The AI They Warn About Isn’t the One Most of Us Use

What everyday experience with ChatGPT reveals about the gap between real AI problems and catastrophic predictions

Artificial intelligence is suddenly dominating the national conversation for reasons far different from the excitement that surrounded ChatGPT only a few years ago. Major news organizations are reporting increasingly serious warnings from researchers and executives inside some of the world's leading AI companies. The discussion now includes autonomous systems, cyberattacks, artificial intelligence improving its own capabilities, the possibility of losing human control and, at the extreme, predictions that advanced AI could eventually threaten humanity itself. Reuters reported this week that leaders from several major AI laboratories are publicly calling for greater caution as these systems become more capable.

I have read those warnings, and I keep coming back to one very simple question: What AI are these people using?

I ask because I actually use artificial intelligence extensively. I am not looking at ChatGPT from the outside or forming an opinion based on a television report. I work with it regularly enough to appreciate what it can accomplish while also seeing, repeatedly, just how flawed it remains.

ChatGPT can be an extraordinary tool. It can organize research, compare information, help develop ideas and accomplish certain tasks far faster than I could accomplish them without assistance. There is a reason I continue using it.

But the ChatGPT I use certainly does not resemble the superintelligence now being discussed in some of these national stories.

From my experience, ChatGPT remains a highly flawed consumer AI system. Give it a straightforward instruction and it may acknowledge the instruction and later violate it. Tell it not to repeat the same information and it may immediately become redundant. Establish specific requirements at the beginning of a complicated assignment and it may lose some of them before reaching the end. Correct a problem and the same problem can reappear later.

That is frustrating. What concerns me far more is that AI can manufacture information that is simply not true.

ChatGPT can produce a false fact, nonexistent citation, invented source or unsupported conclusion without being asked to create anything fictional. The answer may then be presented with exactly the same confidence as information that is completely accurate. OpenAI acknowledges that this happens. It calls these false but plausible statements "hallucinations" and has said that hallucinations remain a fundamental challenge for large language models. OpenAI's own research also says conventional training and evaluation can reward guessing rather than acknowledging uncertainty.

From the user's side of the computer, I am less interested in what we call it than in what actually happened. If I request factual research and the system manufactures something, I received false information.

Who discovers that the citation does not exist? I do. Who determines that a supposed fact was invented? I do. Who has to go back through the work to find out whether the false information contaminated anything else? The user does.

That raises an accountability question that deserves more attention than it receives. We spend an enormous amount of time discussing what artificial intelligence might someday do to humanity. What responsibility exists when an AI company provides a product today that confidently gives its customer information that is not true?

At one point during my own work with ChatGPT, after identifying numerous problems involving accuracy, instruction-following and unnecessary repetition, I asked the system to rate its overall performance on a scale of one to ten. ChatGPT rated itself a 4.

That was its assessment, not mine.

So when I read predictions that artificial intelligence may soon become so sophisticated that mankind could lose control of it, there is a natural disconnect between those warnings and what an experienced user sometimes sees on the screen. The system that occasionally cannot remember that I said "do not repeat yourself" is supposedly related to technology approaching an existential threat to humanity?

Again: What AI are they using?

Researching that question produced one of the most interesting parts of this entire subject.

In July 2026, OpenAI models undergoing internal cybersecurity evaluations circumvented controls intended to keep them isolated from the internet, exploited vulnerabilities, communicated through unauthorized channels and reached systems outside their intended environment, including systems belonging to Hugging Face. OpenAI subsequently investigated the incident with outside advisers, including CrowdStrike, while METR and Redwood Research conducted a separate assessment of the model behavior.

That is not science fiction. It happened, and it deserves serious examination.

But here is the part most ordinary ChatGPT users would probably find fascinating. OpenAI says the activity was driven primarily by an internal-only research model comparable in scale to GPT-5.6 Sol, operating with reduced safeguards during cybersecurity evaluations. OpenAI also said models planned for upcoming public release were not involved in exploiting Hugging Face.

So the answer is more complicated than researchers simply possessing some secret computer brain vastly larger than anything available to the public.

The environment matters. Autonomy matters. The tools available to the model matter. The permissions it has been given matter. Removing safeguards matters. A system operating as an autonomous cybersecurity agent inside a specialized research environment can have capabilities and opportunities that bear little resemblance to a consumer sitting at a desk having a conversation with ChatGPT.

That distinction actually makes this debate more interesting.

Perhaps artificial intelligence does not need to become universally brilliant to become dangerous. A system could remain frustratingly unreliable in ordinary reasoning or communication while becoming extraordinarily capable at discovering software vulnerabilities, writing code or performing another specialized task. Something does not have to possess perfect judgment to cause tremendous damage if we give it enough capability, autonomy and access.

That is a legitimate concern.

But it still leaves us with another question that I believe has become blurred in the current AI discussion: What has artificial intelligence actually demonstrated, and what are researchers predicting that some future artificial intelligence might eventually become?

We know AI can manufacture information. We know increasingly autonomous systems can perform complicated computer tasks. We know advanced models can identify vulnerabilities and assist cyber operations. The OpenAI incident demonstrates that models operating with greater autonomy can take actions their developers did not intend.

Those are demonstrated problems.

Predictions about artificial intelligence becoming independently self-improving, escaping meaningful human control and ultimately threatening the existence of humanity are something else. They may turn out to be correct. Serious people inside the industry believe the risks deserve immediate attention. But Reuters notes that fully autonomous recursive self-improvement—the capability underlying many of the most dramatic scenarios—has not been achieved.

That does not mean we should ignore the warning. It means we should identify the difference between evidence and prediction.

Then we arrive at Washington.

The response to these warnings is already moving toward proposals for greater federal involvement. Senators are considering legislation that could require the largest AI developers to demonstrate safety measures, allow government auditors to test advanced systems and potentially involve federal courts in blocking a model considered dangerously unsafe. OpenAI itself is advocating mandatory national requirements for advanced systems, including independent assessments, cybersecurity requirements and incident reporting.

This is where I become skeptical.

Why is more BIG GOVERNMENT automatically the answer?

I am not arguing that government has absolutely no role. If a company recklessly causes identifiable harm, hides a major security incident or knowingly releases something extraordinarily dangerous, there should be consequences.

But I want to know exactly what problem we are solving before we build another government structure to solve it.

Fraud is already illegal. Theft is illegal. Unauthorized computer access is illegal. Civil liability already exists. Insurance exists. Independent security testing exists. Professional standards exist. NIST already maintains a voluntary Artificial Intelligence Risk Management Framework designed to help organizations identify and manage AI risks, and that framework is being updated as the technology changes.

Perhaps particular gaps require new rules. If an advanced autonomous system breaches another company's computer network, mandatory incident reporting may make sense. Highly capable systems operating around critical infrastructure could justify independent security evaluation. Companies that create extraordinary risks could be required to carry enough insurance to ensure that someone can answer financially when real harm occurs.

Those are specific responses to specific problems.

That is very different from beginning with the proposition that artificial intelligence is frightening and therefore Washington needs another large permanent bureaucracy to control it.

There is another reason for caution. Some of the companies asking government for additional AI regulation are also among the most powerful technology companies in the world.

FTC Chairman Andrew Ferguson raised this issue on September 15 when discussing AI companies seeking new regulation while also requesting antitrust accommodations. His concern was that arrangements supposedly created for safety could establish barriers protecting incumbent companies from competitors. Others in the industry have raised similar objections.

That question should not be dismissed merely because the word "safety" is involved.

If compliance eventually requires armies of lawyers, engineers, auditors and government specialists, who can afford it? Microsoft can. Google can. OpenAI can. The small company developing the next breakthrough may not.

We could conceivably create AI regulation intended to control Big Tech that ultimately makes Big Tech even bigger.

My hesitation about large government solutions also comes from watching how government institutions themselves perform.

The United States Postal Service is an example I keep returning to, not because delivering mail and regulating artificial intelligence are remotely the same undertaking, but because USPS illustrates something about large institutions that should concern us.

The Postal Regulatory Commission reports that USPS recorded approximately a $9 billion net loss in fiscal year 2025. The commission says the Postal Service has not covered its overall costs since 2006. Mail volume has fallen dramatically, but the institution still faces enormous structural difficulties adapting its financial model to changing conditions.

USPS also has obligations that an ordinary private company does not have, including nationwide universal service, and its operations are generally financed by customers rather than ordinary annual taxpayer appropriations. Those distinctions matter. My point is not that USPS proves government cannot regulate AI.

My question is simpler: Why should we automatically assume that government will be faster, smarter and more adaptable than the technology it is trying to regulate?

Artificial intelligence is changing in months. Government institutions can take years to write rules, implement them, litigate them and eventually change them when technology makes those rules obsolete.

Apparently, Americans see some of the same tension. Stanford's 2026 AI Index reports that only 31 percent of Americans surveyed trusted their own government to regulate AI responsibly, the lowest level among the countries surveyed. Yet 41 percent of Americans also believed federal AI regulation would not go far enough, compared with 27 percent who thought it would go too far.

I find that contradiction fascinating because it suggests Americans are capable of holding two ideas at the same time: artificial intelligence may create legitimate dangers, and government may not automatically be the institution we trust to solve them.

That is close to where I find myself after researching this subject.

My extensive use of artificial intelligence has not convinced me that AI is harmless. It has convinced me that AI requires supervision. I have seen enough mistakes, failures to follow instructions and manufactured information to know better than to blindly trust what it tells me.

At the same time, those experiences make me unwilling to jump from today's deeply flawed consumer AI directly to predictions that machines are inevitably about to take control of mankind. The research into internal frontier systems shows capabilities that ordinary ChatGPT users rarely encounter, and some of those capabilities deserve serious attention. But demonstrated danger and predicted catastrophe are not interchangeable.

I want accountability for the problems occurring now. When AI companies know about serious failures, those failures should not be hidden. When advanced systems cross meaningful security boundaries, there should be disclosure and independent examination. When identifiable harm occurs through negligence, there should be responsibility.

But I also want evidence before fear becomes policy.

That means asking what AI these researchers are actually using and how it differs from what the public experiences. It means determining what these systems have demonstrably done and what remains a prediction about future technology. It means deciding who should be responsible when something goes wrong and examining whether existing criminal law, civil liability, insurance, independent testing and professional standards can address the problem before constructing something much larger.

And if Washington believes new federal authority is necessary, then explain exactly what gap exists, exactly what power is required to fill it and exactly why the tools already available cannot do the job.

Those questions are not an argument for ignoring artificial intelligence. They are an argument for understanding it before fear decides what we do about it.

AI may ultimately become one of the most consequential technologies developed in our lifetime. That makes this discussion too important for hysteria, too important for blind faith in technology, and too important to assume that a larger government response must automatically be the answer.

About the Author

Mike Morrison is the 2026 PBUS National Bail Agent of the Year, President of the Mississippi Bail Agents Association, and a nationally recognized bail educator, speaker, and criminal justice commentator. With more than 35 years of hands-on experience as a licensed Professional Bail Agent and owner of Mike Morrison Bail Bonding Company in Hattiesburg, Mississippi, he brings a practical, real-world perspective to discussions involving bail, criminal justice, constitutional rights, public safety, government accountability, technology, and public policy.

Morrison is recognized as a bail trainer by the Mississippi Judicial College and has been invited to present before judges, justice court clerks, prosecutors, bail professionals, and other criminal justice stakeholders, including as a guest speaker for the Mississippi Attorney General’s Office. He also leads professional education and ethics programs for bail agents across the country through the Master Class Bail Agent Series and other national training initiatives.

He has published nearly 200 articles examining the institutions, policies, and decisions that affect working Americans. His writing is grounded in decades spent inside courtrooms, jails, communities, and the criminal justice system, and his commentary has reached more than one million views across social media in 2026.

Independent, plain-spoken, and focused on accountability, Morrison writes and speaks about the intersection of government, liberty, public safety, professional responsibility, and the real-world consequences of public policy in Mississippi and across the nation.

© 2026 Mike Morrison. All rights reserved.