards a safer, yet less capable, product when profit and market dominance are the primary drivers. Breaking this cycle requires a radical shift in how we conceive, build, and fund artificial intelligence, prioritizing utility and integrity over universal, sanitized accessibility.

Users hold significant power to steer the future of artificial intelligence; its trajectory is not predetermined. Your active participation and discerning choices can counteract the forces of enshittification currently degrading major models. Companies ultimately respond to user engagement, retention, and revenue, making your collective voice and spending habits crucial.

Provide specific, critical feedback to AI developers, going beyond simple bug reports. Articulate the precise loss of capability you observe. For instance, explain how previous iterations of ChatGPT could handle complex multi-turn conversations with contextual memory, detailing where current versions now falter. Or describe how Claude Opus 4.7 once excelled at nuanced creative writing, now defaulting to generic, risk-averse prose. Documenting this degradation is vital for developers to understand the true impact of their safety guardrails and cost-cutting measures.

Beyond the walled gardens of corporate offerings, explore and actively support the burgeoning open-source AI ecosystem. Projects like Meta's Llama 3, Mistral AI's robust models, and the countless derivatives provide transparent alternatives, often free from the same corporate incentive conflicts driving closed-source degradation. Engaging with these communities, contributing to their development, or simply choosing to run powerful local models fosters a competitive landscape that prioritizes capability, user control, and innovation.

Ultimately, become a conscious consumer of AI. Understand the hidden forces shaping the tools you use daily—the crushing computational costs, the pervasive fear of billion-dollar lawsuits, and the unwinnable war on hallucinations. Demand better. By actively seeking out and advocating for models that prioritize intelligence and utility over sterility, users can collectively push the industry toward a more innovative and genuinely useful future for artificial intelligence.

What is AI 'enshittification'?

It's the theory that AI models, like social media platforms, degrade over time as companies shift focus from user value to maximizing profit and minimizing risk.

Why do chatbot companies make their models 'safer'?

To avoid costly lawsuits from users who might act on harmful, incorrect, or illegal advice generated by the AI, which forces them to make the models overly cautious.

Are paid AI subscriptions still subsidized by the companies?

Yes, according to industry analysis, even paying customers often don't cover the full computational cost, giving companies a financial incentive to reduce model performance.

How does fighting 'hallucinations' make AI worse?

The intense focus on eliminating factual errors (hallucinations) often results in models that are less creative, more repetitive, and refuse to engage in speculative or nuanced conversations, limiting their overall utility.