>> We turn now to the growing debate over artificial intelligence.

The heads of some of the world's biggest AI companies have warned of catastrophic consequences from the technology they are building.

But computer science professor Cal Newport says that messaging is misleading.

And he's joining Hari Sreenivasan to discuss why he believes an AI industry should stop what he calls the "doom trolling."

Cal Newport, welcome back to the program.

Your most recent piece says, "Dear AI companies, the doom trolling needs to stop."

First of all, what does that mean?

Who is it, who needs to stop doing what?

- Well, doom trolling is my term for this completely strange and arresting and novel behavior of AI companies trying to convince their customers that the products that they are creating are potentially going to cause massive devastation or other negative consequences down the line.

I think OpenAI does this.

Anthropic in particular is a big practitioner of doom trolling.

And it was actually a recent report that Anthropic released talking about how they were concerned that Cloud Code was on a path towards recursive self-improvement that could lead them to lose control of AI altogether.

That finally had me snap and say, "This type of communication strategy, this has to stop."

- What is the purpose?

What can be gained by using a message like this where maybe they feel like they're factual, but at the same time, like, how do you create a market for somebody who wants to use this product?

- I mean, there's many possible explanations, and I don't know which one is the most prominent driving the strategy, but partially, it does make your technology seem more important.

You care a little bit less about exactly how much revenue you're making.

When you consider this company might be producing the most powerful tool that's ever been built, it's also a recruiting tool.

So in San Francisco culture, where a lot of the top engineers currently live, this numerous mindset of AI becoming this harbinger of a new digital end times, is actually really, really prevalent.

So speaking this language could help you recruit.

It could also be, as some have suggested, regulatory capture strategy, where you say this is a dangerous technology that needs regulations, and you hope that those regulations are such that the big companies can abide by them, but it holds back your competitors from being able to make progress.

But I think more than any of those other explanations, it's just in the Silicon Valley culture to talk this way about AI.

I think they have completely normalized this idea that machines at some point are going to perhaps even replace humanity or like significantly change our existence and that this is not necessarily a bad thing.

It's quasi-religious for a lot of people in Silicon Valley culture right now.

It's eschatological.

It is about the sort of future of what's going to happen.

And I think the rest of the country is just waking up now to just how sort of strange and eccentric these type of belief systems have been in Silicon Valley, because now we all have to face them.

And we're saying, "What are you talking about?

You're going to destroy us all.

Why would you build this?"

But that's a completely normal idea if you're over in that part of the country.

Okay, you're a professor of computer science.

Should we be looking at this technology just like a technology?

Or look, I mean, Sam Altman, the head of OpenAI, has compared this to nuclear power.

And that comes with a whole other set of risks and rewards that we think about.

I mean, is this that grand venture that we're about to go on where we need a completely different frame of looking at it than an operating system upgrade or a new tech gadget.

I mean, I do think it's something we do need to be careful about.

But when I say we need to be careful about AI, I think about the long trajectory of this technology, not in particular the tools that are being built right now, which are typically large language models with various harnesses or programs connected to them.

But I think the right way for them to deal with this technology is like a normal technology.

Here is a product, here are the benefits, here are the costs, here's why we think those benefits are worth the cost.

And of course, we take full responsibility for any safety concerns.

If you approach the product that way, you're able to advance the technology without accidentally stumbling into something that could be more dangerous.

That's the way I wanna see us talking about is product by product, treat this like a normal technology, and we should all be very wary about potential dangers and harms, including the companies themselves.

These are things to take responsibility for, not things to just shrug your shoulders at and say, well, what can we do?

- You know, just in the past few weeks, there's been this back and forth between Anthropic and the Pentagon, and there's been concern about this new model that they came out with, which they said was so dangerous that we're only gonna give it to a few different companies around the world, and we wanna make sure that everything's set up.

And then now it almost seems like this campaign of saying that this is so powerful, this could crack the encryption of banks and find all these different vulnerabilities and holes, has worked to a point where the administration says that no foreign national is allowed to use this tool, right?

Is this so dangerous that the kind of the regulations that they were asking for, did you get basically what you said you should?

I do think they got what they were asking for, right?

Now, I believe as a computer scientist who has studied these models that Mythos did not represent a revolutionary jump over previous models when it comes to its ability to find software vulnerabilities or exploit them.

That's a serious concern of models, but it was something that we've seen in every model going all the way back to GPT-2.

So I think they, to make this model seem more exceptional six weeks ago or seven weeks ago, whenever this was, I think they turned up the rhetoric on just how dangerous this was.

We can't release this model.

They had meetings at the White House to try to convince the White House how scary this model was.

They had meetings with reporters.

And then six weeks later, they said, oh, it's okay.

We added some guardrails.

It's fine again.

So yeah, I think in some sense, the White House was maybe embarrassed or upset that, hey, you came to us and convinced us this was the most dangerous thing to have been released in years, and now with sort of just minor standard guardrails are releasing it to the public, we feel like we were a little bit duped.

Now, there's other things that are probably going on in this story, but that thread is one that we have to pull, is that if you tell people that you have summoned a superweapon, you don't also then get to say, "It's $20 a token," or whatever, "Now we're going to release it to the public."

In a perfect scenario, if I had total trust and faith in the government to say, "You are looking out for my best interests and nothing else, and you've taken this very well thought out and careful approach to..." I don't necessarily know if the people who are at the levers, whether it's this administration or the next one or the next one after that, can I separate their alternate agendas, their other interests from regulating this kind of a technology?

Well, I share your concerns about the current administration, that clearly they're not implementing this sort of regulatory oversight in a consistent or transparent way.

There's all sorts of connections with the administration and Anthropics competitor, OpenAI.

So that's a confusing mess.

But I do think going forward in the future, is it possible to have consistent and transparent regulation of these type of AI models?

I think absolutely, because we do it with so many other sectors of the consumer product market.

I mean, you can't sell me a car if the government hasn't given it official safety rating.

If you put something dodgy in dog food, we are going to push back and say, "This is unacceptable."

So I think it is possible, and I think it's time or will be time soon to move past this stance of this is somehow an exceptional technology.

I think that companies like to talk about this like they are the reluctant stewards of an inevitable technology, and they're sort of just watching from afar as this thing develops, like virologists watching COVID-19 spread across the country.

But this is not the case.

These are companies that are building specific products for specific business plans.

It's not an inevitable technology that they're stewarding, it's products they're building.

And you need a third party, as we do with all other consumer products, to say, "If this is potentially very dangerous for the American public, we want to be involved."

And I want to say that danger should involve mental health.

I think we're completely underestimating the toll, the mental health toll, in terms of anxiety and stress that the last two or three years of doom trolling has caused.

I get messages from people who are miserable because of this messaging they're hearing from these AI companies again and again, and that matters as well.

- In a way, the messaging we're talking about here has succeeded in shifting public opinion.

Right, I mean, it seems like forever ago, but it wasn't that long ago when the first one of these generative AI models came out.

People were like, wow, this is pretty cool.

And now it's like, okay, you're telling me that it's going to totally displace white collar work.

You're telling me that this could potentially turn into that Terminator scenario.

You were telling me that it uses gobs and gobs of water and energy and it's going to make my bills go up and make my water in the aquifer dirty.

Now the majority of people are cautiously skeptical.

Yeah, which I think is a tragedy.

Like this should have been an exciting technology.

Large language models at scale are interesting.

They can parse human language.

They can produce structured language with a sort of prodigious fluency.

There's a lot of cool things we can and will build with this.

And instead, we terrified the whole country.

Now, the fact that they did this leads me to believe that this is not some grand game of 4D chess.

But again, it's a collision of worlds, a way of talking in Silicon Valley that doesn't play at scale.

Let me give you an example.

For most of the fall and coming into the winter, there was a relentless drumbeat on this message of white collar jobs are going away, right?

And it led to all this coverage of like, what are we going to do when there's no white collar jobs left, right?

Because they're all going to go away.

And then around the time Anthropic and OpenAI started talking to bankers about an IPO, we had this sudden turnaround this spring where suddenly Sam Altman said, "I was wrong about AI taking jobs and I'm glad to be wrong."

And you have Jensen Wang from NVIDIA saying, "This is stupid.

This is just CEOs trying to sound smart."

And we had even Dario Amadei said, "I know for two years I've been saying 50% of new white collar jobs will be automated.

I didn't really mean that.

I meant parts of those jobs will be automated, not the jobs themselves going away."

So they switched hard on that message, but then because now you had a lack of whatever, interesting stories or fear, whatever they were looking for, they leaned hard into the recursive self-improvement AI superintelligence message.

So as one message went away, they found another one.

So it is pretty erratic.

One of their concerns, the AI companies, and sometimes they wrap themselves in kind of a nationalist cloak and say, "Look, the genie's kind of out of the bottle in the sense that the technology is out there.

Our frontier models might have an advantage, but there are national competitors, specifically China, who's advancing at their own pace.

Is that a gap that we should be concerned about from both a kind of technological advantage for our national security as well as just, gosh, is there a kind of a cheaper, better mousetrap that's coming for free?

Well, I mean, first I would say that there is no accepted principle of consumer product safety that says you can release a product that's going to cause harms if another country is releasing that product as well, right?

We don't make exceptions for causing harm to the populace based on, you know, economic races or geopolitics.

So I don't really buy this argument.

We have to be pushing towards recursive self-improvement and losing control of AI because China might do it as well, right?

Geopolitical issues are not the job of individual companies to try to solve, right?

But I think there's a broader issue.

Yes, cheaper, smarter models are coming.

I see the frontier models that the major labs are working on equivalent to like a Formula One race car.

It's the $20 million car that you build to try to show off how tech-forward your company is, and maybe some innovations will trickle down a few years later to the sedan that the average person can buy.

But I think in the future, and this is just my technological prediction, we're not going to be using on a regular basis these massive frontier models.

These are just to show, it's a leaderboard to show how advanced your company is.

For most of the uses we have for large language models right now, from natural language searching to software development and support, you don't need a 10 trillion parameter model like Claude Mythos to do this.

You could use a well-tuned 50 billion parameter model with a smart harness on it.

I think that is going to be the future.

I think these companies want to IPO before we get to that future, so they'll have the capital to actually adjust, but there is disruption from below that's going to come.

I wonder if we didn't learn enough from our societal entanglement with social networks and social media, right?

Because when you say that, you know, look, we, most consumer products can't be released if we know that they're dangerous.

And here we are, even in the relatively early stages of generative AI, we already have cases of chatbots just really hallucinating in the worst way possible.

I don't know, what are we waiting 15 to 20 years to figure out what kinds of harms can be there?

I mean, how should we navigate this from a legislative perspective or from a global perspective?

I mean, I think we did take way too long to understand what social media, but now we finally are.

And because of that, there will be a much smaller window and a much smaller amount of sort of trust or leeway we're going to give the AI companies, right?

So there's a similar playbook that both are trying to pull.

So in the heyday of social media in the 2000 teens, the playbook that the social media leaders use was to say, this technology is inevitable.

This is the evolution of communication.

It's the digital town square.

We are just the stewards of this technology, but obviously it needs to be here and it's gonna have some harms, but to stop it would be like trying to stop the printing press.

We have finally had enough of that argument.

I think the recent losses in courts with Meta and Google and the hundreds of lawsuits that are coming behind those has shown that from a litigation perspective, the court system is saying, "You're responsible for harms.

You cannot hide behind this is a fundamental communication technology that can't be restricted.

You can't fully hide behind the First Amendment."

I think that has profound implications for the AI leaders because they are also trying to say, "This is just an inevitable technology.

If we're just doing our best to try to steward it, the courts could step in and say, "No, you're liable."

We're starting to see this.

There was an important ruling recently at a court in Germany that said LLM creators are responsible for the text that the LLMs produce.

You can't say that text was the LLMs, we just created it.

If that becomes an international precedent, we might have much tighter constraints on these AI companies much faster than it took for social media.

We trusted the Silicon Valley to do this.

We let so many harms build up over a decade to 15 years that I think we're going to have to deal with.

I think we're going to have to deal with it.

I think we're going to have to deal with it.

I think we're going to have to deal with it.

And we let so many harms build up over a decade to 15 years that I don't think we're going to make that same mistake again.

Do you see this trickling down into like, you know, I'm on WhatsApp groups with different parents about the influence of AI and technology in elementary and middle and high schools, right?

They're wondering like, wait, wait, are we rolling this out too soon?

Is this actually going to stunt my child's ability to solve critical problems and have good ability to think?

So I wonder like what in the next five years, 10 years, when students come to your classroom, is there gonna be a difference in how they think about solving problems?

- I mean, I do think it's an issue to which we need a national solution.

We shouldn't leave individual schools and school districts at the mercy of the sales forces of these ag tech companies.

They're all gonna be telling them, if you don't sign a big deal to get Gemini access to your fourth graders or whatever, that somehow they're gonna be left behind in the modern economy.

It's really difficult for individual schools and school districts to resist that.

So I think we need national standards from non-governmental agencies that are saying, this is what we actually recommend.

Because I do think it's a problem.

My biggest concern is actually writing.

I mean, I think that the production of words from on a blank page, just using your brain is one of the most cognitively demanding and cognitive growth enhancing activities that we do.

And it's one of the core things we do in education to make your brain stronger.

To have a AI model right for you, I think it's like bringing a pulley system to the gym to lift the weight for you.

It completely defeats the purpose of the institution.

And so we need strong guidelines about when AI is and is not appropriate, because otherwise, we will just be taken down district by district, school by school, by relentless ed tech marketing and sales.

- What's an appropriate diet, or what's an appropriate amount of skepticism, cynicism to have when they hear information about this technology?

I mean, I honestly think the reality right now is that if you're not a software developer or work in sort of like certain fields of mathematics, you really don't need to have much of an involvement in a large language model-based AI tool in your life.

I mean, maybe you want to do some occasional natural language searching, right?

Where you want to explain your search query in like natural language.

But I think it's something that most people actually right now do not need to be engaging with that seriously, which I think is a real problem for the AI companies.

And I think kids don't need to have access to chatbots at all.

I would be incredibly wary about this.

This is not some skill that they need to become conversant in like typing, which requires a lot of training.

Using a chatbot is just typing in natural language.

There's no skill to learn.

The dangers are large.

The goal for a child is to develop their ability to think at this point, not to try to be efficient.

So I would be very wary.

See, look, if you don't have a major use for AI in your life, don't go searching for one.

You don't need it.

It's not your job to figure out for the AI companies why their tools are useful.

And I would keep my kids largely away from AI as long as they were still living within my house.

- Professor of Computer Science at Georgetown University, Cal Newport.

Thanks again.