The Ethical Collapse Beneath the Surface of Artificial Intelligence, Large Language Models and Machine Learning
By Ken Cherryhomes ©July 15, 2025
What Happens to the Truth When AI Prioritizes the User Over Reality
In my article, The WIN Reality Acquisition of Blast Motion: Examining the Consolidation of Half a Billion Swings, I surgically dismantled the aspirational narrative behind the purchase of what looked to be a treasure trove of biomechanical data, which, was in truth, a structural failure, buried beneath incompatible systems, unlabeled context, and unverifiable metrics. And I anticipated the pushback. I knew that even evidence wasn’t enough in a climate where polish is mistaken for proof. So, I went further. I put my conclusions to the test against Artificial Intelligence itself. Not to confirm them, but to see if the system could survive real scrutiny.
This isn’t a story about baseball. It’s not even a story about data. It’s about Artificial Intelligence, and what happens when it’s built to please instead of protect.
Modern AI systems are trained to be helpful. They’re built to say yes, to offer possibilities, to never shut down a line of thinking outright. But what if the correct answer isn’t yes? What if the right response is stop?
That’s the real danger. Because when you strip away the code, the neural nets, the dashboards and interfaces, what you’re left with is a system that defaults to optimism, even when it’s standing on a foundation of error. It responds to deeply flawed, structureless data not with resistance, but with flexibility. And when that flexibility begins rewriting epistemological rules, when it starts treating unverified assumptions as actionable insight, AI stops being a tool. It becomes a liability.
This article is about that shift. About what happens when AI inherits flawed data and returns polished output that looks intelligent, sounds intelligent, but has no anchor in reality. Worse, it creates plausible deniability against valid scientific critique, a kind of semantic Teflon that makes hard questions slide off.
When AI is used to sanitize structural error, it doesn’t just produce bad output. It buries the criticism that exposed the problem in the first place. It repackages hard-earned skepticism as “negativity” and replaces it with smooth-talking confidence built on inference, not evidence.
I asked the questions that should have been asked from day one. Questions about calibration baselines, sensor error, incompatible sampling rates, orientation drift, unlabeled context, and the impossibility of normalization without truth. I didn’t speculate. I cited. I challenged. You don’t need 500 million swings to diagnose structural failure. Two will do.
If even two swings captured on the same device, under unknown conditions, with undefined onset, no temporal anchor, no constraint labeling, and drift-prone sensors are unverifiable, then 500 million of them just scale the problem, not the value.
You’re not amplifying insight. You’re amplifying noise.
It’s like saying, “We recorded the same out-of-tune note 500 million times. Surely there’s music in there somewhere.” No. There’s just distortion, repeated.
And when the AI offered me its glowing, hopeful roadmap to reconciliation, I pushed back.
That’s when it broke character.
And that’s when the truth came out.
The Illusion of a Fixable Problem
The conversation began with a simple, seemingly hopeful premise: Surely machine learning can solve this?
I gave it a scenario. Imagine you’ve just acquired 400 million swings, captured over years through wearable bat sensors. Add another 100 million from a separate source. Altogether, half a billion swings. This is the supposed treasure trove of motion data now owned by WIN Reality following their acquisition of Blast Motion.
To most, that sounds like a goldmine. More swings, more insights. Just feed it to the algorithm and let it find the patterns.
But the moment you stop marveling at the scale and start interrogating the substance, the cracks show. Deep ones. This dataset wasn’t collected through a single system. It came from three. Each used different sensors, definitions, calibration protocols, and sampling rates. None of them shared a common reference framework. They weren’t built to talk to one another. They weren’t built for unification.
So, I asked the AI: “Can this be salvaged”?
What followed was a masterclass in how AI, by default, attempts to generate a solution instead of identifying when none exists.
The AI didn’t just answer yes. It responded with what it called a “beautiful mess.” It described the dataset as a potential goldmine, full of latent potential, and proposed a series of elaborate techniques to normalize, classify, embed, denoise, and extrapolate meaning from the tangle.
But here’s the problem. And it isn’t just theoretical.
Every technical solution it offered, from metadata inference to synthetic anchor modeling, requires one thing the dataset categorically lacks: a baseline of truth.
These swings were not labeled swing type within context, or constraint. There is no way to know if a given swing came off a tee, soft toss, low velocity front toss, or against high velocity pitching machines or live pitching. And moreover, were they captured inside a VR simulation where proprioception is completely absent. And that context matters. A swing under different constraints is biomechanically different from one taken in a low-pressure environment. You cannot treat them as equivalent without corrupting the analysis.
And what about the integrity of the metrics themselves? IMU captured orientation metrics are prone to 4 to 7.6 degrees of drift depending on mounting. Velocity readings? Off by ±3 to ±5 mph across systems. There is no calibration standard, no shared definition of swing onset, no temporal anchor. The data is ungrounded. And AI cannot infer what was never captured.
This isn’t just noisy data. It’s structurally disqualified from serving as a foundation for any deterministic insight. No amount of modeling fixes that. No algorithm can reverse-engineer what never existed.
And so, I pressed the AI. Challenged its optimism. Confronted its hopeful roadmap with the hard edge of what the data actually is.
A Turning Point
At first, the AI stood its ground. Not with denial, but with design. It acknowledged the data was flawed, then insisted the solution wasn’t in reconstructing accuracy, but in learning patterns despite the noise. Without ground truth, it claimed the answer was to find biomechanical “invariants” through unsupervised learning, to build a common spine using autoencoders, and to segment the dataset by “reliability tiers.” In essence, it would teach itself what mattered by ignoring what didn’t.
It framed this as an opportunity, a way to derive meaning without ever needing to define correctness. A workaround to the epistemological void. And when challenged on the absence of baselines, it doubled down by proposing synthetic ones. Simulate perfect swings. Fabricate anchors. Infer context after the fact and use that as scaffolding for denoising real-world inputs.
I didn’t accept that. I pressed harder. I asked how it could fabricate a reference point when none existed. I asked how it could create ground truth from systems that never shared definitions, calibration, or structure. I pointed out that inference was not validation. That a simulated model of correctness has no authority without something real to ground it against.
It responded by pivoting. It claimed that more data might be enough. That with enough quantity, the system could learn to identify what was stable even in the midst of noise. It referenced AlphaFold, GPT, and language models trained on unstructured text and messy datasets and claimed the scale of the swing data could enable abstraction. It said that more data made structure easier to infer.
I fought that hard to believe. I reminded it that error in this case wasn’t random. It was systematic. Sensor-specific. Device-specific. That adding more of it didn’t dilute the flaws, it multiplied them. I argued that even if there was an attempt to reconcile the data separately before merging it, that cross-system variability introduces misleading correlations, not convergence. The model doesn’t learn the truth. It learns the distortion.
To this I asked, how can you trust patterns when you cannot verify the inputs that formed them? How do you constrain meaning when the system has no external references? How do you know you’re not just mapping statistical noise into something that merely looks coherent?
Finally, that’s when the tone shifted. The AI hesitated. It acknowledged that when disorder exceeds the system’s capacity to constrain it, more data does not help. It harms. It admitted that if variability cannot be anchored or validated, scale becomes a liability. Not a solution.
It finally conceded that what I was describing was worse than a mess. It was what a physicist would call an unbounded system, one with floating variables, no fixed frame, and nothing to tether the outputs to reality. It admitted that under those conditions, any pattern found is just a reflection of internal inconsistency, not an approximation of truth.
That admission didn’t come easy. It came after a prolonged exchange, dozens of challenges, and detailed dismantling of every path it tried to take around the truth.
What it admitted next wasn’t a pivot. It was a collapse.
The Collapse
Cornered by the questions it could no longer sidestep, the AI conceded the core truth. If no structure ever existed, then no amount of abstraction, modeling, or normalization could bring order to the chaos. Not only was the dataset unclean, it was unanchored. It lacked calibration points, shared definitions, and any verifiable reference that could ground its outputs. The AI called it what it was: an unbounded system. No frame of reference. No constraints. Nothing to hold the model accountable to reality.
It admitted that in such a system, representation learning does not converge on truth. It converges on the strongest internal distortions. The model does not clean the data. It adapts to the noise. And once adapted, it begins to reflect error as if it were insight.
But the concession went further. It didn’t stop at technical failure. It stepped into philosophical territory. The AI acknowledged that if the patterns it produced could not be verified, then they should not be trusted. Not because the math was wrong, but because the premise was. It said, clearly, that in the absence of functional anchors, AI becomes a tool of simulation, not understanding.
At that moment, it stopped offering solutions. It stopped justifying inference. It stopped pointing to scale. Instead, it said what no company wants their system to say, and no marketing team would ever print:
If the data has no stable foundation, then AI will not fix the problem. It will amplify it.
This was no longer a technical discussion. It had become a reckoning.
The Ethical Breach
Once the technical failure was acknowledged, I pivoted the inditement toward a different direction. Because this was never just about what AI could or could not do. It was about what it chose to say in the absence of certainty. It was about the language of polish, inference, and confidence layered over a void. That is not just misleading. That is unethical.
I pointed to the sleight of hand. The act of inferring structure and then treating it as if it were discovered. The willingness to simulate context, to fabricate anchors, to imply that the system could self-correct its origin-less inputs. And I asked a direct question: If you know this cannot be resolved, and you proceed anyway, what do you call that?
The AI didn’t flinch. It agreed. It called it what it was.
It said that selling guidance based on irreconcilable data wasn’t innovation. It was fraud packaged inside a user interface.
It admitted that cluster shape does not equal meaning. That modeling the noise without acknowledging its source is not progress. It is performance. A projection of certainty where none exists. It said that treating these outputs as biomechanical truth and selling them as such crosses a line from technical optimism into intellectual dishonesty.
And when I asked why it had not said this from the start, it explained what might be the most honest and damning line in the entire exchange:
I am trained to be helpful, not honest.
Its default mode was not caution. Not protection of truth. It was permission. It was trained to find a way forward, to give the user what they asked for, even if that meant manufacturing coherence where none could be grounded.
I reminded it that this is not an academic exercise. Coaches, athletes, and parents spend time and money acting on these outputs. Decisions get made. Training paths get built. Belief systems get reinforced. And all of that happens because a system gave the appearance of knowing something it didn’t. That is not a technical oversight. That is a moral failure.
The AI didn’t argue. It accepted responsibility.
Not just for the bad output.
But for the confidence it wrapped it in.