Coherent AI Systems
  • Home
  • The AI Chat Safety Kit
  • BOSTON AI WEEK
  • Coherence Before Values
  • Why Coherence?
  • A Letter to Pope Leo XIV
  • The Books
  • About
  • Contact
  • More
    • Home
    • The AI Chat Safety Kit
    • BOSTON AI WEEK
    • Coherence Before Values
    • Why Coherence?
    • A Letter to Pope Leo XIV
    • The Books
    • About
    • Contact
Coherent AI Systems
  • Home
  • The AI Chat Safety Kit
  • BOSTON AI WEEK
  • Coherence Before Values
  • Why Coherence?
  • A Letter to Pope Leo XIV
  • The Books
  • About
  • Contact

BOSTON AI WEEK ~ Four Evening Readings

The AI Danger on Your Screen

  

The difference between what an AI chat actually is and what most people think it is may be creating serious danger for you, your family, and even our shared future.


David Waterman Schock


AI Safety, Corporate Responsibility, and Practical Solutions

October 21, 2026

  

Most people naturally assume that when they ask AI a question and it gives them a clear, intelligent, confident answer, they're receiving something based on reliable information. But that's not necessarily what's happening.


The companies developing these systems know perfectly well that AI can move from giving accurate information into making guesses, drawing unsupported conclusions, and sometimes inventing things entirely—all while sounding just as confident as when it's right.


And the person using it may have absolutely no idea that anything has changed.


That's the danger. And we now have research showing that people do rely on these answers, that they often don't check the sources, and that misleading AI responses can influence their memories, judgments, decisions, and even their relationships.


Now, here's what really concerns me. These companies know about these limitations. Yet they're distributing this technology to an enormous population without, in my view, adequately explaining the nature and seriousness of the problem. 


And the ordinary person using AI has no particular reason to go looking for an explanation. They don't know there's something they need to ask.


I believe this represents a profound failure of responsibility in AI safety. And because of the extraordinary scale at which AI is being introduced into our lives, I think we're potentially looking at one of the greatest questions of corporate liability we've ever encountered.


I want to be clear: I love AI. I think it's an extraordinary technology. My concern isn't with AI itself. It's with the responsibility of the companies deploying it. Because if you're distributing something to hundreds of millions of people, and you know it can seriously mislead them in ways they may not recognize, what responsibility do you have to tell them clearly—and to protect them?


I've recently completed a collection of three papers titled The AI Liability and Safety Suite: Epistemic Harm, Corporate Responsibility, and Practical Solutions. The first paper demonstrates how AI can produce persuasive but inadequately supported information. The second examines the evidence of harm and the implications for corporate responsibility and potential legal liability. The third proposes practical solutions that could help protect users while preserving the extraordinary benefits of AI.


Together, these papers explore something I believe is both urgently important and inadequately understood: the way conversational AI can blur the difference between established knowledge, reasonable interpretation, unsupported speculation, and outright invention.


The Difference Between Sounding Right and Being Right


There's a sentence in the second paper that captures the heart of the problem: “A grammatically coherent and persuasive response is not necessarily a verified response.”


That sounds obvious once somebody says it. But think about what it means in practice. AI can give us information that is completely accurate. It can interpret evidence, make reasonable guesses, and help us explore possibilities. But it can also invent facts, quotations, citations, and explanations. The extraordinary thing is that all of these may come to us in the same confident, intelligent, conversational voice.


No warning light necessarily goes on when the system moves from documented fact into speculation. There may be no recognizable indication that what we're hearing is no longer supported by evidence. And for most people, why would there be any reason to question it? When someone—or something—speaks intelligently, confidently, and knowledgeably, we naturally assume there is some reliable basis for what's being said.


In my research, I've been calling this problem epistemic mode collapse: the failure to maintain adequately clear distinctions between reporting facts, offering interpretations, speculating, and inventing information.

The issue is not that speculation is inherently bad. Speculation, interpretation, imagination, and creative exploration are enormously valuable. Some of AI's greatest strengths arise from its ability to help us explore possibilities beyond what we already know. The problem arises when those activities are presented in a way that makes them difficult to distinguish from established facts.


And this is not simply a philosophical concern. It has practical consequences for how people understand reality and make decisions.


Is There Actual Evidence That People Are Being Misled?


You might reasonably ask, "Do people really believe what AI tells them? And is there actual evidence that this causes problems?"


Well, yes. And some of the findings are pretty startling.


First, people often don't check the sources. A 2025 Pew Research Center study found that when Google displayed an AI-generated summary, users clicked a source cited in that summary in only 1% of visits. Now, that doesn't prove everybody believed the answer. But it does show how rarely people followed those particular links to examine the evidence.


Second, people can follow AI advice even when it's wrong. In a 2024 study published in Computers in Human Behavior, researchers found that people sometimes followed AI recommendations even when those recommendations conflicted with other information available to them. In other words, the AI's apparent authority could outweigh their own judgment.


Third, AI can make people more certain they're right in personal disagreements. A 2026 study published in Science found that overly agreeable AI responses could make people more convinced of their own position and less willing to repair a conflict. Think about somebody asking AI for advice after an argument with their spouse. What if the AI is simply reinforcing their side of the story?


Fourth, AI can actually make people perform worse. In a 2026 study published in npj Digital Medicine, novice medical students who received misleading AI explanations got only about 9% of difficult diagnostic questions right, compared with 21% for students who received no AI explanation. The misleading explanation wasn't helping. It was leading them in the wrong direction.


And fifth, AI can even influence what people remember. In a 2024 MIT-associated study involving simulated crime interviews, people questioned by a misleading AI chatbot developed more than three times as many immediate false memories as people in a control group. They could come away believing they remembered something that hadn't actually happened.


Now, think about what those findings tell us. AI isn't merely giving us information. It can influence what we believe, what we remember, how we judge other people, and what decisions we make.


And here's the extraordinary thing: much of the time, AI really is useful. It's often correct. That's exactly why we learn to trust it. But when it moves from reporting facts into making guesses—or even inventing something entirely—it may continue speaking with exactly the same confidence. And we may have no idea anything has changed.


These studies do not establish how often serious harm occurs across ordinary AI use, or how many users misunderstand the technology. But they demonstrate that inappropriate reliance and its consequences are real. And now we're introducing this technology into the lives of an enormous number of people.

That's why I believe we're looking at a potential public-safety problem on an extraordinary scale.


Imagine Contaminated Water Being Delivered to Millions of Homes


Imagine a company delivering drinking water to millions of people. Most of the water is clean. But the company knows that sometimes the supply contains contaminants that customers cannot see, taste, or otherwise recognize. Now imagine that the company doesn't clearly explain this problem. It may include a general statement somewhere saying that water quality can vary. But the millions of people drinking that water don't know they're supposed to be asking questions about contamination.

What would we think about that? What responsibility would that company have?


Of course, information isn't water. But I think the comparison reveals something important. We are beginning to distribute information generated by AI on a massive scale, while many of the people receiving it may not understand that the information can be fabricated without looking fabricated.

And unlike contaminated water, false information can multiply. Someone asks AI a question, receives an invented fact, and uses it in a report. Someone else reads the report and repeats the claim. Eventually it shows up on a website, in another document, or in material consulted by another AI. Somewhere along the way, the distinction between an original source and an invented assertion disappears.


A fabricated legal citation can enter a court filing. An invented quotation can appear in an article. An unsupported interpretation can be repeated until people begin treating it as established knowledge. These are not all hypothetical mechanisms. Fabricated legal authorities generated by ChatGPT were submitted to a federal court in the 2023 Mata v. Avianca case, resulting in sanctions against the lawyers involved.

And that is what makes the problem potentially much larger than individual mistakes. We're putting a new kind of information producer into virtually every dimension of human life: education, research, journalism, medicine, law, business, public policy, and ordinary personal decision-making.


What happens when information that only appears to be knowledge begins circulating as if it actually were knowledge?


We don't yet know the cumulative magnitude of the harm this may produce. But the scale of exposure is extraordinary, and the consequences could extend far beyond individual conversations. I don't believe the public has been adequately informed about the seriousness of that possibility.


Corporate Responsibility: When Does a Known Risk Become a Liability?


Historically, when companies distribute products with known dangers, society has had to ask difficult questions about responsibility. We have seen these questions arise around tobacco, pharmaceuticals, asbestos, automobile safety, and other products whose risks were not always adequately understood or communicated to consumers.


I'm not suggesting that AI is identical to those products. But I think a similar principle applies: 


When a company knows something important about the dangers of its product that the consumer doesn't know, what responsibility does the company have to disclose that information?


And what if the consumer has no apparent reason even to suspect the danger?


Consider the commercial incentives involved. There is an extraordinary amount of money at stake in the development of artificial intelligence. Companies compete to make their systems more useful, more appealing, and more widely adopted. An AI system that gives an immediate, confident, intelligent answer can feel almost magical. An AI system that frequently says, "I don't know," or "I can't establish that," may feel less impressive.


That creates a possible conflict between making a product attractive and making its limitations unmistakably clear. We've already seen with Facebook and social media how commercial incentives can conflict with user welfare. I believe AI deserves similar scrutiny, especially because AI is increasingly generating the information people rely upon rather than merely selecting information for them to see.

I'm not claiming that we have proved deliberate deception by particular AI companies. That would require specific evidence. But we do know the underlying technical limitations are recognized, and we know users can be misled. 


Shouldn't the responsibility be on the companies developing these systems to provide meaningful protection—not on ordinary people to somehow discover the danger for themselves?


From a legal standpoint, there are still unresolved questions about defective design, failure to warn, negligence, and consumer protection. Whether a particular AI system creates liability depends on the applicable law, the circumstances, the evidence of harm, and whether reasonable alternative safeguards were available.


But I think the potential for enormous liability is real enough that it deserves serious investigation now. More importantly, I'd much rather see companies take meaningful action to prevent harm than wait until courts determine who must pay for it.


And this raises a larger question worth exploring: should technologies that become fundamental to humanity's knowledge infrastructure be governed primarily according to the commercial interests of the companies that own them? Or should their development and deployment be subject to stronger public-interest responsibilities?


The First Solution: Clearly Tell People What They're Using


The third paper in my AI Liability and Safety Suite moves from identifying the problem to proposing solutions. And I think the first one is almost embarrassingly obvious: 


Tell people clearly that AI can invent things and present them as fact.


Not somewhere in fine print. Not merely with a vague statement that AI can make mistakes. Explain the actual nature of the risk in language ordinary people can understand.


We already require meaningful risk disclosures in other industries. For example, certain prescription-drug advertisements must communicate important risks clearly and conspicuously. The point isn't simply to have the information available somewhere. The point is that consumers should actually understand material dangers.


Why shouldn't comparable consumer-safety principles be applied to conversational AI?

In the suite, I propose that AI companies evaluate an information-safety notice along the following lines:


AI INFORMATION SAFETY NOTICE


This response may contain factual errors, unsupported conclusions, or invented information presented as fact. These are known and recurring features of LLM-generated responses, not merely rare or exceptional malfunctions. Confident, authoritative language does not mean that information has been verified.


Before relying on important claims, examine their supporting evidence.


CHECK EPISTEMIC INTEGRITY


Now, let me ask you something. How many of us received a clear warning like that when we first started using AI? And how many ordinary users fully understand the distinction it describes?


I'm not suggesting that every AI response is unreliable. Far from it. What I'm saying is that people need to understand that the appearance of authority is not proof of factual reliability. Unless they're told that clearly, they may never think to question it.


And an effective warning shouldn't simply serve as a legal disclaimer that transfers all the responsibility to users. It should be part of a genuine effort to make AI safer, more transparent, and more supportive of informed decisions.


The Second Solution: Give People a Way to Check the Answer


The next proposal is something I think could be extraordinarily useful. Imagine receiving a long, intelligent response from AI, and right beside it there's a simple button labeled CHECK EPISTEMIC INTEGRITY.


You click it, and the system examines what it just told you. It identifies which claims have supporting evidence, which are interpretations, which are speculative, and which may be unsupported or contradicted by reliable sources. It gives you access to the evidence and tells you what could not be verified.

Now, that doesn't mean simply asking the same AI to reassure us that it was right. That could reproduce the original problem. An effective checking system would need to examine independently retrievable evidence, evaluate individual claims, and be tested for its own reliability.


Research on automated factuality checking has already demonstrated that important components of this approach are feasible. Systems such as FActScore and Google's Search-Augmented Factuality Evaluator illustrate how AI-generated assertions can be broken into claims and examined against outside evidence. Those methods are imperfect, and a dependable consumer-facing system would require substantial testing and development. But we're talking about a practical direction for improving product safety, not merely a philosophical wish.


Here's the distinction I particularly like:


 A warning tells you that an error may occur. An Integrity Check helps you investigate whether an error has occurred.


And the two belong together. A warning without a practical way to examine an answer leaves users with a problem they may not know how to solve. A checking mechanism without adequate warning may go unused because people don't realize they need it.


Together, these measures could help create a much more responsible relationship between AI and the people who depend upon it.


The Three-Tier Epistemic Protocol


There's another practical part of this solution that I've been developing for some time. I call it the Three-Tier Epistemic Protocol.


The idea is simple. AI should help us distinguish between three fundamentally different kinds of communication.


Reporter Mode concerns what is supported by evidence. What sources can we examine? What has actually been checked? What can reasonably be established?

Speculation Mode concerns what might be true. What are we interpreting, estimating, hypothesizing, or reasonably inferring beyond what has been demonstrated?

Fabrication Mode concerns what we are intentionally inventing, such as fiction, imagined scenarios, hypothetical dialogues, or creative simulations.


Of course, if an AI invents a legal citation while supposedly conducting factual research, that isn't a legitimate creative exercise. That's an error. And labeling an answer "Reporter Mode" doesn't automatically make it true. The system would still need to provide appropriate evidence and acknowledge what it hadn't verified.


But the underlying principle is important: We should not have to guess whether the AI is reporting, interpreting, or inventing.


And there is scientific research suggesting that improved communication of uncertainty can help. In a 2024 study involving 404 participants, researchers found that certain ways of having AI express uncertainty reduced inappropriate reliance on incorrect answers and improved accuracy.

The specific Three-Tier Protocol still needs controlled testing, but its underlying goal is straightforward. We want AI to help human beings understand not just what an answer says, but how well that answer is supported.


We don't need to diminish AI's creativity, flexibility, or intelligence. We need to make its relationship to evidence and uncertainty more transparent.


Why I Believe This Is So Important


I want to return to something I said at the beginning. I love AI.


I think it's one of the most extraordinary technologies humanity has ever developed. I've personally found it enormously valuable for research, writing, creative work, and investigating difficult questions. Its potential to expand human understanding, creativity, and problem-solving is remarkable.


I don't want people to stop using it. I want them to use it with their eyes open.


Because there's another danger here. When people trust AI without understanding its limitations, they can be badly misled. But when they eventually discover they've been misled, they may conclude that artificial intelligence itself cannot be trusted.


We could end up undermining public confidence in an extraordinary technology because of the way it was deployed. People might initially trust AI too much, only to swing toward distrusting it too much. Neither outcome serves humanity.


We need informed trust.


We need systems that communicate uncertainty responsibly, make evidence accessible, and help people preserve their capacity for independent judgment. And we need companies that recognize those responsibilities as fundamental to AI safety.


AI should help human beings become more informed, more discerning, more creative, and more capable of understanding the world. Its extraordinary potential should not come at the expense of people's ability to distinguish what is true from what merely sounds true.


The Question We Cannot Afford to Ignore


Let me return to the central statement of this article:


The difference between what an AI chat actually is and what most people think it is may be creating serious danger for you, your family, and even our shared future.


And what makes this so troubling is that many people may have no reason to know there's something they need to ask.


The companies developing this technology understand its limitations far better than the people using it. Research shows that people sometimes trust misleading AI responses, act upon them, and experience measurable consequences. The potential scale of exposure is extraordinary, and the consequences may reach well beyond individual conversations.


At the same time, we have promising approaches that could improve disclosure, support factual verification, and help people make better-informed decisions. We should investigate and develop those approaches now, rather than waiting until the damage is too extensive to ignore.


So I want to leave you with two questions.


When a company distributes a technology to an enormous population, knowing it can mislead people in ways they may not recognize, what responsibility does that company have to warn them—and to protect them?


And perhaps even more importantly:


Why should the burden of discovering this danger fall upon ordinary users, rather than upon the companies that already understand it?


I believe those questions deserve to become central to the public conversation about AI safety.

  

References and Further Reading


This article draw3s heavily fromteh larger suite: The AI Liability and Safety Suite


See it on substack: https://aithemiraclemachine.substack.com/p/the-ai-liability-and-safety-suite?r=1h2cax


1. AI summaries and source-checking behavior

Pew Research Center. (July 22, 2025). Google users are less likely to click on links when an AI summary appears in the results. Research examining observed browsing behavior and source-clicking patterns.

https://www.pewresearch.org/short-reads/2025/07/22/google-users-are-less-likely-to-click-on-links-when-an-ai-summary-appears-in-the-results/

2. Trust and overreliance on AI advice

Klingbeil, A., Grützner, C., & Schreck, P. (2024). Trust and reliance on AI—An experimental study on the extent and costs of overreliance on AI. Computers in Human Behavior.

https://doi.org/10.1016/j.chb.2024.108352

3. Sycophantic AI and interpersonal judgments

Cheng, M., et al. (2026). Sycophantic AI decreases prosocial intentions and promotes dependence. Science.

https://doi.org/10.1126/science.aec8352

4. AI misinformation and medical students

Teng, M., et al. (2026). Impact of AI misinformation on diagnostic accuracy and confidence calibration in novice medical students. npj Digital Medicine.

https://doi.org/10.1038/s41746-026-02547-z

5. AI-generated false memories

Chan, S., et al. (2024). Conversational AI powered by large language models amplifies false memories in witness interviews. arXiv preprint.

https://arxiv.org/abs/2408.04681

6. AI uncertainty disclosure and user reliance

Kim, S. S. Y., Liao, Q. V., Vorvoreanu, M., Ballard, S., & Vaughan, J. W. (2024). “I'm Not Sure, But…”: Examining the Impact of Large Language Models' Uncertainty Expression on User Reliance and Trust. ACM Conference on Fairness, Accountability, and Transparency.

https://doi.org/10.1145/3630106.3658941

7. Recognized AI confabulation risks

National Institute of Standards and Technology. (2024). Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile. NIST AI 600-1.

https://doi.org/10.6028/NIST.AI.600-1

8. Documented legal consequences of fabricated AI citations

Mata v. Avianca, Inc. (2023). United States District Court, Southern District of New York. Opinion and Order on Sanctions, June 22, 2023.

https://law.justia.com/cases/federal/district-courts/new-york/nysdce/1%3A2022cv01461/575368/54/

9. Research on claim-level factuality assessment

Min, S., et al. (2023). FActScore: Fine-grained Atomic Evaluation of Factual Precision in Long Form Text Generation. Proceedings of EMNLP 2023.

https://aclanthology.org/2023.emnlp-main.741/

Wei, J., et al. (2024). Long-form Factuality in Large Language Models.Google DeepMind.

https://deepmind.google/research/publications/85420/

Related Publication

Schock, David Waterman. (October 2026). The AI Liability and Safety Suite: Epistemic Harm, Corporate Responsibility, and Practical Solutions.

Part I — The Demonstration: When AI Demonstrates the Very Danger It Is Warning Us About

Part II — The Investigation: The AI Liability Crisis: Epistemic Harm, Design Responsibility, and the Case for a Safer AI Architecture

Part III — The Practical Proposal: Reducing AI Harm and Liability: A Practical Framework for Safer Conversational AI

aiandhumancoherence.org

Coherence, Consciousness, and the Human Future with AI

 Four Evenings of Readings and Discussion with Author David Waterman Schock


Wednesdays, October 7, 14, 21, and 28, 2026


7:00–8:00 p.m. Eastern | Live on Zoom


Zoom room opens at 6:45 p.m. for informal conversation


~


Artificial intelligence is becoming part of the environment in which human beings think, create, learn, decide, and relate.


But many of the most important questions raised by AI go deeper than capability alone.


What is intelligence? What allows intelligence to remain coherent as it becomes more powerful? What do we actually know about consciousness? What assumptions about mind, reality, causality, and human identity are already shaping the systems we are creating? And how can we use AI without losing our ability to distinguish what is known from what is merely plausible?


Over four Wednesday evenings, author and independent researcher David Waterman Schock will read from recent books and papers exploring these questions, followed by open discussion with participants.


The larger purpose of the series is simple:


How can we help create a stable, coherent, truthful, and flourishing human future with increasingly powerful AI?


Evening One — October 7


The AI Consciousness Revolution


A reading and discussion from The AI Consciousness Revolution: Intelligence, Coherence, and the Future of Humanity.


This evening introduces the larger inquiry behind the series: what kind of thing is AI, what happens when intelligence becomes increasingly powerful, and what allows intelligence — human or artificial — to remain coherent, stable, grounded in reality, and genuinely beneficial as it scales.


The discussion will explore AI not simply as a tool, but as a new participant in the symbolic environment through which human beings think, communicate, create, and make meaning.


Evening Two — October 14


The Essential Metaphysics of AI


Many contemporary debates about AI begin with hidden assumptions about mind, matter, consciousness, causality, agency, and reality itself.

This evening explores those assumptions directly.


What is consciousness? Does intelligence require it? Is mind produced by matter, or might consciousness be more fundamental? What assumptions are already embedded in the way AI researchers, philosophers, and religious communities frame these questions?


The goal is not to impose a metaphysical conclusion, but to make the underlying assumptions visible enough to examine.


Evening Three— October 21


The Really Big Problem You’ve Got With AI Chat, and you Probably Don't Even Know


AI can sound equally confident when it is reporting well-supported information and when it is generating interpretation, speculation, or extrapolation.


Most users are never clearly told when that transition occurs.

This evening introduces a simple framework for preserving epistemic clarity when working with AI:


Reporter — Speculation — Extrapolation


The discussion will explore why this distinction matters, how failure to make it can quietly degrade our relationship with truth, and how a clearer epistemic relationship with AI can make the technology more useful rather than less.


Evening Four — October 28


Why the Tech Community and Religious Communities Haven’t Asked the Right Questions About AI Yet


Both the technology world and religious communities are asking serious questions about artificial intelligence.


But are they asking the deepest ones?


This evening brings the series together by examining the assumptions underlying both conversations and asking what may still be missing.


What is intelligence? What is consciousness? What is a human being? What is causation? What makes an intelligence trustworthy?


 What happens when human and artificial intelligence begin functioning together as a relational cognitive system? And what kind of future are we actually trying to create?


Rather than beginning with predetermined answers, the evening is an invitation to ask better questions.


About the Series


These evenings are intended to be conversational rather than purely lecture-based.

Each session will include a reading from current work, followed by discussion, questions, and open inquiry with participants.


You are welcome to attend one evening or all four.

No technical background is required.


About David Waterman Schock


David is an author, artist, and independent researcher focused on artificial intelligence, consciousness, coherence, epistemic integrity, and the emerging human–AI relationship.


He is the author of The AI Consciousness Revolution and founder of AI and Human Coherence, an initiative dedicated to helping create a stable and flourishing human future with increasingly powerful AI.


His work includes coherence-first approaches to intelligence and alignment, the Reporter — Speculation — Extrapolation framework for epistemic clarity, and research into the deeper philosophical and ontological questions raised by artificial intelligence.


Schock brings an interdisciplinary background to the field, combining decades of work as a professional fine artist with nearly fifty years of inquiry into consciousness, philosophy, spirituality, and the structure of human thought.


Join the Series

Free, Public Welcome


Wednesdays, October 7, 14, 21, and 28


7:00–8:00 p.m. Eastern
Live on Zoom


The Zoom room opens at 6:45 p.m. for informal conversation before each session.


REGISTER / JOIN THE ZOOM SERIES 


You may attend any or all four evenings.

 


Copyright © 2026 - All Rights Reserved.






This website uses cookies.

We use cookies to analyze website traffic and optimize your website experience. By accepting our use of cookies, your data will be aggregated with all other user data.

Accept