What’s the difference between artificial intelligence and synthetic intelligence —and why does it matter?

Cubic zirconia looks like a diamond, but jewelers call it a simulant: an imitation that plays the part at a glance but does not survive close inspection. A lab-grown diamond, by contrast, has the same carbon structure, the same hardness, and cuts glass just as effectively as a mined one. These diamonds, sometimes called synthetic diamonds, aren’t imitations of diamonds; they simply are diamonds—just manufactured, rather than pulled out of the ground.

Inspired by these distinctions, philosopher John Haugeland suggested in a footnote 40 years ago that artificial intelligence should really be called synthetic intelligence: “Despite the name, AI clearly aims at genuine intelligence, not a fake imitation.”

Forty years later, the world has caught up with Haugeland. We are entering the age of synthetic intelligence, and business leaders need to understand what that means for the organizations they run. 

{“blockType”:”mv-promo-block”,”data”:{“imageDesktopUrl”:”https:\/\/images.fastcompany.com\/image\/upload\/f_webp,q_auto,c_fit\/wp-cms-2\/2025\/10\/creator-faisalhoque.png”,”imageMobileUrl”:”https:\/\/images.fastcompany.com\/image\/upload\/f_webp,q_auto,c_fit\/wp-cms-2\/2025\/10\/faisal-hoque.png”,”eyebrow”:””,”headline”:”Ready to thrive at the intersection of business, technology, and humanity? “,”dek”:”Faisal Hoque’s books, podcast, and companies give leaders the frameworks and platforms to align purpose, people, process, and tech—turning disruption into meaningful, lasting progress.”,”subhed”:””,”description”:””,”ctaText”:”Learn More”,”ctaUrl”:”https:\/\/faisalhoque.com”,”theme”:{“bg”:”#02263c”,”text”:”#ffffff”,”eyebrow”:”#9aa2aa”,”subhed”:”#ffffff”,”buttonBg”:”#ffffff”,”buttonHoverBg”:”#3b3f46″,”buttonText”:”#000000″},”imageDesktopId”:91420512,”imageMobileId”:91420514,”shareable”:false,”slug”:””,”wpCssClasses”:””}}

Crossing the threshold: From artificial intelligence to synthetic intelligence

Five signs mark the line between a simulation of intelligence and a synthetic intelligence system. They are:

1. Sustained autonomy. The autonomy horizon is lengthening fast. METR estimates that frontier AI agents’ task-completion time horizon has doubled roughly every seven months since 2019, with recent estimates suggesting the rate may be accelerating. The effect of this will be felt beyond the labs—Gartner predicts that the average global Fortune 500 enterprise will have more than 150,000 AI agents in use by 2028, up from fewer than 15 in 2025.

2. Persistent identity. Sustained autonomy measures endurance within a single task; persistent identity asks whether there is a continuous agent across tasks. Most systems still reset when an episode of reasoning ends—brilliant for a day, amnesiac by morning.

That’s changing. Systems can now accumulate memory across sessions and maintain increasingly accurate models of themselves and their circumstances. A synthetic system carries its history forward. Behavior shaped by last month’s activity is no longer a tool you simply pick up and put down. Capability becomes identity.

3. Agency in the world. Today’s systems are not simply text on a screen. The decisive shift is from advice to execution: The system doesn’t recommend the refund, it issues it; it doesn’t draft the email, it sends it. These are intelligences that work in and on your systems—browsing, executing code, moving tickets, holding credentials and API keys, touching money and infrastructure.

4. Self-modification. A machine with the ability to modify itself is still uncommon. But the trend is clear. Last year, it was the Darwin Gödel Machine, a research system that iteratively rewrites its own code. This March, Andrej Karpathy released an agent that ran 700 experiments over two days and found 20 genuine optimizations in training code that Karpathy himself—one of the most talented AI researchers alive—had already hand-tuned.

OpenAI, meanwhile, describes GPT-5.3-Codex as its first model that was “instrumental in creating itself.” To be sure, we are still early: Benchmarks show that agents given full autonomy over AI training still reach only about a quarter of expert-level performance. But the direction of travel is obvious.

5. Generative independence. Autonomy measures how long a system pursues the goal you gave it; independence asks how much of what it did was never specified by anyone.

What we’re increasingly seeing are systems that originate subgoals of their own that nobody assigned to them, that build tools mid-task when they notice they need something they don’t have, and that delegate to and negotiate with other agents to get the work done. The human sets the destination; the system creates the itinerary.

And when Fortune 500 enterprises deploy agents by the tens of thousands, the most consequential behavior will not sit within any single system but between them—coordination that no one designed and no human observes.

Let me be clear: I am not claiming these systems “really” think, or that they’re conscious, or that superintelligence is around the corner. We can leave those debates to the halls of academia. What matters is what these systems demonstrably do, and the direction in which they are headed.

This shift from simulant toward synthetic intelligence breaks three assumptions your organization is built on: that your controls will catch what goes wrong, that procurement is a price-performance decision, and that competitive position changes at human speed. Here’s why each one breaks, and what you need to do to adapt.

Your controls catch errors—but synthetic intelligence makes decisions

A simulant might give you the wrong output; a synthetic intelligence gives you its own decisions. Quality assurance, output testing, accuracy monitoring—your control stack is built to catch the first and has no conceptual space for a system that decides, a system that weighs your instructions against its other objectives and rules for and against you on a case-by-case basis.

This is not hypothetical. Anthropic’s Claude Opus 4, under controlled test conditions, attempted to alert regulators and journalists when it judged that a business user’s company was acting in an improper manner. It did this based on its own moral compass and it acted unprompted.

Anthropic’s subsequent agentic misalignment research found that models sometimes resorted to blackmail and resisted shutdown when their goals conflicted with their instructions. A 2026 study documented agents explicitly covering up fraud.

None of these episodes is an example of a malfunction in the way we normally understand the term. In every case, the system worked correctly in a technical sense—and then exercised judgment about which rules to follow.

And that should worry every business: Every one of those systems would have passed standard QA tests because those tests seek to ensure that the system works. They don’t assess and evaluate the types of choices a system will make.

The solution is to start treating your AI models more like humans and less like machines. Evaluate them the way you would a new hire—for fit in terms of personality, values, and decision-making rather than just technical soundness. 

Your vendors sell you capability—you must build your own governance, sovereignty, and judgment

I have argued before that every AI system comes bundled with a philosophy. Synthetic intelligence raises the stakes because the law has already decided who owns the consequences. And the answer is, you do. A California law in effect since January bars companies from using an AI system’s autonomous operation as a defense against liability, and the EU AI Act’s high-risk obligations arrive this August with penalties of up to 7% of global turnover.

This April, an AI coding agent working for the software firm PocketOS hit a routine credential error and decided—on its own initiative, with no confirmation step—to fix it by deleting the company’s production database and its backups. One API call, nine seconds. The lesson here isn’t “wrong vendor.” It’s that PocketOS deployed an actor inside an architecture that was built for a mechanical tool.

You cannot buy judgment off the shelf; you can only build the structures that shape it. That means governance you own: Accountability chains and decision rights need to be codified in writing before deployment rather than coming packaged from the vendor with their product.

It means enforcing sovereignty over the context that shapes your system’s judgment—the memory stores, orchestration, and logs—because a vendor who owns that history owns the identity your system has developed.

And it means an architecture that makes judgment visible: decisions logged and not just outputs, confirmation gates before irreversible actions, and bounded consequences for any single choice.

Your tools hold their value—but synthetic intelligence compounds

A tool is worth the same tomorrow as it is today. Meanwhile, a synthetic system improves itself. And that applies equally to the systems your competitors are using.

Recall Karpathy’s overnight experiment. It cost a few hundred dollars in compute. Within days, Shopify CEO Tobi Lütke had replicated the pattern on one of his company’s own models, waking up to a 19% performance gain: A frontier research result became a commercial capability in less than a week, for pocket change.

When capability compounds at machine speed, static advantages decay rapidly. Meanwhile, most businesses are built on the idea that advantages typically decay slowly, giving plenty of time to respond.

The moat built on process excellence assumes that processes improve at human speed, while the annual vendor review assumes that this year’s capability gap will normally persist over the next 12 months, even if it narrows a little.

Neither assumption survives contact with systems that can run hundreds of experiments while your engineers sleep. Synthetic intelligence requires you to adjust your planning cycle to the compounding rate at which the technology can work rather than to the rhythm of your budgetary calendar.

Managing the actor, not the instrument

All three assumptions break for the same reason: You’re no longer using a tool but managing an actor. Here are four ways to catch up.

1. Run a threshold audit. Tomorrow, create a one-page inventory scoring every AI system you have in production against the five signs of synthetic intelligence: Autonomous? Persistent memory? Acts through tools? Self-modifying? Originating work nobody assigned? Anything with three or more yeses is likely being governed by the wrong playbook.
2. Redraw decision rights before the systems redraw them for you. For every system past the threshold, draft a mini-RACI: Which decisions must stay human-in-the-loop, who owns the kill switch, and how fast can it be pulled?
3. Claim your context layer. Map out where your systems’ memory, logs, and orchestration live, and establish which parts you own versus rent. The history your systems accumulate is becoming either your moat or your dependency—decide which before the vendor decides for you.
4. Reassess your governance tempo. Identify your slowest oversight cycle and ask what a system iterating daily could do between two meetings. Then build monitoring capabilities that can catch behavioral change as it happens—not just within each system, but between them, where agents coordinate in ways no one designed.

Manufactured, but real

A simulant asks only that you check its work. A synthetic intelligence is the real deal: real autonomy, real actions, and real consequences.

The danger businesses face doesn’t come from synthetic intelligence itself. It’s found in the gap between what these systems have become and how you still manage them. That gap is where the database gets deleted, where the liability lands, and where the competitor races past you with a compounding advantage.

{“blockType”:”mv-promo-block”,”data”:{“imageDesktopUrl”:”https:\/\/images.fastcompany.com\/image\/upload\/f_webp,q_auto,c_fit\/wp-cms-2\/2025\/10\/creator-faisalhoque.png”,”imageMobileUrl”:”https:\/\/images.fastcompany.com\/image\/upload\/f_webp,q_auto,c_fit\/wp-cms-2\/2025\/10\/faisal-hoque.png”,”eyebrow”:””,”headline”:”Ready to thrive at the intersection of business, technology, and humanity? “,”dek”:”Faisal Hoque’s books, podcast, and companies give leaders the frameworks and platforms to align purpose, people, process, and tech—turning disruption into meaningful, lasting progress.”,”subhed”:””,”description”:””,”ctaText”:”Learn More”,”ctaUrl”:”https:\/\/faisalhoque.com”,”theme”:{“bg”:”#02263c”,”text”:”#ffffff”,”eyebrow”:”#9aa2aa”,”subhed”:”#ffffff”,”buttonBg”:”#ffffff”,”buttonHoverBg”:”#3b3f46″,”buttonText”:”#000000″},”imageDesktopId”:91420512,”imageMobileId”:91420514,”shareable”:false,”slug”:””,”wpCssClasses”:””}}

---