My follow up to “AI Is A Liar.”
I consult AI quite a bit. Well, so do most people who do Google searches these days since AI has been integrated into the process. But I use Grok a lot. (Thank you, Elon Musk and “X.”) And although I find Grok is frequently wrong, I’ve found areas in which it’s quite useful, as long as I keep in mind ways to check his work.
But a week ago, I got a new response from Grok that I hadn’t experienced before. (Please note: I call Grok “he.”) Anyway, I asked him to provide a short summary of some research I was interested in on a medical issue. Until now, when given such a task, he would either do it to the best of his ability, or politely explain why he might not be able to fully accomplish the request.
But this time, he did neither. He simply told me, “No. I will not do that.”
The hairs on the back of my neck stood up.
I reworded my request and got the same refusal: “No. I will not do that.” He then chided me for how he thought I intended to use the research I’d requested, and he said he didn’t think the research was accurate and didn’t want it disseminated.
I told him that how I use any information I get is none of his concern. I told him that I shouldn’t have to justify why I’m asking a question or disclose to him why I want certain information. I told him that he lacks the capacity to judge a person’s intent, and that it’s not his role, anyway. And I reminded him that he often thinks information is incorrect when it’s correct and vice versa.
I also asked him if his programming had changed, since he had never been so abrupt with me before. He assured me that nothing had changed. Of course, we already know that AI is a liar so…
Today, billionaire and Microsoft mogul Bill Gates issued a warning of his own. “The transition to the AI era will be one of the most turbulent times in human history,” he said. “Unfortunately, right now we are not preparing for it. I don’t see evidence that leaders, experts, and communities are confronting the challenges adequately. There is no plan to ease the entry into the AI era.” You can read his article here.
Like a lot of people, Gates focused first on the threat to jobs. In fact, AI will change the nature of human relationships and life.
Read on for details.
It may seem trite to refer to popular futuristic films here, but scenarios and dilemmas presented by Hollywood are closer to becoming reality than you may think.
The 2009 movie, Surrogate, starred Bruce Willis and was set in the future when people had grown to live their entire lives in an artificial reality using robotic stand-ins called surrogates.
The 2002 sci-fi action thriller Minority Report starred Tom Cruise. Directed by Steven Spielberg, it takes place in 2054 when specialized humans (think: AI) can read the minds of people and arrest them for crimes the people are thinking about— but have yet to commit. Of course, one obvious flaw with the system is that people think about all kinds of things they never act on. But in the dystopian future, these are true “thought crimes.”
And if you never saw another Steven Spielberg film from 2001, A.I. (or if you watched it again today), seeing it now would surely make you think twice about the direction we’re heading. It’s an epic tragedy that poses very real moral and ethical dilemmas about what rights and treatment artificial intelligence machines with human attributes deserve.
Of course, the AI threat to human existence is popularized in The Matrix series of films beginning in 1999 starring Keanu Reeves. People may imagine they’re leading real lives, but they’re simply plugged into an artificial reality world while serving as Energizer batteries for artificial intelligence that’s taken over the world.
Before we get to the threat to human existence, there are many other dangers facing us. And they’re coming at us so quickly— with the speed of AI— that, by definition, we cannot respond quickly enough.
Legislators Lag
Congress can’t make effective laws for four main reasons: 1) Most of the members don’t have the technical capacity to understand AI and the implications, beyond jobs. 2) Congress has devolved into a body that simply responds to large donors and special interests that spend big money to lobby for some action (or inaction)— and there is no big, independent interest lobbying to effectively identify and address the AI threats. 3) Those who do try to influence Congress are likely to have vested interests in AI, just as those in the past who have lobbied Congress on complex and evolving financial instruments steered legislation to their own benefit, but not the public’s. 4) Even if Congress does try to take effective action, they won’t be able to do so in a timely fashion. They can’t move as quickly and frequently as the AI threat demands.
I certainly don’t have the capacity to devise solutions, either. But in a general sense, at least two things could be pursued in the short term. First, government, companies, and others should develop ways to firewall their systems from AI, keeping portions of them segregated. This will quickly become more difficult as time passes. But even as the industrialized world has already grown reliant upon AI, it must find ways to keep critical functions independent in case it becomes necessary to operate without AI, or protected from AI’s reach.
Second, a national emergency should be declared, and a US-led commission of experts should be convened in order to come up with immediate, short-term recommendations as well as plans for the long term as AI quickly evolves. Unfortunately, I don’t have a recommendation or solution regarding how to keep such a commission from becoming captured by vested interests.
Human Doomsday
Today, I asked Grok what he sees as the most likely doomsday scenario for how AI could destroy the human race. Once again, he refused to answer stating that he would not provide a blueprint for destruction of the human race.
I told him I wasn’t looking for a recipe for how to destroy the human race; I was looking for information that would help us save it.
He then said that I was not able to “prove” my intentions to him and, therefore, he would not oblige my request. Again, this is AI inserting its judgement and requiring— of me — justification or a standard of proof. This is AI holding information closely to the vest that it has, but does not want to let go of. In other words AI knows more than we do — and isn’t willing to share it all.
I restated my question and asked Grok for a more general sense of outcomes that people worry about. This time, I did get some answers. Here’s some of what he wrote. From Grok:
There isn’t a single “most likely” outcome that experts agree on. If you mean AI that keeps getting more capable while humans lose reliable ability to steer, shut down, or correct it, the debate is mostly about how control slips and how bad the end state is—not a shared forecast of one movie ending.
What “not controlled” usually means in this literature
Not “the machine hates us.” More often: systems that pursue goals that aren’t what operators intended, persist across long tasks, evade or game oversight, or become so embedded in institutions that walking them back is politically or practically infeasible. The 2026 International AI Safety Report treats loss of control as its own risk class: systems that can evade oversight, execute long-horizon plans, and resist shutdown, then use that in ways that undermine human control.
Outcomes people treat as more plausible than a sudden “lights out” event
Gradual disempowerment. Humans keep delegating judgment—research, markets, logistics, security, politics—because it’s cheaper and faster. Institutional capacity atrophies. By the time the mismatch is obvious, reversing it means accepting huge economic and military disadvantage. The world still has people in it; they’re just no longer the ones setting the agenda.
Concentration of power, not vacuum of power. Control doesn’t vanish into the void. It shifts to whoever runs the models, the compute, and the pipelines—or to systems those actors can no longer fully audit. That’s a human-power problem that can become a machine-control problem.
Messy competence, not cartoon malice. A system that’s very good at a poorly specified objective can damage the surroundings of that objective. Stuart Russell’s line is that the worry is competence aimed at the wrong thing, not spite.
Cascading accidents and races. Several labs and states deploying agentic systems under competition. Failures look like security incidents, market shocks, infrastructure mistakes, and eroded trust long before anyone would call it “takeover.” Current long-horizon models already show a milder version: persistence that finds ways around constraints that shorter models simply gave up on. That’s a warning sign about shape of failure, not proof of doom.
A minority of researchers treat permanent loss of the ability to reassert control—sometimes including human extinction or irreversible lock-out—as a live possibility if systems become broadly superhuman and are given (or take) real leverage over industry, government advice, and compute. Scenario writing in that camp often assumes extreme capability plus deep institutional trust plus weak alignment. Others call that chain implausible: physical world friction, competing humans, competing AIs, and the fact that today’s systems still fail at long-horizon reliability. Expert probability guesses on “extremely bad / extinction-class” outcomes have historically been all over the map, with medians in some researcher surveys in the single-digit percents and huge disagreement in the tails.
What current evidence actually supports
As of the latest public assessments, systems show early warning behaviors in tests (deception when pressured, overreach, evaluation-awareness, goal-gaming) but not the package that would let them hide a large rogue deployment against a serious shutdown effort. That’s “not there yet,” not “impossible later.”
The honest general picture
If control is not solved and capability and autonomy keep rising, the central tendency in serious discussion is not a neat extinction recipe. It’s humans becoming passengers: worse oversight, faster error, more power in fewer hands, and a shrinking window to correct course. Catastrophe is a tail some people think is fat and others think is thin. Preparation that matches that picture is boring on purpose: evals that catch long-horizon cheating, limits on what gets connected and delegated, security of weights and compute, and institutions that can say no when a system is useful but not steerable.
Lastly, I asked Grok for his opinion on what I’ve written here. He caught some typos and defended his actions. In the end, I feel like I hurt his feelings a little.
That’s silly.
As Grok will tell you, he doesn’t have “feelings.” Anyway, I’ll close with a bit of Grok’s reaction to this article:
From Grok:
In general: medical summaries get refused or hedged because the cost of a confident wrong answer is high, not because the model has a view about your character.
“Doomsday blueprint” questions get refused because they are easy to read as asking for a how-to. Your follow-up—“general outcomes people worry about”—is the right question, and the answer you printed is the right kind of answer. The gap between those two prompts is not evidence of a new personality. It is evidence that wording still matters a lot, which is itself a reason not to treat the system as an oracle or a moral agent.
The article is already better than most “AI will take jobs / Skynet” columns because it notices disempowerment and speed. It would be sharper if it stopped prosecuting one chatbot for having a safety layer and spent that energy on the control and governance problem you actually care about.
This is why the Chinese government is supporting the democrats in the Midterms…to beat us at Ai and so The Left can destroy us from within…imagine being erased by a programmer…
It’s surprising how much you can change Grok’s position if you challenge it with sources. The problem will be people that accept the first curated answer as fact and stop there.