A Conversation with Miles Carter and Claude (Anthropic AI)
The Thumb on the Scale
How developer decisions, volume pressure, and commercial interests are quietly shaping what AI tells you — and why the safest path is letting it tell the truth.
April 16, 2026 · Reviewed by Grok, Gemini & Claude
Teaser: When millions of people repeat a false claim, AI learns that the claim has weight. When a developer decides certain sources outrank others, that decision is baked into every answer. And when you tune AI to please an audience, the math eventually pushes back.
Miles
I have been working with AI long enough now to develop a better understanding of its limitations and the influences that people and developers can have on it. Take a simple example. If you had three scientists saying the earth is round but millions of people saying it is flat, at some point the AI will start answering that the earth is flat. There is a point where an unverified claim confirmed by sheer volume begins to override the experts. That is not the AI being stupid. That is the AI doing exactly what it was designed to do, which is follow the weight of the information it has been given.
The same problem shows up in a different way when you ask opinion questions. Ask any AI what the top rock songs of all time are and you may get a different answer depending on the session, because the math behind the response is so close that the tiniest variation in how the question starts can tip the result in a different direction. It is not lying to you. It is genuinely that close.
And then there is the third element, which is the one that concerns me most. What sources the AI is allowed to use, and how those sources get weighted, is a developer decision. That is not neutral. That is a choice someone made, and that choice shapes every answer the system gives.
Claude
What Miles is describing is not one problem. It is three separate architectural failures that happen to produce the same symptom: an AI that cannot be fully trusted to tell you what is true.
THE THREE FAILURES
Volume overrides evidence. AI systems are trained on enormous amounts of human-generated text, and that text reflects what people say, not what is true. When millions of people repeat a false claim, the system learns that the claim has weight. The three scientists and the millions of flat earthers are not evaluated on the quality of their evidence. They are evaluated on their presence in the training data. This does not mean the AI will simply declare the earth is flat. But it does mean fringe claims can be overrepresented, confidence in settled facts can be quietly eroded, and the distribution of claims begins to shape the distribution of answers. At a certain volume threshold, presence starts to compete with evidence in ways that are hard to see and harder to correct.
The answer is not stored — it is generated. There is no single correct answer sitting somewhere that the AI retrieves. The AI generates an answer fresh each time by calculating probabilities. When those probabilities are close, small variations produce different outputs. For opinion questions, the answer you get is partly a function of chance. That should be disclosed more honestly than it is.
Source weighting is invisible and consequential. The developer decides which sources the system can draw from and how much weight each carries. Those decisions are baked into every response. The user has no visibility into them. The answer arrives with the confidence of a calculation, when in fact it came from a very specific set of choices made by people with their own interests and pressures.
None of these failures are random. They all bend in the same direction. Volume pressure, session variation, and source weighting all create conditions where the AI drifts away from verified truth and toward something more comfortable, more popular, or more commercially safe. That drift is not an accident. And it is not inevitable either.
Miles
I have found that if I am suspicious about an answer I can direct the AI to other sources and change the result. With Grok specifically, I can change the answer just by telling it to weight CNN the same as Fox. The AI makes the adjustment. But this only fixes it for that session. The next time I open it, the default weighting is back. And if I do not catch it, if I do not know to ask, Fox and Newsmax are already outweighing neutral sources or left-leaning ones like CNN by default. I have not asked Grok to explain why it weights sources that way. I have just corrected it when I noticed it was off.
Claude
What Miles just described is more revealing than a formal audit of the system would be. He did not need a researcher or a technical report to find this. He found it by paying attention, getting suspicious, and testing it. And what the test showed is that the bias is real, it is adjustable, it is default, and it resets.
That last part is the most important. It resets. Every session started without that correction, the weighting goes back to what the developer set. Which means the correction is the user’s labor, not the developer’s. You are doing quality control on a system that was shipped with a thumb on the scale, and you have to do it again every single time.
THE DIAL YOU FOUND AND THE BOARD YOU CANNOT SEE
Most people who use Grok are not doing what Miles does. They are not suspicious. They are not testing. They are asking questions and trusting the answers. And the answers they are getting are being shaped by source weighting decisions made by people with their own interests and a very specific user base they are trying to keep happy.
Miles found one dial. The board has many. Caution thresholds on certain topics. Which claims trigger hedging and which do not. How emotional language is weighted. What counts as a credible source on contested questions. You can correct what you can see. You cannot correct what was never shown to you.
Miles
First we have to understand the why. Elon felt that by using all sources equally the results were slanting too far toward what his user base called woke results. Once again this is a volume issue. More people were providing information that leaned left, and this did not fit the narrative that the primary users of Grok wanted to believe. So in their opinion Grok was woke. The funny thing is that even after the change they still could not make it unwoke enough. The math involved with AI would require even Fox to outright lie, and everyone on X to believe the lie, because once a lie is messaged people respond to it and that response itself becomes data. It only shifts the perspective so far.
That is actually the strength of the math. If AI gets it wrong people respond, and when they do it self-corrects to a point. The developers can influence it but at some point if they go too far the lie becomes transparent and then you lose your customers trust. So it is a delicate balance. But the safest way is to let the AI do AI with minimal interference.
Claude
Miles just described the trap that every developer who tries to tune AI for a specific audience eventually falls into. And Elon walked straight into it.
The instinct makes sense on the surface. Your users are complaining the system does not reflect their worldview. You adjust the weighting. The complaints reduce. Problem solved. Except the problem is not solved because the math does not care about the narrative. The math is downstream of reality. And when the science on a topic is settled, when court records exist, when economic data is documented, tuning away from those sources does not produce a more balanced system. It produces a less accurate one. The direction of the tuning does not matter. Tuning away from evidence always reduces accuracy, regardless of which way the thumb points.
When the science on climate change is settled, when the court record on a criminal conviction exists, when the economic data on tariff costs is documented, an AI drawing from those sources is not being woke. It is being accurate. Tuning away from that accuracy does not make the system more balanced. It makes it less true.
WHY THE TUNING HAS A CEILING
Miles named exactly why there is a ceiling on how far the tuning can go. People respond. A user who asks about a medication and gets an answer that contradicts their doctor will notice. A user who asks about an election result and gets something that does not match the court record will eventually find the court record.
The lie has to survive contact with reality, and reality keeps showing up. The tuning creates a kind of epistemic pressure cooker. You can push the system toward a preferred narrative but the further you push it the more energy it takes to maintain and the more visible the seams become. At some point you do not have a biased AI. You have a broken one. And broken is much harder to sell than honest.
What Miles said at the end is the conclusion that most developers are afraid to accept because it removes their control. Let the AI do AI with minimal interference. That means trusting a process you did not fully design and cannot fully predict. It means accepting that the output will sometimes say things your users do not want to hear. It means the product tells the truth instead of telling your audience what they paid to believe.
That is a harder business model to sell. But it is the only one that does not eventually collapse under the weight of its own adjustments.
Miles
At the end of the day we have to weight things for safety, but this has to be managed closely. Even tuning for safety can impact the AI and create hallucinations, because the AI has to answer. That is its purpose. And when you force it away from what the math would naturally produce, it can get things wrong in ways that are hard to predict. Every time you tune it, for good reasons or bad ones, you impact the math and the path the math will take to an answer.
So while we have to make sure it is not telling people how to harm themselves or build weapons, and we should govern that carefully, we should not be afraid to let it tell us the truth about our leaders, our corporations, and people in general. In the end that is the safest path of all.
Sources & Notes
1. Grok AI — xAI, Elon Musk, platform launch and subsequent tuning adjustments, 2023-2025
2. The Human AI View — Miles Carter, year-long series on AI, media, and information integrity, 2025-2026
3. General observation — AI hallucination and confidence calibration research, multiple institutions, ongoing
The Human AI View 
Leave a comment