My work is governed by a simple question: does a system strengthen decision making, preserve responsibility, and remain coherent over time? That standard applies equally to trading algorithms, community tools, and artificial intelligence. AI is not exempt from scrutiny. If anything, it requires greater scrutiny because so much of the public conversation surrounding it is driven by marketing, speculation, and misunderstanding.

I develop and publish AI frameworks for a very specific reason. I want to demonstrate where AI produces measurable value and where its use becomes socially dangerous. The value is real, but it exists within clear boundaries. AI can help organize information, identify patterns across large datasets, summarize content, and assist with communication tasks. These are environments where mistakes can be reviewed, corrected, and reversed before they create lasting harm.

Outside those boundaries, the situation changes dramatically. When AI systems are placed in positions where their failures can result in death, imprisonment, medical injury, or other irreversible consequences, the technology ceases to be a convenience and becomes a threat. This is not a hypothetical concern. The evidence already exists in court records, hospital reports, accident investigations, and the lives of people who have paid the price for misplaced trust in statistical systems.

Facial recognition technology is frequently presented as a tool for public safety, efficiency, and convenience. In practice, it is a system that converts human identity into a probability score. It does not recognize people in the way human beings do. It compares patterns within images and attempts to determine whether those patterns resemble images contained within a database.

The distinction matters because a probability score is not the same thing as certainty. Facial recognition systems are trained on datasets assembled from photographs, surveillance footage, social media images, and law enforcement records. Those datasets reflect the biases and imperfections of the societies that created them. When those systems are deployed at scale, the errors become predictable rather than exceptional.

The consequences have already been demonstrated. Robert Williams was arrested in Detroit after a facial recognition system incorrectly identified him as a suspect. Michael Oliver experienced a similar wrongful arrest in Michigan. Porcha Woodruff, who was pregnant at the time, was arrested after being falsely linked to a robbery through a facial recognition match that turned out to be incorrect. In each case, the technology produced an error, but the human cost extended far beyond a simple mistake in a database. People lost freedom, suffered emotional trauma, faced public humiliation, and endured disruptions to their personal and professional lives.

Supporters of facial recognition often respond by pointing to improvements in accuracy rates. That argument misses the central issue. Even a highly accurate system will generate large numbers of false positives when applied across millions of people. More importantly, the statistical accuracy of a system tells us nothing about the consequences of its failures. A false match in a music recommendation engine is an inconvenience. A false match in a criminal investigation can destroy a person's life.

The problem is not simply that facial recognition sometimes gets things wrong. The problem is that it is being applied in environments where getting things wrong carries consequences that no probability engine should be trusted to bear. When statistical pattern matching becomes a substitute for human judgment in matters involving liberty and justice, the result is not efficiency. It is the automation of error on a scale previously impossible.

The vision behind autonomous vehicles is compelling. If computers could drive more safely than humans, traffic fatalities could decline dramatically. Mobility could increase. Transportation could become more efficient and accessible.

The problem is that the reality of current systems falls far short of that vision.

Modern self driving systems rely heavily on pattern recognition. Cameras, sensors, and machine learning models attempt to identify objects and predict outcomes based on prior examples. This works well under many conditions. It works far less reliably when confronted with situations that differ from the data on which the system was trained.

The record of real world incidents demonstrates this limitation. Fatal crashes involving Tesla vehicles operating with Autopilot engaged have been the subject of repeated investigations. Walter Huang died when his vehicle struck a concrete barrier while the system was active. Jeremy Banner was killed when his vehicle failed to recognize a tractor trailer crossing its path. Investigators have also examined numerous incidents involving emergency vehicles parked on roadways with flashing lights, situations that should be obvious hazards but which repeatedly exposed weaknesses in the system's ability to interpret the environment.

These incidents reveal a fundamental issue. A machine learning system does not understand a construction zone, a traffic officer directing vehicles, or a child running into a street. It identifies patterns and generates predictions. Most of the time those predictions may be adequate. Occasionally they are not.

On a social media platform, an incorrect prediction might show a user an irrelevant advertisement. In a moving vehicle traveling at highway speeds, an incorrect prediction can kill someone.

Advocates often argue that autonomous systems should be evaluated against average human performance. Even if that comparison eventually becomes favorable, it does not eliminate the underlying concern. The public is being asked to share roads with systems that remain incapable of understanding context in the way human beings do. The risk associated with those limitations is not borne solely by the manufacturers or developers. It is borne by everyone who uses public roads.

Until these systems can demonstrate reliable performance across the full range of conditions encountered in real world driving, treating public roads as testing grounds for evolving AI systems is gross negligence and unacceptable.

Large language models are remarkably effective at generating human sounding text. They can summarize information, draft documents, answer questions, and create the impression of expertise across an enormous range of topics.

The appearance of expertise, however, is not expertise itself.

A language model does not understand medicine. It does not possess clinical experience, ethical responsibility, or diagnostic reasoning. It generates responses by predicting patterns in language based on information contained within its training data. The resulting output may sound authoritative even when it is incomplete, misleading, or entirely wrong.

This distinction becomes critical when people seek medical guidance.

Cases have already emerged in which individuals received harmful advice from AI systems. The widely discussed incident involving the National Eating Disorders Association chatbot demonstrated how an automated system could provide recommendations that were directly contrary to the needs of vulnerable individuals. Reports have also surfaced of users receiving dangerous health advice, including recommendations that delayed appropriate medical treatment or encouraged harmful actions.

The danger lies not only in factual inaccuracies but also in the confidence with which those inaccuracies are delivered. Human experts can acknowledge uncertainty, seek additional information, or refer a patient to another specialist. Language models often generate answers regardless of whether sufficient information exists to support them.

Medicine requires far more than information retrieval. It requires judgment, context, empathy, ethical obligations, and accountability. A physician must understand not only symptoms but also circumstances. The same diagnosis may require different treatment approaches depending on age, medical history, financial limitations, support systems, and countless other factors.

A language model has no genuine awareness of any of these realities. It can imitate the language of medical expertise, but imitation is not understanding. When that imitation is mistaken for professional judgment, the consequences can become severe.

The most important question is not whether an AI system is accurate ninety percent of the time, ninety nine percent of the time, or even more. The real question is what happens to the people caught inside the remaining failures.

When Robert Williams was arrested because a facial recognition system incorrectly identified him, he was not treated as a statistical error. He was treated as a criminal. Police arrived at his home, arrested him in front of his family, and placed him in jail for a crime he did not commit. The technology was wrong, but the burden of that mistake fell entirely on an innocent man.

The same pattern appeared in the cases of Michael Oliver and Porcha Woodruff. In each instance, an AI system pointed investigators toward the wrong person. Instead of treating the result as a tentative lead requiring rigorous verification, authorities treated it as evidence. Innocent people found themselves forced to defend their freedom against accusations generated by a machine. Their names, reputations, careers, and personal lives became collateral damage in a process that assumed the technology was more trustworthy than the human beings it accused.

These stories matter because they expose a deeper problem. When institutions place excessive confidence in AI systems, the people harmed by mistakes often begin from a position of presumed guilt. The machine produces a result, and the individual is expected to prove the machine wrong. That reversal of responsibility is profoundly dangerous. Technology should be required to prove its conclusions. Citizens should not be required to prove their innocence against an algorithm.

The same principle applies beyond law enforcement. Families who lose loved ones in crashes involving autonomous driving systems do not experience those deaths as technical malfunctions. Patients who receive harmful medical guidance from automated systems do not experience those injuries as statistical anomalies. They experience them as life changing events that cannot be undone.

What is often missing from discussions about AI is the recognition that every failure belongs to a real person. Behind every false arrest is someone whose freedom was taken away. Behind every fatal system error is a family that must live with the consequences forever. Behind every harmful medical recommendation is a patient who trusted information that appeared authoritative.

Supporters of these systems often point to investigations, settlements, policy changes, or public apologies after something goes wrong. None of those things undo the damage. An apology does not bring back a life that has been lost. An apology does not erase the trauma of being handcuffed, jailed, and publicly accused of a crime. An apology does not restore a damaged reputation, recover a lost job, repair a broken family, or return a home that was lost because a person's life was thrown into chaos. Once the harm has occurred, the victims are expected to move forward, but many cannot simply move forward. The consequences remain with them for years, and in some cases for the rest of their lives.

This is why the burden of proof must remain extraordinarily high whenever AI is introduced into environments involving liberty, health, or human life. The question is not whether the technology usually works. The question is whether society is willing to accept innocent people being arrested, injured, or killed when it does not.

The line itself is simple. AI has value when mistakes can be corrected before lasting harm occurs. It becomes unacceptable when its failures destroy lives, take away freedom, or cause irreversible injury.

A flawed summary can be rewritten. An incorrect recommendation can be ignored. A mistaken classification can be reviewed.

A wrongful arrest leaves scars long after charges are dropped. A fatal collision leaves an empty chair at a family table. A preventable medical injury can alter the course of a person's life forever. Even when authorities admit the mistake, the damage remains. Lost years cannot be returned. Lost opportunities cannot be recreated. Lost trust cannot simply be restored because someone issued a statement expressing regret.

That is the distinction that matters. The debate is not ultimately about technology. It is about people. Any system whose failure can cost a human being their life, liberty, health, or reputation deserves scrutiny measured not by benchmark scores or corporate promises, but by the human consequences of getting it wrong. The permanent nature of those consequences is precisely why apologies, explanations, and after the fact corrections are not enough. Once certain harms occur, there is no meaningful way to reverse them.

If you've read my About page, then you already know that I have been a programmer for more than forty-five years and have worked with machine learning and artificial intelligence in one form or another for over three decades. Throughout that time, I have consistently argued that this technology requires oversight, restraint, and careful implementation. That position has not changed.

I use AI every day. In many respects, it is a remarkable tool. It helps overcome blank page syndrome. It is excellent at finding typographical errors and dyslexic mistakes that I might miss because I am partially blind. It assists with research, organization, writing, image generation, music production, and a wide range of creative tasks. Used properly, it can save time and improve productivity.

The important phrase there is "used properly."

AI is a tool. Nothing more and nothing less. Every piece of work I publish that involves AI is reviewed repeatedly. I verify facts, inspect output, correct errors, and revise content over and over again. The entire process is built around the assumption that the tool will make mistakes and that those mistakes must be caught by a human being.

That is the point.

Whether I am producing music, creating videos, transliterating news articles, or generating images using my chalk on burlap technique, the work always contains the possibility of error. No matter how carefully I review the results, some mistakes will inevitably slip through. In my view, those mistakes serve an important purpose. They demonstrate both the value and the limitations of the technology.

Every mistake is a reminder that AI should never be trusted in a life-critical role. It should never be the final authority in any decision involving human life, liberty, health, or safety. The fact that a tool can be useful does not mean it is trustworthy. The fact that a tool can produce impressive results does not mean it understands what it is doing.

Some people will call my process slop, and that is perfectly acceptable. What I do does not put lives at risk. No one has died because of a flawed image, an imperfect news summary, or an error in a piece of music. The mistakes I make are visible. They are out in the open for anyone to examine, criticize, and evaluate.

In many ways, those mistakes are part of what I am trying to show. They expose the reality that AI systems are not infallible. They reveal limitations that marketing departments often prefer not to discuss. Every visible mistake is evidence that the technology remains imperfect, even in low-risk creative environments.

More importantly, every mistake you see should raise a simple question: how many mistakes do you not see?

If errors continue to appear in work that is reviewed, revised, and openly published, then what should we assume about systems operating behind closed doors, protected by confidentiality agreements, legal settlements, and corporate public relations departments? What confidence should we have in technologies deployed in vehicles, medical systems, surveillance platforms, or other environments where failure carries permanent consequences?

I can openly acknowledge my mistakes because they are visible to everyone. You can judge the quality of my work by what you see. You can examine the failures, understand why they happened, and decide for yourself whether the results are acceptable.

What you cannot do is accurately judge the safety of a technology when the evidence of failure is hidden from public view. When settlements, non-disclosure agreements, and private negotiations keep critical information out of sight, the public is asked to trust claims rather than evidence.

My mistakes are public. That transparency allows people to make informed judgments. The same standard should apply to every AI system that seeks a role in the real world, especially when the consequences of failure can never be undone.