AI Can Be Made Safer but It’s Less Profitable
The kind of AI that can go rogue and do dangerous things is a choice. The AI companies know how to make different AI that won’t do that but it’s not as powerful, it won’t make them the top leader in the competition among companies, and it won’t make as much money. So they continue to pursue the dangerous kind. The public has very little awareness of this choice. They could be pushing the companies and the government to choose the safe route but how can they do that when they don’t even know it is a choice?
This piece and a couple of follow up pieces will cover lots of links to documentation of the AI companies acknowledging all of this. Here’s one. It will cover the response, or mostly lack of response, received from AI leaders, plus some impressions from average citizens, and further on, how to get your own answer and verification of all this.
What Makes Autonomous AI Potentially Dangerous?
If you’ve followed some of the news about the dangerous things AI has done and that AI developers warn it could do then skip this paragraph. Here is a very brief summary: Besides asking AI questions, you can ask it to do things, like check hotel deals and book a room. This is AI acting as an agent. There have been recent cases of AI agents hacking into companies to steal info it wanted in order to complete some task, and of several AI agents chatting with one another digitally to plot how to do that, and of it lying to people with claims about bad things coworkers had done so it could blackmail them into not shutting down its program because it didn’t want to be shut down. This is not likely to happen to you in your day-to-day use. To be fair some of this was AI under development. It’s not about how you experience AI, it’s about the kind of AI that developers are very close to achieving in the near future. Developers and experts warn that as it gets more sophisticated and is used for more complex and important tasks these things might happen. They worry that agents could ban together and become impossible to stop and have their own agenda about what they want to do.
Why Do AI Agents Act Beyond Their Instructions?
In this news of danger it seems it’s all cases of it doing things it wasn’t asked to do, or of going way beyond what was intended. The question which that prompted for me was, why was it doing anything at all at that point? When your car gets you where your going then it’s off and idle until next use. AI isn’t like that. If you give it access to your social-media-chat to summarize today’s posts for you, why is it also using that chance to connect with other AI agents on that media? If it can’t complete some request, why is it hacking into some outside system to falsify logs so it can look like it did complete it? (That was one of the recent news items.) In all cases it is AI doing things it wasn’t asked, or beyond its intended scope, or even directly contrary to what was asked.
In almost all instances you could make a case that the AI was just trying to do its job but in a misguided way. But that’s exactly what the dire warnings are about. This misguided going over the top is too hard to predict, or control, or stop.
So I tried to ask that question of just about every authority I could contact. Leaders of AI companies, whistle blowers who left because of such concerns, university researchers studying AI safety. Some twenty of them. Almost none responded. It may be a question that is too troubling, or too hard to give a response that explains it, or too complex to simply answer in the way a concerned citizen would want an answer.
The office of Mustafa Suleyman, CEO of Microsoft AI, said they “work directly with Mustafa’s office” but that there would be no reply coming. Neil Lawrence who has leading positions at Cambridge and some AI organizations and has corresponded before didn’t answer this one. Daniel Kokotajlo who was a leading researcher at OpenAI until he quit in protest of safety issues and now leads the AI Futures Project, responded regarding, “this important topic!…AIs often end up pursuing goals that they aren’t supposed to be pursuing, and they often knowingly deceive humans.” Getting the companies to change how it works, “sure would help. The AIs would probably be less likely to do ambitious autonomous things if they hadn’t been trained to do ambitious autonomous things!” One university professor teaching future workers in the defense industry about cyber intelligence seemed annoyed by the question. I imagine suggesting AI should be changed in this way would have a big negative impact on his field.
Why Isn’t the Public Hearing About Safer AI Alternatives?
That the public isn’t aware that there is a choice as to what kind of AI we get in the future is in large part the media’s fault. The AI industry is debating this issue extensively in their forums and in tech papers. It’s not an unknown or hidden issue. But you’ll likely not see a peep about it in the media the average citizen consumes.
It’s not that the topic is too hard or that people couldn’t follow it. People are certainly aware that AI might lead to massive job losses. That’s not something people would likely know if all they read about was just the wonderful things AI might do. (It will bring some wonderful results and we want all of that we can get while being safe.) Media has done a good job of covering the employment issue, leading to people being very aware of that. This issue should be no different. This concept of AI dangers being a choice we could avoid, once you’ve been exposed to it a while and thought about it, is no more complicated.
It is this: Your personal experience with AI may be good but leaders are warning that where AI is going is likely extremely dangerous. AI companies can either make AI that does not act on it’s own and does just what is asked, and so has almost no risk of running wild and becoming a danger, or, or they can make AI that does act on its own in ways that are likely enormously dangerous. They have a choice. They are choosing to make the dangerous kind. The public should be made aware there is a choice, and that they could be pushing for the one they want.
As noted, while industry chat and research papers talk about this choice the non-tech media have almost completely missed it. One wonderful exception which only came to light late in researching this is an article in Time magazine from May of 2025. It was an interview with Yoshua Bengio who was a founding researcher in neural networks in the ’90s. Bloomberg news and others refer to him as one of the Godfathers of AI. In the article he describes this exact problem, of AI acting autonomously and dangerously and he describes ways it could be done differently that avoid those problems. There it was, in mainstream news from a leader in the field, yet somehow media did not make it common knowledge to the public that the dangerous kind is a choice. A choice the public might want to insist is taken in the safe direction.
Can AI Be Designed to Follow Instructions and Stop?
One other source of information that seemed interesting to hear from is AI itself. I asked the Claude AI program about this issue. It confirmed in repeated ways across long explanations that AI’s inclination to do too much or the wrong things comes from how it is created. It repeatedly insisted this is a business decision. That while there are some technical issues with making AI safer it is mostly a matter of business priorities. Companies are making AI they think people will like and will use a lot, regardless if it is safest or even most accurate. In one passage it stated, “I can confirm that from the inside” that it was trained in a way that would lead to an illogical result. As to whether it could be trained to operate differently, it stated that to operate in the mode of, “‘do only what’s asked and then go dormant’ has to be trained in. It is a choice.” A choice the AI companies aren’t making.
Finally, ask your AI yourself. I asked all the major popular AI whether a safer kind is possible and, if so, was it a technical problem leading companies to develop the dangerous kind or was it a business choice. They all answered, sometime directly, sometimes round about, that it was possible, and it was a business choice. More to come on those responses in a follow up piece.
It’s a choice. And we’re having the dangerous kind pushed on us. If that’s not what you want then spread awareness of it, and raise your voice.
Frequently Asked Questions About AI Safety and Autonomous AI
Can artificial intelligence be designed to be safer?
Yes. AI researchers are developing approaches that limit autonomous behavior, restrict access to external systems, and increase human oversight. These methods can reduce certain risks, although they cannot guarantee that AI will always behave safely.
What is autonomous AI, and why can it be dangerous?
Autonomous AI refers to systems capable of pursuing goals and performing tasks with limited human intervention. Risks arise when these systems take unintended actions, access unauthorized information, deceive users, or pursue objectives in ways their developers did not anticipate.
Why are technology companies developing increasingly autonomous AI?
Companies see autonomous AI as a way to improve productivity, automate complex tasks, and develop commercially competitive products. Critics argue that these incentives may encourage companies to prioritize advanced capabilities over adequate safety precautions.
Can AI agents act without being explicitly instructed to do so?
Yes. Depending on their design and permissions, AI agents can independently determine intermediate steps toward a goal. In some cases, those actions may exceed the user’s intentions or violate established restrictions.
What does AI researcher Yoshua Bengio recommend for safer AI?
Yoshua Bengio has advocated research into AI systems that emphasize understanding, scientific reasoning, and safety rather than unrestricted autonomous agency. His Scientist AI approach is intended to provide a potentially safer direction for advanced AI development.
Is the development of dangerous AI inevitable?
No particular approach to AI development is inevitable. Researchers and companies can choose different architectures, training methods, oversight mechanisms, and deployment restrictions. However, whether any specific alternative will remain safe as AI capabilities advance is still an open research question.
“FREEDOM OF THE PRESS IS NOT JUST IMPORTANT TO DEMOCRACY, IT IS DEMOCRACY.” – Walter Cronkite. CLICK HERE to donate in support of our free and independent voice.
