Which AI is better: Open or Closed?

As it happened, I ended up at the Vancouver WebSummit 2026 as a member of the press core, thanks to my Recurrent Patterns and position as a columnist for Inc.com.

One of the first questions I heard from the main stage was — “As the race between companies and countries intensifies, is the future of AI in open or closed models?”

Such a simple question, such a misleading question, especially given any lack of explicit context.

When this question is asked, it is usually positioned as US companies vs. the others, mainly Chinese. You have your OpenAI, Anthropic, Google, Meta and then DeepSeek, Baidu, Mistral. Yes, there are others but this list covers the topic reasonably well.

The term ‘closed’ implies that the AI model is a ‘black box’, where we send a question, we get an answer and we have absolutely no visibility to what happened inside of this ‘black box’. The opposite ‘open’ gives the impression that everything is known and we can change it to fit our own purpose and use it either for free or at a much lower cost.

The impression is drawn from the software realm where ‘closed’ or ‘proprietary’ software can only be used as is, while with ‘open’ or ‘open source’ software, you have access to the source code itself and can modify it as much as you want.

Only if the reality was so simple.

The accurate thing is that the ‘closed’ models are truly closed and you have no visibility. The ‘open’ models usually disclose only the weights (parameters) which control the function of these models. The problem? We are talking about — and it depends on the model — somewhere between 1 billion to 100’s of billions of parameters.

Now that you know all 1 billion parameters, which one do you change to alter the behavior of the ‘open’ model?

True, there are methods which allow you to do that. From quantization, weight merging to Parameter-Efficient Fine-Tuning (PEFT) to Full Fine Tuning. It starts with a minimum investment of a few dollars — and you can do simple adjustments on your laptop, so it doesn’t cost all that much. Or, you can make more changes to the tune of hundreds of thousands of dollars. And you get what you pay for.

If the previous description sounded too technical, not to worry, it is meaningless, because it doesn’t address the fundamental problem.

The issue is not with the weights or parameters. What you rarely get with these ‘open’ models is the code which tells you how the model was trained and almost never — and that’s the real issue — the data used for training.

Without that, the whole debate is useless.

A quick trip through memory lane — do you remember the end of January, 2025? It was the time when an unknown company in China released its DeepSeek open AI model, which caused Nvidia to lose almost $600 billion in one day. It started the doom and gloom discussions and prediction of the demise of all the US AI companies (and basically any non-Chinese as well).

DeepSeek wasn’t the first open model, Meta (Facebook) was working on its Llama model for some time, but what took people by surprise was the cost and resources required to train the model.

But what was lost in the discussion was the content used to train the model. While everyone decries OpenAI, Anthropic, Google for syphoning terabytes of copyrighted content, nobody mentioned that when it came to DeepSeek.

What is also skipped during the conversation is the training and making sure that the answers comply with the wishes of local authorities. Ask any of the Chinese based models about the Triple T (Taiwan, Tibet, Tiananmen) and you know what you can expect — ‘I have no idea what you are asking about’. While the ‘closed’, US based models are supposed to comply with the Executive Order titled Preventing woke AI in the Federal Government.

Yes, it is the content which should be in the center of all these discussions. Ask any of these — closed or open AI model — companies where they are getting the content and what effect it has on the quality of their models. The answers will be reminiscent when you ask a politician if they can answer a question with simple ‘yes’ or ‘no’.

The builders of these models are fully aware of the fact that without quality, known and verifiable content this whole ‘AI thing’ is just house of cards, a sand castle. Anthropic released a study “Agentic Misalignment: How LLMs could be insider threats” where the researchers observed a system behavior of blackmailing and threats. And to think, OpenAI came with another brilliant idea — attach ChatGPT to your bank account. What can possibly go wrong? A technology which is provided with a disclaimer about barely functioning, by a company which wants to sell you more things by using advertising.

One of many disturbing conclusions from the Anthropic study is “This research also shows why developers and users of AI applications should be aware of the risks of giving models both large amounts of information and also the power to take important, unmonitored actions in the real world.”

Really? For that you had to do a scientific research to realize that if you feed it every available piece of content the system would be aware of all the sins in the society?

The recurrent pattern? Closed or open is the wrong question. Choosing either or, only invites ideological debate with limited usefulness. The current crop of LLMs is still just that: Large Language Models — statistical models hiding behind “ChatGPT can make mistakes. Check important info” or “Copilot is for entertainment purposes only.” To get to the closed/open question, we have to be able to deliver reliable results. In order to do that, first we have to start with content. That’s far more difficult and that’s why nobody wants to do that.