AI companies shouldn't get Section 230 immunity, and hallucination isn't a bug waiting to be fixed
Ken argues that because generative AI synthesizes new content with the source in the model itself, it looks more like a publisher than a neutral platform; and hallucination is closer to confabulation in neuropsychology, a risk built into how the system works.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
AI companies shouldn't get Section 230 immunity
Ken's answer to whether AI companies should be protected by Section 230 of the U.S. Communications Decency Act is an unambiguous no. His reasoning: the premise of the 1996 law is a simple distinction — users create content, platforms merely host it. Generative AI doesn't retrieve or display something someone else made; it synthesizes new content, sometimes producing sentences that don't exist anywhere in the training data, and it will state fabricated things confidently as fact. The source of a defamatory post a user publishes is the user; defamatory content generated by AI was never written by a human, so the source is the model itself. At that point the AI company no longer looks like a neutral intermediary and looks more like a publisher of content.
— Ken FordHallucination isn't a bug waiting to be fixed
Ken stresses that hallucination is more accurately called confabulation. It isn't a software defect waiting for a patch; it comes from the probabilistic nature of today's autoregressive large language models — they generate text that looks plausible rather than retrieving verified facts. Retrieval, reasoning and verification techniques can reduce it substantially but have not eliminated it, and there is good reason to think the problem is intrinsic to this generation of models and to the underlying computer science. If Section 230 or similar protections are narrowed, AI companies may bear more liability for false, defamatory or harmful output, which would create a powerful incentive to develop architectures that anchor more reliably to facts — possibly even to fundamentally different approaches to AI.
— Ken FordAI legal personhood has already been sentenced to death
Ken thinks the idea of AI gaining legal personhood doesn't hold up: corporations are "legal persons" not because they are intelligent or autonomous but because they are structured proxies for real humans — shareholders who own stock, employees who act on the entity's behalf, capital assets that can be seized to pay debts. AI has none of these. This isn't his personal opinion; it has been tested and rejected. The DABUS system was put forward as a patent inventor for years, and the U.S. in 2022, the U.K. in 2023, Germany in 2024 and Japan this March all said no — only natural persons can be legal inventors. The "electronic personhood" concept proposed in the EU in 2017 died almost immediately, dismantled by an open letter signed by more than 150 AI researchers, roboticists and legal scholars.
— Ken FordData center panic is understated, not overstated
Responding to the June 12 Atlantic article arguing that data center panic is exaggerated and that critics are driving up costs, Ken's judgment is that it reads more like a rationalization for "the largest transfer of public resources to private companies in modern history" than a balanced analysis. He points out the article understates the scale of the infrastructure: current and planned AI facilities are measured not in megawatts but in gigawatts, and a gigawatt-scale data center at full load consumes as much electricity in a year as a small state or even a major metropolitan area, while a single hyperscale facility may use millions of gallons of water a day for cooling. Noise, waste heat and heat island effects are also rarely mentioned by expansion advocates.
— Ken FordCommunities aren't even allowed to know the terms
Ken points out that a great deal of decision-making happens behind closed doors: hyperscalers such as Google, Amazon, Meta and Microsoft and data center developers often sign nondisclosure agreements with local officials first, and only afterward do communities learn how much water and power a project will need, what infrastructure upgrades it requires and what public subsidies are involved. In Virginia, where data centers are most concentrated, a review found that 25 of 31 localities with existing, approved or proposed data centers had signed such agreements. One NDA in Bessemer, Alabama even required municipal officials to destroy confidential project records when the agreement expired or the developer asked. Microsoft later pledged to stop signing NDAs with local governments, which Ken sees as an admission that secrecy had become a serious problem.
— Ken FordCompanies inflate the danger and it backfires
When the developer of a technology is also the primary source for claims about how powerful and dangerous it is, there is a subtle tendency to exaggerate the capabilities of current systems, because that reinforces the impression that the company has achieved extraordinary technical leadership. The result is that public discussion gets distorted twice over — risks are overstated and the technology looks more revolutionary than the evidence supports. And the strategy undermines itself: the more successfully you convince the public and policymakers that your products are extremely dangerous, the stronger the political case for regulating those products. Ken stresses this isn't to say companies deliberately invent risks that don't exist, nor that AI companies oppose all regulation — Anthropic often argues for stronger AI oversight.
— Ken FordCall your model a national security technology and don't be surprised
This isn't hypothetical. Earlier this summer, the U.S. government issued an export control directive requiring Anthropic to suspend foreign nationals' access to its Fable 5 and Mythos 5 models, restoring access only after additional safeguards were added. Ken's conclusion: once a company publicly frames its models as exceptionally powerful and dangerous, it shouldn't be surprised when the government starts treating those models as national security-relevant technology. He argues for a more restrained mode of communication from the industry — no utopian promises and no doomsday prophecies, but evidence-based assessments of capability and risk; if a system truly has unprecedented dangerous capabilities, the supporting evidence should be public for independent evaluation.
— Ken FordThe word hallucination lets the model off the hook
Ken objects to using hallucination to describe large model errors because the word activates a whole frame of thought. Hearing that "a person hallucinated," we naturally imagine an otherwise competent mind briefly perceiving something that isn't there, implying an abnormal malfunction of a reliable cognitive system. But today's large models don't work that way: when it gives a wrong answer there is no temporary cognitive breakdown, and in most cases it isn't a malfunction at all — it's doing what it was designed to do, generating a plausible continuation of text based on patterns learned in training, the prompt and the context. The error isn't an exception to how the model works; it's a risk built into how it works.
— Ken FordFabrication is wrong too, so he settled on confabulation
Ken at one point preferred fabrication over hallucination, because a fabrication is "something constructed," closer to what a language model does. But the more he thought about it the less he liked it: fabrication usually means someone deliberately invents something in order to deceive, whereas a model making an error has no intent, no awareness that the output is false, and no attempt to mislead anyone. The word he ultimately settled on is confabulation, borrowed from neuropsychology: a person confidently gives a detailed, coherent, entirely believable account of something that never happened, without realizing it is false and without any intent to deceive. Healthy people also confabulate when they reconstruct memories and unconsciously fill gaps with plausible detail.
— Ken FordChatGPT didn't program itself
On the claim that "ChatGPT-5 was programmed by earlier ChatGPT," Ken says this is a misleading conflation: what OpenAI announced was ChatGPT-5.3 Codex, a dedicated coding and software engineering model, not the chatbot the general public interacts with. What "played a role in creating itself" refers to is OpenAI engineers using earlier versions to help debug training, monitor training runs, diagnose evaluations, optimize parts of the harness and support deployment. That's AI-assisted engineering, not AI autonomously deciding to build its own successor. He also notes that from OpenAI's wording to tech blogs' "builds its own models" and "recursive self-improvement," these phrases never appeared in OpenAI's announcement — they were amplified layer by layer through multiple rounds of secondary reporting.
— Ken FordDistillation is legal, but don't train rivals on my output
Frontier AI companies themselves routinely distill models into smaller, cheaper versions, but they are far cooler about it when competitors use their outputs to train rival models. Anthropic, on one hand, calls distillation a legitimate technique when used properly, and on the other names specific companies it accuses of abuse: in February 2026 it said DeepSeek, Moonshot and MiniMax generated more than 16 million interactions with Claude through roughly 24,000 fraudulent accounts; in a June 2026 letter to two U.S. senators it accused Alibaba of conducting more than 28.8 million interactions through roughly 25,000 fake accounts, specifically to extract Claude's agentic reasoning, software engineering and long-horizon task completion capabilities. Anthropic's help center also explicitly prohibits using Claude outputs to train competing general-purpose models.
— Ken FordThe 50-meter car wash: the model answers the wording, not the task
In February 2026 a Mastodon user asked four frontier language models: I want to wash my car, the car wash is 50 meters away, should I walk or drive? All four suggested walking — but the correct answer is obviously to drive, because the car has to physically be at the car wash. The AI infrastructure company Opper then tested 53 leading models systematically: in the first round only 11 answered correctly and 42 suggested walking; on repeated tests only 5 got it right every time. Ken's analysis: the prompt contains a highly salient corpus cue — 50 meters is a very short walking distance, and in training text "short distance + walk or drive" is strongly statistically associated with "walk" — but that association is irrelevant here, because the purpose of the trip isn't to transport a person, it's to transport a car.
— Ken FordThe 50-meter car wash exposes more than a lack of common sense
A preprint from Carnegie Mellon University published a few weeks after the Opper test gave a stricter version: the short-distance cue influenced model answers roughly 9 to 38 times more than the explicit goal of washing the car; at the token level the patterns looked more like keyword association than compositional reasoning. When the car as an object was made more explicit in the prompt, model performance improved substantially — showing the relevant pattern is in the system, it just isn't strong enough to override the "50 meters" cue. Ken stresses this isn't the old refrain that "AI lacks common sense," but what that absence looks like in practice: an answer can be fluent and locally plausible while being wrong at the level of the task.
— Ken FordFluent confidence and task understanding are different abilities
Humans automatically represent an errand in terms of physical constraints — the car has to be at the car wash. What the model outputs is the strongest linguistic association in the prompt and training data. From this Ken draws a line: large language models are extremely good at generating what a thoughtful person would say, but far less reliable at checking whether this particular goal imposes a constraint that a generic answer would miss. This failure mode has already appeared in more consequential settings: stringing together a sequence of steps that each look reasonable while missing a requirement buried in the original request. And this gap doesn't announce its own existence.
— Ken FordHe won't sign the 200-person statement because it's too vague
Recorded on July 17, five days after nearly 200 researchers and economists, including 16 Nobel laureates, signed a statement warning that AI could replace a large share of human work and cause widespread unemployment. Ken says he probably wouldn't sign it — not because AI doesn't matter, but because the statement is too vague to be useful. Its three claims (AI could become extremely powerful, could bring an economic transformation larger but faster than the Industrial Revolution, and policymakers should act now) are all just "could," with no accompanying analysis, no probabilities, no specific mechanisms and no concrete policy proposals. His own verdict: it's more a headline than an argument.
— Ken FordThe statement omits the physical constraints of compute and power
Ken notes that the statement's "incentives, guardrails, institutions" are never specified: is it education reform, retraining, tax policy, labor market adjustment, antitrust, AI safety standards or liability rules? Who does what? He thinks a serious AI statement should address the resource demands of the current scaling paradigm — most recent progress comes not from elegant breakthroughs in understanding or reasoning but from brute-force scaling: bigger models, more data, more GPUs, ever larger data centers. Electricity demand, water use, grid capacity, siting and compute concentration are already becoming real constraints. He would rather see emphasis on more efficient learning, better reasoning architectures, hybrid symbolic and statistical methods, causal modeling, smaller specialized models, and better memory and retrieval.
— Ken FordWe're better at predicting which jobs disappear
The statement emphasizes mass unemployment but doesn't explain why that outcome should be seen as likely rather than merely possible. History has plenty of examples of technology displacing workers, and plenty of examples of technology creating industries, occupations and forms of work that couldn't have been imagined beforehand — the Industrial Revolution, electrification, the computer and the internet all destroyed jobs and created far more new ones. The lesson Ken draws from economic history is that we are usually much better at predicting which jobs will disappear than at predicting which new jobs will appear. It's easy to look at an existing occupation and imagine how AI automates it; it's much harder to imagine what new occupations, services, institutions and industries will grow once the technology is widely deployed.
— Ken FordIn their own words · checked verbatim
That's a good question and my answer is generally no. They should not be.
Ken Ford3:06
Hallucinations, which are as I mentioned earlier, are actually much closer to confabulations, are not simply software bugs waiting to be patched.
Ken Ford7:11
Changing the location of a data center does not change the underlying physics.
Ken Ford28:27
That is not an exception to how the model works. It's a risk built in to how the models work.
Ken46:41
It isn't lying and it isn't hallucinating in the human sense. It is constructing a plausible answer from incomplete or imperfectly grounded information. The same general process it uses when it happens to be right.
Ken49:43
Yet, no one standing next to the computer gets wet.
Ken54:54
The invention may be centralized, but adoption is likely to be decentralized.
Ken1:14:14
The significance is that advanced AI capability is becoming abundant, cheaper, more portable, and harder to contain.
Ken1:21:22
is more of a headline than an argument.
Ken Ford1:31:32
But, the core uncertainty here is not primarily economic. It is really deeply technical.
Ken Ford1:32:32
Figures
| Virginia localities that signed NDAs | 25 of 31 | 14:20 |
| Jurisdictions that rejected the DABUS case | U.S. 2022, U.K. 2023, Germany 2024, Japan this March | 8:11 |
| Signatories against the EU electronic personhood concept | more than 150 AI researchers, roboticists and legal scholars | 8:11 |
| Publication date of the Atlantic article | June 12 | 11:15 |
| Models tested by Opper | 53 leading models | 1:23:24 |
| Models that answered correctly in the first round | 11 (42 suggested walking) | 1:23:24 |
| Models that answered correctly every time on repeated tests | 5 | 1:23:24 |
| Interactions Anthropic accused Alibaba of | more than 28.8 million, via roughly 25,000 fake accounts | 1:12:11 |
| Factor by which the short-distance cue influenced model answers (relative to the car wash goal) | roughly 9 to 38 times | 1:26:27 |
| Researchers and economists who signed the AI unemployment risk statement | nearly 200, including 16 Nobel laureates | 1:29:28 |
Glossary
- confabulation
- A neuropsychology concept: confidently recounting something that never happened, with no intent to deceive and no awareness that it is false.
- Section 230
- 1996 U.S. legislation that spares platforms from publisher liability for content their users post.
- DABUS
- An AI system put forward as a patent inventor; courts in multiple countries have ruled that only natural persons can be inventors.
- distillation
- Using a large model's outputs to train a smaller, cheaper model — something frontier companies themselves do routinely.
- agentic reasoning
- A model's ability to plan autonomously, call tools and complete tasks over multiple steps.
How to listen
Founders, investors and engineers watching AI regulation, data center siting and the boundaries of model capability — especially anyone trying to figure out whether hallucination can actually be fixed.
The clarification around 1:00 about ChatGPT programming itself is fairly basic and can be fast-forwarded.