The AI Industry’s ‘Good Enough’ Problem

Back in December 2022, I wrote a post explaining that the real danger of AI was not that it would write better posts or create better images, but that it would be “good enough” for most to not care.

In that post, I likened AI content to a cell‑phone camera. Cell phone cameras are now ubiquitous and are “good enough” for most people. Situations where, decades ago, people would have hired a professional to take images are now routinely covered by cell‑phone cameras.

Could a professional photographer do a better job? Absolutely. But cell phone cameras are good enough for most people in most circumstances.

Nearly four years later, that article has largely come to fruition. We are in a world flooded with AI-generated content including text, images, songs and videos. Though most people agree that AI content is worse than human-created content, AI music is charting, AI videos are taking over YouTube and social media, and even mainstream media outlets, such as Sports Illustrated and CNet have been publishing AI articles.

AI still can’t beat human quality. But, for its supporters, it’s “good enough.” That is, and always has been, where AI is most dangerous.

However, now the tables have turned. The AI industry is getting a taste of what it’s like to be on the wrong end of the “good enough” argument. As they seek to justify billions of dollars in investment, they are repeatedly being undercut by other AI models and approaches that, while not as good as their frontier models, are “good enough” for most people.

The Chinese Model Problem

The first cracks began to show in early 2025. That was when DeepSeek published its R1 model, which was allegedly trained for a fraction of the cost of comparable AI models. OpenAI (and others) accused DeepSeek of distilling their models to create DeepSeek.

This was met with some extreme schadenfreude from the internet. OpenAI was accusing DeepSeek of doing to them what they had done to human creators, taking their output and training a new model on it. Simply put, there was not a lot of sympathy to be found for OpenAI.

Depending on the comparison, DeepSeek’s tokens cost a small fraction of OpenAI’s tokens, especially when you compare higher-end models.

However, it’s not just DeepSeek. An advisory from U.S. cybersecurity and intelligence agencies claims that Moonshot AI, Alibaba, MiniMax, StepFun, and Z.AI have all done something similar.

To be clear, distillation is only one possible reason for the models to be cheaper. Chinese companies have made legitimate innovations in the field of AI that have also reduced costs. Still, distillation does drastically reduce the cost of training an AI model, even if it limits the model’s capabilities.

If you look at LLM leaderboards, you’ll see that they are still dominated by US companies, mostly OpenAI and Anthropic. However, as of this writing, several Chinese companies have models that are competitive with the top-tier models and are only a fraction of the cost.

However, it’s not just the cost that should have the major AI companies worried. Many of those aforementioned models are open-weight, meaning that anyone can download and use them for free.

That, in turn, may be the biggest long-term threat to the AI industry.

Rolling Your Own AI

The relationship between lawyers and AI systems has been extremely fraught. We’ve seen countless lawyers get citations for AI hallucinations in their filings, including one case where both sides were sanctioned. We’ve also seen an expert witness tank a case because of their reliance on AI.

In spite of (or perhaps because of) these issues, law firms have been doubling down on AI, but not in the same way.

Morgan & Morgan, the nation’s largest plaintiff law firm, announced that it is spending $1 billion over the next 10 years to build its own in-house AI system dubbed MX2. The system is already being used by over 5,000 people, and performs tasks such as writing first drafts, extracting medical records and searching for relevant cases.

To make matters worse for the AI industry, Morgan & Morgan is planning to offer MX2 as a service to other law firms by the end of 2027. Though we don’t know what models MX2 is based on, it’s likely that it’s an example of an open-weight model creating competition for the existing AI industry.

Kirkland & Ellis previously announced that they are doing something similar, spending $500 million to build its own, internal, AI platform. However, they have no plans to offer it as a service.

Both of these announcements are a major blow to Harvey and Legora, the two major players in the legal AI space. While both companies do have some in-house models, both rely on OpenAI and/or Anthropic for their services. Lawyers, concerned about costs, loss of intellectual property and other issues have been increasingly looking at bespoke AI solutions.

Even if these systems aren’t as good on paper, between the cost savings and the other concerns, it’s easy to see why a law firm would consider building its own AI system, especially if it provides similar results.

But it’s not just large law firms, or even large companies, that are turning to in‑house AI. Depending on what you want to do with AI, there’s a good chance that you can simply use the device that you have.

Macs, for example, ship with Apple Intelligence. It can edit text, summarize documents and generate images on the device itself. If you want to go further, there are tons of smaller AI models that can be downloaded and run on your device.

If you’re willing to take the step of setting up a home server, you can go even further, running larger models that are more capable. You can also run a VPS (virtual private server) and use a free AI model, paying only for the hosting costs.

Bottom Line

AI costs are on the rise and, in many cases, AI is costing more than the people it was supposed to replace. As I see it, this creates a “K” shaped trajectory for the AI industry.

Those who use AI minimally or don’t use it at all will simply have their needs met by whatever is on their device. In fact, they will most likely not be aware that they are using something that others consider to be AI. To them, it will just be grammar checking, notification summarization, etc. They will never pay a cent for AI.

On the other extreme, large corporations have every incentive to invest in their own AI infrastructure if they are going to use AI. Even if their models aren’t as good as the frontier models on paper, the cost savings, IP concerns, and other issues make it a natural choice for large companies.

Even worse for the AI industry, they could follow the example of Morgan & Morgan and create a competitor to other AI companies.

So where does this leave OpenAI, Anthropic and the other large AI companies? A big part of their proposition to investors was that they would be providing AI as a service to just about everyone. But now most have an alternative that is “good enough’ for them that is either significantly cheaper or free.

To be clear, this is not a new idea. In May 2023, a leaked Google memo discussed how the company has no “moat” and neither does any of their competitors. Open alternatives can and do catch up to the closed models within months or even weeks.

But the bigger problem is that a slightly less capable model can be “good enough” for most uses. Why should someone spend an ever-rising amount of money to stay on the cutting edge when their needs and wants can be met by something that is just a few months behind?

In the end, I find it strangely fitting that the AI industry polluted the internet (and the world) with the promise of fast and easy content that was “good enough” for the purpose. Now, it may be undone by that same mentality and that same promise, this time from cheaper and open alternatives.

It doesn’t do anything to fix the damage that AI has already done. But it is a cold comfort in a time where there are few comforts to be had.