This article is the first in what will be an ongoing inquiry. The subject is enormous. AI ethics is a puzzle too vast to solve in one sitting, but we can begin mapping its outlines: the illusion of inevitability, the ethics of restraint, and the possibility of choosing another path.
On Outshining AI, we’ll keep exploring that path, wherever it leads.
How important is it for us to get artificial intelligence right? How crucial is it that we figure out how to govern AI so it is fully constrained to do good—rather than evolve into something that manages us… or worse.
Obviously, this is as crucial as it gets. But is “fully constrained to do good” even remotely possible? After all, aren’t “bad” and even “very bad” applications of AI inevitable?
Some at the cutting edge of AI research are asking the same questions. One of them is Tristan Harris, a technology ethicist and co-founder of the Center for Humane Technology. Harris believes we cannot afford to buy into what he calls the myth of inevitability.
Tristan Harris first came to prominence warning that social media’s “race for attention” would fracture societies and erode individual well-being—a warning that, looking back, feels prophetic. Now, in his 2024 TED Talk “Why AI Is Our Ultimate Test and Greatest Invitation,” he warns that we may be repeating the same mistake, only on a much greater scale.
His central question is as simple as it is unsettling: What if the greatest danger isn’t what AI can do, but what we believe we can’t stop?
A power without precedent
According to Harris, AI’s promise is unmatched. “Imagine there’s a map and a new country shows up on the world stage and it has a million Nobel Prize level geniuses in it… [who] don’t eat, don’t sleep… work at superhuman speed… for less than minimum wage,” he says. “That is a crazy amount of power.”
Applied well, it could bring “truly unimaginable abundance.”
“But what’s the probable?” he asks. AI’s power now eclipses every other technology, and investment in AI tech continues to increase at an unprecedented scale. What are the odds these AI applications will be built with care unless we design rules and rewards that favor serious caution and rigorous transparency, as well as painful penalties when rules are ignored? That’s the core question Harris raises.
Yes, used ethically and safely, AI could accelerate discovery of life-saving medical treatments, new clean energy solutions, and all manner of wonder materials.
Harris calls this “the possible.”
The essence of his premise isn’t about dreaming up possibilities. It’s about what happens when the work the market rewards collides with the most powerful technology in human history.
Let’s imagine what those more probable paths could look like.
Online engagement without guardrails
Picture a video platform whose revenue lives and dies on minutes watched. Now give the platform’s owner an AI that can spin up any deepfake of a face or a voice on demand.
On tightly moderated platforms with strong proof-of-origin requirements and disclosures, YouTube among them, much of this gets caught. But on faster-moving short-video feeds or closed networks, where credits are missing and trace-back tools are weak, a different pattern emerges. The recommendation engine learns a blunt lesson: outrage and novelty hold attention.
On tightly moderated platforms with strong proof-of-origin requirements and disclosures, YouTube among them, much of this gets caught. But on faster-moving short-video feeds or closed networks, where credits are missing and trace-back tools are weak, a different pattern emerges. The recommendation engine learns a blunt lesson: outrage and novelty hold attention.
And they pay.
The result? An assembly line of believable, personalized deepfake shorts: a “revelation” about a celebrity, breaking “news” from a TV personality who isn’t real, a friend’s cloned voice telling a half-truth—all designed to hook viewers.
As advertisers buy ads where viewers spend the most time, money flows to clips that rack up watches and clicks—not to straight-shooting storytellers or honest brokers of information. In short, the system rewards watch time, not integrity and accuracy.
The safety net that never lets go
It begins with an accident. A runaway trading algorithm triggers a chain reaction that erases trillions of dollars in value overnight. The public demands protection. Within months, regulators, banks, and tech companies join forces to build something that can “see the next crisis coming.”
They call it the Safety Cloud.
It’s a global system linking payment networks, data centers, communications platforms, and banks under a shared goal: prevent harm before it happens. The first version of the Cloud watches for fraud and financial abuse. Later updates add content screening to reduce panic, and behavioral alerts for signs of “escalation.”
At first, the new AI-powered oversight feels reassuring. Transactions clear faster. Conspiracies fade from view. Emergency rooms see fewer overdoses tied to viral misinformation. Every update brings smoother safety metrics.
Then small things begin to happen.
A journalist’s access card stops working at the border. A doctor’s license renewal is delayed without reason. A parent’s post about a side effect disappears after “further review.” Support tickets go unanswered—not from neglect, but because no one is authorized to overrule the system’s decisions. The logic behind each block is locked behind “safety compliance protocols.”
People adjust. They rewrite their posts, speak more carefully, and avoid topics that might “trip the filters.” Businesses redesign websites to stay inside the recommended tone range. Writers learn to test their drafts against AI “risk checkers” before publishing.
No laws are broken. No one is arrested. Yet over time, the room for error—and for dissent—shrinks. The AI-driven network doesn’t punish; it simply withholds. What was once open access becomes conditional trust, automatically updated, and continuously scored.
Most don’t notice the change. The system keeps learning. And because it rarely makes a visible mistake, no one dares question its grip.
Why this is a real possibility
Because the early pieces of such a system already exist. Financial systems use predictive tools to block suspicious transactions. Hospitals use automated triage to approve or deny treatments. Social networks use ranking models to slow the spread of rumors and “hate speech.” Stitch those systems together—and add one layer of shared oversight after a global shock—and the Safety Cloud practically builds itself.
Its power comes not from malice but from momentum. Each sector adopts it for convenience. No one company owns it; every company connects to it. Once millions rely on it for credit, care, and communication, it becomes too risky to ever switch off.
Transparency fades by design. As the models grow more complex, few can explain their reasoning. Appeals become automated, then circular. The system is always right, because it defines what “safe” means.
Where this is likely to show up first
Finance and insurance. Anywhere money moves, risk scoring already is at work on a daily basis. Expanding those models to assess “systemic safety” would be a natural next step after a major geopolitical shock.
Healthcare and biotech. Predictive models now assess everything from cancer risk to mental stability. A unified safety network could merge these with public health surveillance “for the common good.”
Global communications. Large platforms already cooperate to filter “harmful” content. Adding shared trust scores would make moderation more uniform—and less visible.
Infrastructure management. As AGI begins to coordinate power grids, transport, and supply chains, access itself could depend on safety scores, locking entire communities out of basic services during “elevated risk” periods.
This is the kind of control that doesn’t need intimidating enforcement, because it works through compliance. It’s a net woven in the name of safety, pulled tighter by every new lesson learned, until one day it’s no longer clear who’s being protected, and from whom.
That, too, is “the probable” if we pursue safety without equal attention to transparency, individual rights, and basic accountability.
The myth of inevitability
When people say today’s AI is “inevitable,” they usually mean something simple: Artificial intelligence keeps showing up in more places, whether we’ve asked for it or not. One week your phone suggests AI replies in texts. The next week your email, search box, and work software add new AI buttons. And because these tools spread fast through app stores and plug-ins, it feels like the only choice is to keep up or be left behind.
Harris pushes back on that scenario. He asks: If literally no one wanted this pace or pattern, would the laws of physics still force it on us? His point: this isn’t fate, it’s a set of business choices we could make differently.
But why does it feel unstoppable? It’s because everyone believes everyone else will keep racing. Labs tell themselves, “If we don’t move first, a rival will,” so they speed up—and their rivals do the same. Product developers do the same.
When inevitability becomes insanity
However, Harris vigorously protests this rush to market. “We’re releasing the most powerful, inscrutable, uncontrollable technology we’ve ever invented,” he says, and doing it under maximum pressure to cut corners. “This is insane.”
History shows we can choose otherwise, Harris argues. Banning above-ground nuclear testing and phasing out ozone-killing chemicals didn’t happen by accident. Nations acted once the risks were clearly understood.
Harris’s point is the same here: before we can practice wise restraint with AI, we have to be willing to examine with scientifically rigorous clarity the risks of staying on the path we’re on now.
The narrow way forward
If “let it rip” leads to chaos and “lock it down” leads to Orweillian control, the job is to steer a middle way of AI development and deployment where power is matched with responsibility. That means shaping AI, not freezing its progress. Some steps Harris recommends:
- Put skin in the game. If an AI product causes foreseeable harm, the maker should be held liable. Strong product-liability rules incentivize safer releases up front, instead of conjuring up remedies later after AI genies are already out of the proverbial bottle.
- Label what’s synthetic and make it traceable. Simple “who made this and when” tags and basic trace-back tools would help slow the flood of fakes without blocking artists or researchers.
- Protect kids from manipulative AI companions. Set strong, clear guardrails so these systems can’t steer children toward harm.
- Back whistleblowers. People shouldn’t have to risk their livelihoods to warn the public about real AI-related risks. Strong whistleblower protections make honest and timely warnings possible.
That’s the beginnings of a “narrow path”: common-sense steps that line up with how people actually behave—and with how powerful these tools really are.
What wise restraint looks like in practice
Harris’s point about wisdom is simple: real wisdom includes restraint. Or as he puts it, “There is no definition of wisdom… that does not involve restraint.”
Here’s how that could look in the real world: short, concrete moves that slow the right things without stalling the good:
- Stage releases—then earn the next step. Start small, with a limited group, clear “kill switches,” and outside testers. Widen access only after a model passes agreed-upon checks (not just internal ones). Harris’s larger point: don’t ship first, fix later.
- Make risks common knowledge. Share what goes wrong, so all AI developers learn faster and the public isn’t left in the dark.
- Draw hard lines where risks are greatest. For example, kids’ AI companions need strict guardrails. Ubiquitous surveillance should be off-limits. These are two of the places where restraint matters most.
- Back the people who speak up. Strong whistleblower protections let insiders warn the rest of us without losing their livelihoods. This is part of the “collective immune system” Harris calls for.
In short: wisdom isn’t just a state of mind. It’s also a set of default rules: go small first, study, learn and react to risks up front, build brighter boundary lines, demand and enforce accountability.
Choosing coordination over competition
We’ve done this before. When the risks were clear, nations coordinated—nuclear test bans, genome-editing pauses, phasing out ozone-destroying chemicals. None of that was “inevitable.” People and nations chose it once the science and stakes were understood.
Harris’s call is to do the same with AI: swap the race mindset for shared guardrails and a “collective immune system.” In practice, that looks like a few simple, collective moves: regular briefings on frontier risks, baseline safety checks before wide releases, real paths of appeal when systems go wrong, and protections for those who speak up. The idea is to make the risks common knowledge so we can act together, not react alone.
The invitation before us
Tristan Harris frames this moment as both warning and welcome. It’s a test of whether we can match power with responsibility. It’s also an invitation to do so. “Your role,” he says, “is to be part of the collective immune system… to say that this is not inevitable.”
At Outshinging AI, we’ll keep tracking the narrow path with open eyes, and reporting on what works and where it fails. No single article settles it. But each step toward full risk assessment, real accountability, and practical restraint moves the “probable” toward the “possible.”
Or, as Harris puts it, AI is “humanity’s ultimate test and greatest invitation.” Let’s answer it wisely.
Primary keyword: responsible AI development Supporting keywords: AI ethics, AI restraint, AI governance, AI inevitability, responsible AI
Leave a Reply