The Blunder of Subscription-Based AI

Something I’m watching closely, as the pressure mounts on AI firms to demonstrate a path to profitability, is how subsidized subscriptions influence user behavior. As some of you may know, I’ve been delivering AI literacy workshops to help non-technical leaders and their staff make better decisions about where and how to leverage the tech. This gives me an opportunity to observe how people outside of my group of expert technical peers use LLMs in real time.

One thing that is immediately noticeable is how many people just fire away with a half-baked, vague prompt, get a mediocre answer, and then refine from there. But this is expected behavior because most people are on a subscription plan. Today, these plans are heavily subsidized, and within them, there is no real cost to this behavior. The meter doesn’t feel like it’s running, and if you hit the limit, you just wait a few hours and start up again.

However, to have even a chance at profitability, this must change, and I don’t think the industry, with all its misconceptions about how real humans outside the tech bubble behave, realizes the can of worms they’re about to open.

Building the Wrong Habits

When a resource, like these subscription models, feels unlimited, people don’t optimize. This is rational behavior. AI subscriptions are structured to reward volume over intention. There is no incremental cost to inefficient prompting; in fact, if you don’t hit your limits, in a way, you’re wasting money.

The result is that millions of users are being trained to interact with AI in the least efficient way possible: exploratory, low-stakes, stream-of-consciousness prompting. Toss something in and see what comes out.

I’m often guilty of this as well. After all, I am a human.

That behavior makes sense under the current subsidized subscription model, but Sam Altman’s recent comment about treating intelligence as a utility got me thinking… he’s right; subscription models aren’t the future. Metered usage/tokens are the only way they have a chance of achieving profitability.

The behavior of subscription users makes sense today, but in a metered future, it will set them up for a rude awakening.

The Token Economy’s Economic (and Emotional) Toll

So, let’s think this through by doing something big tech seems increasingly incapable of: thinking about things from the user’s perspective. Intelligence as a utility means that you pay for exactly what you use, just like electricity or cloud computing. What does that mean for these subscription users? I have a few thoughts.

Wordiness, Engagement, and Sycophancy

Today, most LLMs are tuned for engagement. They do this by asking follow-up questions in their output, praising the user, and, because it fools people, often using excessive wordiness when making a point, because to certain people that sounds like intelligence.

As soon as users start watching a cost meter, those verbose, unnecessarily padded AI responses go from eye-rolling and annoying to a measurable extra charge. Those multiple paragraph preambles before the actual answer are going to feel like being charged for filler. Maybe “feel like” is the wrong phrase, because users will really be paying for the fluff.

The upcoming backlash from users on a budget isn’t hard to predict, but no one at these firms is talking about it because, again, they don’t think about the emotions of their human customers.

Hallucinations and Low Quality Output as a Billing Dispute

A hallucination today is moderately annoying, but again, because of the illusion of unlimited prompting in subscriptions, you just toss the output and start again. Whoops, guess I didn’t write that prompt correctly. Oh, I guess the training data isn’t up to date on that topic. Ew, I can’t cite THAT source.

In a metered future, every single time this happens, it will cost you money. “I paid $X for this?!” is going to become a standalone social media category. You see Karens throwing tantrums about their food being slightly off, and you think they’re just going to sit back and pay for low-quality or hallucinated output without raising a stink?

Task Anxiety as a New Category of Therapy

I remember the days of metered internet, and I don’t want to go back. Rationing, paranoia about what apps are running based on whether I’m on the wifi or not, the dreaded overage bill? Metered intelligence is a much worse version of this problem because the cost is open-ended. If you’re starting a project, you won’t genuinely know if it will take five prompts or fifty.

Now, think about your average worker in a business. They’re going to be on a budget, their employer is going to monitor the cost of their usage, and mark my words, they will be comparing that usage to their peers. I can easily imagine this conversation behind closed doors:

Manager 1: Have you seen Joe’s token usage? It’s 3x the team average.
Manager 2: Whelp, time to let Joe go and replace him with someone more efficient.

I genuinely believe that if metered intelligence comes to pass, a talking point in job interviews will be how efficiently you use your tokens. Ironically, this will put a premium on expertise, because experts require fewer tokens to produce quality output.

And, what happens when companies set a hard cap on token spend? When an employee has a critical deadline to hit, and they’re about to hit their cap. Anyone with an ounce of empathy can imagine the anxiety of being cut off before the work is finished. You may as well put a Sword of Damocles over each worker.

The only party that benefits from this is the AI providers. People will have to exceed their budget, or not get their work done. And, if they succeed in penetrating education to produce a generation of AI-dependent workers, they can effectively charge whatever they want.

The Subsidized Era will End

This current phase, with its cheap subscriptions, generous limits, and friction-free access, is not sustainable. We’re in the land grab phase. It’s straight out of the playbook of ride-sharing companies like Uber: flout regulations, take massive losses to gain market share, and then jack up the prices once dependency settles in.

To be clear, this is a business strategy, and it has been shown to work, so while I despise what a lot of these AI leaders are doing, I acknowledge that they’re acting rationally in this case.

I see two very different outcomes ahead. One where AI can deliver value to people with expertise who have learned the discipline and judgment of when and how to use it, and another where it becomes an expensive source of frustration for people who don’t. And, a key point to keep in mind: AI companies don’t care if you’re inefficient with your tokens, because that makes you a more profitable customer.

The question is when, not if, the switch is made to only support metered usage; are they prepared for the backlash from users who spent years building the wrong work habits for a pay-per-use environment?

I very much doubt it.