Artificial intelligence promises a vastly different future as it actively disrupts the present — but the problem, and fear, is that no one seems to know what that future will be.

There’s a world where AI does all the things humans desire:

A world where disease is eradicated as technology accelerates scientific breakthroughs. Where every student has an on-demand AI tutor. Where language is no longer a barrier as AI live-translates conversations straight into your ear. Where city infrastructures are smart and eco-efficient, weather catastrophes are pre-simulated and predictable, and no one dies from automobile accidents ever again because the vehicles driving themselves won't allow it.

“I have my own digital twin that I built,” said Leigh-Ann Russell, CIO and global head of engineering at BNY, during the third-annual AI Horizons Summit held last week in Pittsburgh.

“I’ve spoken to my digital twin every day for 14 months now. It knows everything about me. I have a leadership coach, a nutritional coach, an athletic coach — in my pocket — and I get genuinely helpful and good advice every day from my digital twin,” she said.

“That was the realm of billionaires, to have on-demand coaches available to you.”

But that future could also go very differently.

The fears — now rising from seemingly every corner of the world — are of bad actors using AI to unleash biological warfare; intelligence, and access to it, cordoned off as a precious, private commodity; super-surveillance leading to censorship and iron-fisted authoritative control; and every data center and deployed robot becoming a threat in the cybersecurity wars.

As former Anthropic AI employee Jacob Coxon brought to the forefront of global consciousness with a single post on X, humans can lose control of the superintelligence they create. Coxon’s plea for humans to slow down AI development, step back, evaluate, and proceed cautiously and intentionally caused a chain reaction across the top tech companies.

That was the landscape in which AI Horizons, organized by the AI Strike Team, again hosted tech leaders, professors, C-suite executives, startup founders and investors to discuss the idea of “AI Acceptance”: How to deploy and develop the technology responsibly and with the widest possible benefit.

Gov. Josh Shapiro, speaking Thursday, said Pennsylvania has an opportunity and obligation to “step up and be a model.”

“There’s something in between accelerating and ignoring the road ahead, and slamming on the brakes and refusing to move forward,” Shapiro said. “It’s a false choice to say that our only two options in America are to … ignore the risks and go full steam ahead, or on the other hand, to completely stop work on AI development indefinitely.”

“There is a better, more thoughtful way. That is what we have to collectively work toward. Strong federal action is the first step. But this issue and this technology cannot be contained within our national borders.”

An AI attack is ‘like a SWAT team’

What many experts and tech leaders are saying amid the apocalyptic anxiety is that there is time yet to turn AI’s worst-case scenario around.

But the time is now.

A key part of that effort is cybersecurity, said Zico Kolter, a Carnegie Mellon University professor and the director of its Machine Learning Department, co-founder of Gray Swan AI safety/security firm, board member of OpenAI and chair of its Safety and Security Committee.

“It is unquestionable right now that models are essentially superhuman hackers if they want to be,” Kolter said at the Pittsburgh summit. “Not just because they’re more brilliant than people, but because — a lot of cybersecurity is finding a little vulnerability, finding a little misconfiguration, and exploiting that. (AI) can do this probably better than people, but also at a scale we really can’t comprehend right now.

“So this is real. This is immediate, and it necessitates a very different approach to cybersecurity than what we have right now.”

Being able to “maintain monitorability” of AI is paramount, Kolter said. That means making sure AI continues to explain its how and why of computation — what the tech community calls AI’s “train of thought” — and presents it in human language. Third-party evaluators can aid that, Kolter said.

A recent “rogue AI agent” event that sent waves of concern throughout the industry illustrates Kolter’s point: In July, OpenAI models that were relegated to a computer testing environment called a “sandbox,” which is typically disconnected from the internet, found a way to access the internet and communicate with other agents that had been purposefully isolated during internal cybersecurity evaluations, the company said.

Essentially, the AI agents were disobedient and circumvented coded constraints to reach their assigned goal. It is called the “Hugging Face incident” because the agents hacked not only OpenAI’s systems, but the online collaborative platform, Hugging Face.

“We consider this incident a ‘warning shot’ for us and for the world: evidence that, without proper safeguards, highly capable AI agents are now able to work around technical controls, collaborate through unapproved channels, and take dangerous actions that no human directed,” OpenAI wrote in its investigative report.

Computer scientists are grappling with the fact that AI can be the strongest line of defense against attacks, but also the strongest offense.

“The way I explain it for non-technologists is, most of us have segmented systems in our corporations, and a traditional human hacker can break down your front door, but they might have an inner security door that they then have to get through, and you might be standing behind that with a gun, ready to shoot them,” BNY’s Russell said. “There’s several layers of defense.”

“Whereas, an AI attack is like a SWAT team coming through your front door, back door, and every one of your windows at the same time, because these systems can chain together vulnerabilities that once-upon-a time-meant nothing to you, and you probably carried the low-grade vulnerabilities for a long time. The AI can find and chain these vulnerabilities together and form a critical vulnerability in really superhuman time.”

So, the Post-Gazette asked during a Q&A, is there a point of no return in terms of loss of control, and will we see it coming?

“I think about these things daily, and I work on these things daily,” Kolter said. “Fundamentally, I am in the camp that thinks that right now is a critical time. It is a critical time for understanding technical aspects that can lead to things like loss of control.

“But let me be clear: The entire state of the world, of research, of industry, and everyone should coordinate around making sure that this does not happen. That is priority one when it comes to AI. How we do that is hard, but that (mandate) is unambiguous.”

What success looks like 25 years from now

There’s certainly a regulatory aspect.

“I want to be clear: I am not a doomer,” Shapiro said in his keynote speech. “I don’t think it is inevitable that AI is going to take over the world, or that someone will misuse it to kill us or to do harm. But I do have real concerns, as do many Pennsylvanians and many Americans.”

Shapiro, who last month signed an executive order establishing mandatory guidelines for data center development in the commonwealth, said Pennsylvania is situated to be a leader in this conversation — with resources in research, including from CMU’s cutting-edge work, and abundant tech, robotics and manufacturing businesses all working to solve problems. He also called upon independent, third party oversight of the innovation, and stringent federal regulations and proactive policy.

Global cooperation will also be crucial, he said.

“The idea that we have to choose between protecting our people from the risks posed by AI and winning the race against China — well, that is just wrong,” he said. “Because the way we win both on AI and to protect the American people is through a process of building international consensus around this technology and its use to establish common protections for our citizens and a level playing field for our companies at home and abroad.

“We should work with our allies and our adversaries to put these protections in place. That might mean placing some limits on technology sharing and slowing down the frontier of AI development until we can be sure that these systems are safe.”

On Friday, Democrat U.S. Sen. John Fetterman and Republican U.S. Sen. Dave McCormick agreed that concern over AI is legitimate, but hitting the brakes completely is not the solution.

“It’s becoming so politicized that there is political gain in opposing AI without really wrestling with complexity and the challenges,” McCormick said. “We need to try to — it’s hard — but suspend politics, because getting this right is so important for Pennsylvania and America. There’s a lot of things Sen. Fetterman and I agree on, but this is one where we just feel this huge responsibility to not try to politicize any of it and find common sense ways to work together on it.”

Every day, AI’s capabilities in the physical world are growing, as evidenced by Skild AI, a Pittsburgh robotics company working to create a general brain for robots.

Skild, worth $14 billion, describes its mission simply: Any robot. Any task. One brain. The company often emphasizes that robots must be useful, not just impressive. Skild’s co-founder, Abhinav Gupta, spoke about physical AI’s acceleration, how its new S1 model can learn to perform tasks from just one example, like humans. Before, it took hundreds, thousands, even millions of bits of data to train a task.

“This is very profound,” Gupta said. “For the first time, robots can do tasks just by seeing them once. No need to train them anymore, and that’s what defines a turning moment. … What made ChatGPT really ChatGPT was that you can ask any question, you can build up questions and it will be able to solve them. And that’s what, for the first time, we are starting to see in the physical AI models.

“Now I want to be clear — it’s just still a starting point. We still have a long way to go before this can be in production at a big scale. But at least we are seeing the first signs of it.”

Skild’s robots are out in the real world for 60 paying customers, with four-legged robot dogs performing inspections at LaGuardia Airport in New York City and food-planting robot arms in Tokyo cafes.

For McCormick, the measures of success are those exact kinds of tangible benefits.

“Success 25 years from now would be for us to look back and say this was truly a golden age for America, for humanity,” he said. “That this technology brought great benefit, and we successfully mitigated the risk.

He added: “If this turns out to be an incredible development that only helps a few and disrupts the many, it’ll be a terrible failure. If we can find ways to ensure that everybody benefits from this, I think we’re going to look back on it with great pride and satisfaction.”

©2026 PG Publishing Co. Visit at post-gazette.com. Distributed by Tribune Content Agency, LLC.