The Future of Metro Detroit Auto Care: High Tech Tools, Human Hands

Social media is rife with sensationalist claims regarding the end of work, the irrevocable transformation of white-collar jobs, and a total upheaval of the social contract. The scale of artificial intelligence and its ability to process language is undoubtedly impressive—so much so that even its creators seem entranced by their own creations. They treat these models like Dr. Frankenstein regarded his monster or as a proverbial Golden Calf.

However, a more nuanced approach is required. By detailing the underlying mathematics of these AI models, I will demonstrate why current physics (and meta-physics) does not support the notion that these systems can match or exceed human intelligence. After critiquing the prevailing doomerism surrounding their use, I will outline what the automotive aftermarket should actually expect from this technology and how we plan to integrate these tools into our own products and services.

The Engine of the LLM: A Mechanical Metaphor

To understand Large Language Models (LLMs), it helps to view them through the lens of internal combustion. This comparison clarifies how data is processed and where the limitations of the output begin.

The Air (Input/Data)

"Big" datasets act as the raw intake (e.g The Internet, YoutTube, Spotify). This includes internet sources, books, audio files, and video content. Without a massive volume of generated data, the engine has nothing to compress. It is important to note that these models are not passively synthesizing data from the physical world like a human. We are the source of the oxygen; the AI is merely a vacuum cleaner in a closed room. If humans stop producing original observations, the engine eventually runs out of air. LLMs are not embodied in the physical world. They do not sense passively, or even actively, the same data humans have access to in a physical reality. Not even humans understand everything they are sensing all at once.

The Calibration (Parameters/Weights)

While many view parameters as the fuel, they are more accurately described as the engine mapping or compression ratio. Trillions of parameters act as the static instructions, the valve timing and fuel trims, that allow the model to find patterns within existing data. The training process, which necessitates massive data centers, creates these rules from the digital exhaust of human history.

A key reason the outputs seem creative or intelligent is that the average human cannot conceptualize the scale at which these rules are being computed. It is a brute force illusion. The model is not smart; it is simply performing a trillion lookup operations faster than humans can blink. Furthermore, these models are recombinant. They do not create ex nihilo; they recombine old patterns into a new statistical average. This is why models can experience .model collapse if they are retrained on their own synthetic outputs. Without a fresh intake of human primary observation, the internal logic becomes incestuous and eventually degrades into nonsense.

The Combustion (Inference/Probability)

This is the spark. The model does not think; it calculates the probability of the next token based on the prompt. It is a mathematical explosion that transforms data into a response. This probability is not a single word choice but a categorical distribution—a massive map of every possible word the model knows, weighted by likelihood.

Software can be tuned to choose words with more likely, less likely, or average probabilities. However, a human must perform the initial dyno tune through a process called reinforcement learning from human feedback (RLHF). Humans manually rank responses to ensure the model remains coherent. This human driven calibration creates a statistical gravity that pulls the model toward the center of the road. And, it needs to be mentioned that the human tuners are not arbiters of truth! They are doing their best, but it isn't as if a majority of users and tuners are going to be capable of representing all cultures, knowledge bases, and experiences. Because the engine is tuned to be safe and reliable, the responses are seemingly varied but average in quality, lacking the leaps of true human innovation.

Unintended Emissions (Output)

This is the generated text, pixel, or audio. Like any engine, it produces emissions in the form of hallucinations, errors, or content that lacks original insight. These are not failures of the system but byproducts of its stochastic nature. It is a probabilistic guess that, without a physical anchor in reality, can never be guaranteed as ground truth.

A large language model likened to an engine: Intake (air) with text sources; Architecture (fuel) with transformer layers; Processor (combustion) inferring tokens; Exhaust (output) shows emissions like errors.

Reductio ad Absurdum: Human Cognition and the 'Universal Learning Function'

AI accelerationists are making a knowingly or unknowingly absurd argument: that human intelligence is merely a mathematical function. This worldview asserts that there is nothing unique about human cognition and that it can be fully replicated through human generated rulesets and massive compute. It is a reductive perspective which subversively posits the LLM as a better or more evolved form of intelligence.

The flaw in this logic is a fundamental category error. Mathematics is a language we developed to aptly describe behavior and predicted function in the physical world; it is not the nature of the thing itself. We can use calculus to model the arc of a baseball, but the baseball does not calculate its trajectory. By claiming the mind is a universal learning function, accelerationists are mistaking our descriptive tools for the underlying biological reality.

The Chinese Room Fallacy

To understand why statistical processing is not the same as understanding, we must look to the philosopher John Searle and his 1980 thought experiment known as the Chinese Room.

Imagine a person who knows no Chinese is locked in a room with a massive book of instructions. Outside the room, native Chinese speakers slide slips of paper with Chinese characters under the door. The person inside uses the instruction book to find the correct symbols to slide back out. To the observers outside, it appears the person in the room understands Chinese perfectly. In reality, the person is simply following a ruleset for symbol manipulation without any knowledge of what the symbols actually mean.

This is the exact architecture of a Large Language Model. The model is the room, the trillions of parameters are the instruction book, and the generated text is the slip of paper. The system is manipulating the syntax of human language with staggering efficiency, but it possesses zero semantic understanding. To claim that a statistical prediction of the next word is equivalent to human thought is to mistake the menu for the meal.

The Systems Reply

In the decades since Searle first proposed this experiment, proponents of artificial intelligence have offered several counter arguments, the most prominent being the systems reply. This argument suggests that while the individual person in the room does not understand Chinese, the system as a whole—the person, the room, the instruction book, and the data—collectively achieves understanding.

From this perspective, intelligence is an emergent property of the entire stack. Proponents argue that just as a single neuron in the human brain does not understand English while the entire brain does, a single processor in a data center does not understand a prompt while the entire network does.

However, this still fails to account for the primary observer problem. In a biological system, the system is anchored to a physical body with metabolic needs and sensory feedback. In a silicon system, the understanding remains a closed loop of symbol manipulation. Whether the understanding is located in a single component or the entire system, it remains a secondary synthesis of human derived records rather than a primary observation of the physical world.

The Limits of the Artificial

Our current ignorance regarding the intersection of biology, consciousness, and the physical world precludes us from accurately defining or replicating the very intelligence that created the LLM in the first place. We are attempting to build an artificial representation of a system we do not yet understand, treating a high-fidelity mimic as a peer.

This isn't an evolution of life; it is the ultimate safe tune. It is a system that looks like it is winning the race but is actually just a recording of the leader. While the recording can be played back at incredible speeds and volumes, it lacks the displacement, the heat, and the physical agency required to drive the car.

Physics and the Efficiency Paradox

There is a staggering disparity between biological and silicon intelligence that does not comport with our understanding of physical efficiency. This gap is most evident when comparing the energy consumption required to reach a specific outcome. A massive, global value chain is required to refine the hardware, generate the megawatts for compute, and pump millions of gallons of water to cool the servers.

The 20-Watt Human vs. The Megawatt Model

The human brain operates on roughly 20 watts of power—about the same as a dim incandescent lightbulb. On this meager energy budget, humans created the very AI that now requires an entire power grid to run. In contrast, training and operating frontier AI models requires specialized data centers and hundreds of megawatts of power. If a system requires a dedicated power plant to answer a question about the nearest pizza place, while a human can navigate a complex physical world on a ham sandwich, the definition of intelligence must be questioned.

The Architecture of Waste: Sparse vs. Dense

This efficiency gap is a direct result of how these two engines are built. The human brain is inherently sparse. When you perform a task (like identifying a 10mm socket in a cluttered drawer) the brain does not fire every neuron it possesses. It activates only the specific, relevant circuits required for that visual and motor task. It is a system of just-in-time energy delivery.

Conversely, traditional AI models are dense. Because they are built on massive matrix multiplications, every time you send a prompt, the model has to search nearly every one of its trillions of parameters to calculate the output. It is the computational equivalent of redlining a 1,000-horsepower engine just to pull a car out of a parking spot.

The Limits of Language and Sensing

Language is a narrow part of our existence and describes only a thin slice of reality. The reliance on text and recorded data creates a hard wall that AI may never be capable of scaling. While we have developed sensors to monitor specific variables, they are incapable of capturing the holistic, common senses humans utilize every day. Even if such sensors existed, deploying them at a scale that replicates human presence is economically and practically unfeasible.

Data Correlation vs. Physical Experience

An AI sees a video of a car as a collection of pixels, but a technician sees that same car and experiences a symphony of sensory data. They hear the high-pitched hiss of a vacuum leak, feel a rhythmic vibration through the shop floor, and smell the acrid scent of burning oil.

While multimodal models can now correlate sounds and images—predicting, for example, that a certain sound usually accompanies a certain visual—they still lack physical embodiment. They do not feel the heat radiating from an engine block or the specific, tactile resistance of a seized bolt. Much of our complex evolutionary transformation occurred without formal language, yet AI is entirely trapped within the digital records of it. The failure of LLMs to achieve true intelligence suggests that reality is not just a social construct or a collection of tokens; there is a physical truth that exists beyond the data.

The Efficiency of the Biological Technician

While AI will be capable of automating standard administrative tasks, there is no evidence that current architectures will be capable of performing non-standard physical work. Tasks such as probing a circuit, removing complex interior panels, or remediating a seized bolt require a level of physical intuition that cannot be derived from a dataset.

In fact, attempting to build the immense datasets required to replicate these highly nuanced tasks would be the least efficient use of our resources. Humans are already optimized for these activities through millions of years of biological R&D.

The Role of the AI Assistant

Robots and LLMs will likely be useful as high-fidelity assistants. They can help define which diagnostic tests to perform, provide suggestions based on CAN bus data, and use computer vision to perform basic inspections or highlight damage unseen by the naked eye. However, the actual mechanical work will remain the domain of the human.

We will see companies attempt to build lean, fully automated shops run by robots, but they are unlikely to reach that goal using LLMs and current sensor fusion. The pursuit of AGI through language models ignores the fact that true intelligence is anchored in the ability to move through and manipulate the physical world—a feat that requires more than just the ability to predict the next word or pixel.

Agency and the Closed-Loop Failure

In control theory, a closed-loop system relies on an oxygen (O2) sensor to provide the ground truth of the exhaust gas. Without this feedback from the physical world, the system cannot correct itself. It is simply guessing at the fuel trim based on a pre-programmed map. Human agency is the O2 sensor of the cognitive world. We observe the actual chemistry of reality and adjust our actions accordingly.

The Feedback Loop of Model Collapse

Research by Shumailov et al. (2024) demonstrates what happens when a system lacks this external reference. When AI models are trained on their own synthetic outputs rather than primary human observations, they enter a degenerative feedback loop known as model collapse. Without a human primary observer to provide ground truth, the model begins to forget the tails of the distribution. It loses the rare, nuanced, and complex data points, eventually collapsing into nonsense.

This reality serves as a direct rebuttal to the post-modernist notion that truth is merely a social construct or a matter of language. In the shop, there is an objective physical truth. A technician acts as the essential external reference. They do not just predict what a car should do; they observe what it is actually doing and force a change in the distribution of the repair process.

The Subjective Experience Fallacy

The danger of removing the primary observer is that you are left with nothing but competing narratives. Imagine a technician attempting to tell a customer that their car is working from their own perspective because that is their subjective experience. This would be a laughable absurdity in a service bay where an engine is clearly knocking or smoke is billowing from the exhaust.

In the automotive aftermarket, we cannot narrative a car back into service. We cannot use rhetoric to fix a short to ground. The machine provides the ultimate objective audit. If an AI is trained on data that suggests a specific fault is normal operation, it will confidently gaslight the customer with statistical probabilities while the physical hardware fails. The AI has no skin in the game. It does not face the consequences of a thrown rod, whereas the human technician is tethered to the binary reality of the machine.

The Synthetic Dead End

There is no guarantee that a model creating synthetic auto repair data could reliably abstract new, useful learnings from that data. Training a model on its own generated text is like trying to charge a battery by plugging it into itself. Without the alternator of the physical world providing a fresh charge of primary observation, the system enters a state of information entropy.

In this state, the model does not just fail to learn; it actively erodes. It trades the noisy accuracy of rare edge cases for the smoothness of common averages. It becomes a closed-loop echo chamber that eventually forgets the very nuances that define a master technician's value. True abstraction requires the physical friction of the world—the tactile feedback of a wrench or the specific wave of a scope—to prove a hypothesis right or wrong.

The Problem of the Out-of-Distribution Vehicle

The limitations of statistical inference are most apparent when a model encounters something it has not seen several thousand times. A human technician can abstract by drawing on knowledge from similar vehicles, electrical principles, and physical intuition to handle missing context. This is known as reasoning from first principles.

A model is bound by its training data. Unless a specific conclusion is statistically represented in its weights, the system will fail. It hits a wall because it cannot reason from the basic laws of physics. While proponents argue that retraining will eventually fix these gaps, the automotive industry does not operate on the timeline of a data center. Customers are not going to wait for a model update to get their vehicles serviced. It would be unwise for original equipment manufacturers (OEMs) to ignore the necessity of the human technician as the only agent capable of resolving out-of-distribution problems in real time.

AI in the Automotive Aftermarket

We are decades away from economically viable models and robots that can perform the myriad tasks of a skilled trade. However, AI still offers value as an assistant to professionals. By treating these tools as digital teammates rather than replacements, we can bridge the gap between complex modern engineering and the practical needs of the service bay.

Estimating and Triage

Conversational tools are an excellent way to automate the collection of preliminary data, speeding up the estimating process and identifying upsell opportunities before the vehicle even rolls into the shop. Complex heuristics are hard to encode in a static form and can be fatiguing for customers. AI can identify the main issue and understand which follow-up questions to ask.

Many service advisors are not well versed in vehicle engineering, but they are excellent at sales. By using LLMs to aid in collecting precise data for technicians, the repair and diagnostic process becomes smoother. This digital receptionist protects the shop's calendar by ensuring that appointments are booked based on actual technician availability and parts lead times, rather than just an empty slot on a grid.

The Visual MRI: Automated Inspections

Large dealerships and progressive independent shops are rolling out systems that use advanced 360-degree cameras and machine learning to detect defects visually. These systems act as an "MRI for cars," pinpointing easy to miss leaks, worn suspension components (torn rubber components, leaking mag ride shocks), and maybe eventually suspension components being out of place under load.

Furthermore, identifying damage before the technician even touches the vehicle is essential for liability. Customers should feel confident their car is being returned in the same condition in which it arrived. In 2026, these camera images are converted into descriptive text on the repair order by an LLM, creating an objective source of truth that reduces friction between the shop and the customer.

Diagnostics and Cognitive Off-boarding

Probably our most exciting use for LLMs is the ability to build more effective diagnostic tools. Technicians can only store so much information about every vehicle platform in their brains. LLMs are a means of off-boarding some of that cognitive load to a centralized knowledge base.

While technicians are not typically data analysts, LLMs provide a digital tool that can perform basic but useful analysis of CAN bus data and telemetry. These assistants can scan millions of data points to find signal anomalies—pinpointing complex electrical faults that might otherwise take hours of manual probing. This does not replace the technician; it empowers them to find the culprit in a very complex system.

Resilience of the Trades

Because AI hits a hard wall at sensor fusion and physical agency, the skilled trades remain the most resilient sector of the economy. We will see companies attempt to build lean shops run by robots, but they will fail to reach that goal without the human O2 sensor to resolve the out-of-distribution problems that define real world repair. In the 2026 aftermarket, the shops that win will be those that pair machine speed with human judgment.

Sources

• Searle, J. R. (1980). Minds, brains, and programs. Behavioral and Brain Sciences.

• Shumailov, I., et al. (2024). AI models collapse when trained on recursively generated data. Nature.

• Cole, D. (2023). The Chinese Room Argument. The Stanford Encyclopedia of Philosophy.