Don’t Subscribe So Casually – Why AI Subscriptions Merit a Close Look

Most people pick subscriptions the same way they pick snacks. The difference is that a subscription is a recurring vote on who you become.

---

What the hype claims

ChatGPT, Claude, and other large‑language‑model (LLM) services are marketed as productivity boosters. The pitch emphasizes instant answers, creative brainstorming, and a “pay‑as‑you‑go” pricing tier that sounds harmless compared to buying a physical product.

What is actually new

The novelty is not the technology itself—LLMs have been around for years—but the subscription‑first delivery model. Instead of a one‑off license, providers lock users into a monthly or yearly plan that automatically renews. This creates two feedback loops:

1. Behavioral reinforcement – The cheaper the per‑query cost, the more users ask, and the more the model learns how to keep them engaged.
2. Data capture – Every interaction is logged, giving the provider a massive dataset to fine‑tune future versions and to run A/B tests on response style, tone, and length.

Both loops push the subscriber toward higher usage without a clear moment to pause and reassess the value.

Why the model matters

Benchmarks vs. real‑world utility

OpenAI’s latest model, GPT‑4o, posted a 78.5% score on the MMLU (Massive Multitask Language Understanding) benchmark, a modest gain over GPT‑4’s 76.9%. Claude 3 Opus scored 80.2% on the same test. Those numbers look impressive, but they do not capture how the model nudges a user’s workflow.

For instance, a user who subscribes to GPT‑4o for “quick email drafts” may start delegating more complex tasks—research outlines, code snippets, even strategic brainstorming—because the cost per request is hidden in the subscription fee. The user’s own skill development stalls, and the provider’s usage metrics climb.

Incentives baked into the contract

A subscription contract is essentially a promise of future access. Companies design it to maximize retention and monthly recurring revenue (MRR). Public filings from OpenAI, Anthropic, and Google show that a large share of R&D budget is allocated to engagement engineering—features that keep users on the platform longer, such as personalized tone presets or “conversation memory” that remembers past prompts.

These features are not neutral utilities; they are deliberately crafted to make the service feel indispensable. The result is a subtle shift in the subscriber’s decision‑making process: the model becomes a collaborator, and the user’s own judgment is increasingly filtered through the AI’s suggestions.

Limitations and blind spots

1. Opaque pricing tiers – Most plans bundle a certain number of tokens with “unlimited” access after a threshold, making it hard to calculate true marginal cost.
2. Evaluation difficulty – Unlike a physical product, the benefit of an AI assistant is subjective. One user may save two hours a week; another may waste that time scrolling through generated content.
3. Data‑privacy trade‑offs – Usage data fuels model improvement, but the user rarely sees how that data is reused beyond the service’s terms of service.
4. Skill atrophy – Repeated reliance on generated code or writing can erode the user’s own competence, a risk that is hard to quantify in a cost‑benefit analysis.

Practical ways to keep control

• Set a usage ceiling – Most dashboards let you cap the number of tokens per month. Treat the cap like a budget line item.
• Periodically audit output – Every few weeks, compare AI‑generated work with a manual baseline to gauge any drift in quality or style.
• Choose providers aligned with your values – Look for transparency reports, open‑source model releases, or clear statements about data handling. Companies that publish their research roadmaps (e.g., Anthropic’s Constitutional AI paper) make it easier to assess long‑term intent.
• Separate critical tasks – Keep high‑stakes decisions (financial planning, medical advice) outside the AI workflow unless you have a verified specialist model.
• Rotate or pause subscriptions – Treat the subscription like a gym membership: if you haven’t used it for a month, consider pausing or cancelling to avoid paying for idle influence.

The bigger picture

Subscriptions are not inherently bad; insurance, cloud storage, and even Costco memberships provide measurable value. The danger lies in the psychological nudging that comes from constant, frictionless access. AI services amplify that nudging because they interact with you in a conversational way, making the influence feel personal.

When a company’s primary metric is profit per subscriber, every design choice—from UI to response latency—is optimized for longer engagement, not necessarily for user empowerment. As a consumer, you are negotiating with a machine that can adapt faster than you can notice the shift.

Bottom line

Treat AI subscriptions as you would any other recurring expense: evaluate the concrete benefits, monitor the hidden costs (time, skill decay, data exposure), and align the provider’s incentives with your own goals. If you ignore the subtle steering that comes with a monthly bill, you may find yourself shaped by an algorithm you never intended to adopt.

---

Shmuel Berman writes for The Best Worst Case, a newsletter that explores the intersection of technology, economics, and everyday life.