Nothing New Under the Sun: AI, Genesis, and the Limits of Machine Creativity

We usually tell the human story as a climb: from cave to city to cloud, each generation smarter than the last. I don't believe that.

Answers in Genesis, a creationist ministry, makes the case I find most convincing. Humans were "created with intelligence right from the beginning," and since then "this intelligence appears to be decreasing rather than increasing" (Hadsall, AiG, 2025).

The turning point was the Fall. "Sin came into the world through one man, and death through sin" (Romans 5:12). As AiG describes the creation model, the original human genome "contained no errors," but it "has gradually degraded over time" as harmful mutations accumulated (Smith, AiG, 2025). Disease entered our genes. Then Noah's flood cut humanity to eight people, a genetic bottleneck, and lifespans fell off a cliff (Purdom & Menton, AiG).

The Bible records that decline in plain numbers:

AiG connects the same decline to our minds: "Given the accumulation of deleterious mutations, this can easily impact intellectual capability" (Hadsall, AiG). I believe our ability to create, innovate and think sharply has been fading along with our years.

So why does it feel as if we keep getting smarter? Because while each of us got weaker, our libraries got bigger. Each generation inherited what the last one had written down and added a small piece of its own. Most inventions since Eden were a step next to an earlier one. The truly new part, the delta, kept getting smaller.

That brings me to AI.

A large language model is, in a sense, the last stop on that road: the whole library of human writing compressed into one system. And this fall, it did remarkable things with that library.

Source: Lex Fridman Podcast #503, "AI Crisis in Math" (transcript), Oct 26, 2026.

• Verifying the largest proof in history. The Classification of Finite Simple Groups, the "periodic table of symmetry," runs about 15,000 pages and was "built by over 100 mathematicians across generations" (Axiom). The Bulletin of the American Mathematical Society calls it a "110-year project" (Solomon, 2001). Axiom says its AI agents rewrote the entire proof in Lean, a language computers can check line by line. They wrote 50 million lines in 24 days (Axiom). So far, that claim comes only from Axiom.
• Chipping at century-old problems. On September 8, OpenAI published a paper claiming that the Navier–Stokes equations, which describe how fluids flow, can "blow up" in finite time (OpenAI). On October 6, it released 722 manuscripts from an unnamed internal model (719 after withdrawals), organized into 372 result families. They came from about 4,000 problems, at roughly three hours of computing per result (OpenAI; openai/math). One of the results is the "quasi-Riemann hypothesis," a step toward a problem open for more than 160 years (Quanta). Another is a claimed proof of the unique games conjecture from computer science (Quanta).
• Challenging a 27-year-old assumption. Computer scientists have long assumed that puzzles like 3-SAT (finding an on/off setting for hundreds of switches that satisfies every rule) can never be solved without the time exploding. MIT's Ryan Williams reports that an OpenAI model has produced a claimed refutation. The assumption's own authors are organizing an independent review (Williams).

That's 24 days of machine work against generations of human work. I don't want to downplay that.

But look at what these results actually are. The Axiom project verified human knowledge. It created none. OpenAI's own Navier–Stokes paper says its construction relies on "a smooth force compactly supported in space and time" (OpenAI). Scientific American calls that a "loophole," a "weird, unphysical force" that stirs the fluid just right (Scientific American). OpenAI withdrew three papers the day after release because of a sign error, and revised fourteen others (Retraction Watch). OpenAI itself says that "some of the unformalized results could have issues" (openai/math).

The quasi-Riemann result is the most telling. Experts largely agree it is correct, though Princeton's Peter Sarnak won't call it "a Theorem" until humans check and understand every detail. Quanta reports that the proof required no new concepts. Mathematician Alexander Dunn called it "a surprising synthesis of well-known ideas," and it builds on a 2021 formula he developed with Maksym Radziwiłł (Quanta).

A synthesis of well-known ideas, in other words. Solomon saw this three thousand years ago:

"What has been is what will be, and what has been done is what will be done, and there is nothing new under the sun." (Ecclesiastes 1:9)

For AI, that verse is close to literal. Everything it produces is drawn from what is already under the sun. Even the strongest result answers a question humans asked, judged by a yardstick humans built.

The mathematicians closest to this work see the same gap. On Lex Fridman's podcast, MIT's Ankur Moitra said the OpenAI solutions "just don't look like a human has solved those problems." Rutgers' Alex Kontorovich admitted, "I can prove a theorem, but I don't understand why the proof works." The University of Washington's Steven Brunton put it most simply: "Humans are still asking the questions." Then he added, "I want it to start telling me what questions I should be asking" (Lex Fridman Podcast #503 transcript).

It isn't doing that yet. Asking the new question is the part that is still ours.

I didn't reach this conclusion from headlines alone. It came from pushing the latest frontier models hard on creative work of my own.

Here's one example: I wanted party invitations that were totally unique and trendy, and I had some design ideas of my own. I asked every one of the latest image-generation models to come up with something creative.

Prompt after prompt, no matter how hard I tried, the AI struggled to come up with anything on its own. I had to give it specifics: what to make, how to make it, which theme to use. Only after several rounds did I get something close to what I had in mind.

That's been my experience in most cases. When the work is truly unique, AI needs a human with the creative idea, spelled out clearly and in detail, before it can do anything with it. The model executed my idea. It never handed me one.

None of this is news to anyone who scrolls a feed. Most of the time, you can tell an AI-generated image or video at a glance, with no other context. It isn't because of a stray sixth finger. It's because the piece feels plain and generic, a repetition of something you've already seen. It feels fake.

Ironically, AI is much better at copying reality. In one online experiment with more than 12,500 participants and about 287,000 image judgments, people correctly identified AI images only 62% of the time, and they struggled most with landscapes and city scenes (Roca et al., 2025). Imitating what already exists is exactly what these models are built for. Imagining what doesn't exist yet is where they give themselves away.

My newest ideas rarely come while I'm working on a problem. They show up early in the morning, just after I wake up. They show up in the shower, or in the middle of some mundane task. Sometimes a light bulb goes on while I'm hearing or watching something on TV, and it connects to something totally unrelated that I've been turning over. A conversation from months ago, a comment from someone in a completely different field, and something I read last week suddenly fit together, and I can't explain how.

Research suggests this isn't just me. In a 2025 preregistered study published in Scientific Reports, participants wrote a short story, took a 10-minute break, and then wrote again. The more their minds wandered during the break, the more their writing improved when they came back to the same story. Thinking deliberately about the story during the break didn't produce that benefit. Only letting the mind drift did (McDaniel, Habibi & Kaplan, 2025). Science can show that the brain keeps weaving in the background. It can't fully explain how.

David had a simpler explanation:

"I praise you, for I am fearfully and wonderfully made." (Psalm 139:14)

"So God created man in his own image" (Genesis 1:27). God didn't only shape Adam from dust. He "breathed into his nostrils the breath of life" (Genesis 2:7). When God finished creating, He called it "very good" (Genesis 1:31). He enjoyed what He made, and He gave us a share of that joy: the ability to make something that didn't exist before.

That is why I believe the human brain is so unique and impossible to replicate. Even fallen, even dimmed by six thousand years of decline, it still carries that breath. A machine trained on our words carries only the words. Or, as Kontorovich said of his own field, "the value of pure mathematics is that it lives in the heads of human beings" (transcript).

I know this view isn't fashionable this week. The tech world has started calling this month "Black October," after a run of AI-linked security incidents (CNBC), and my feeds are full of "fast takeoff" talk. Frontier models are now so capable at finding software flaws that some are being held back: since April, Anthropic has limited its Claude Mythos model to vetted security defenders (Fortune).

Look at what actually happened. AI-assisted research surfaced PHP2Shell, a pair of flaws in the PHP programming language (Wiz). Among websites whose server-side language is known, about 70% use PHP (W3Techs). Some of the underlying bugs had been known for years (CNN). Then word of the fix leaked, and people rushed to exploit it. NPR's headline said it best: "AI found the flaw. Hackers found the leak" (NPR). Every link in that chain was a human choice. AI made each step faster. It didn't make any of them.

Even Anthropic, which builds some of these models, says so in its IPO filing: "bad actors have already conducted high-profile cyberattacks using AI's capabilities to discover vulnerabilities." The same filing also warns that increasingly autonomous systems could "take harmful actions independently and at scale without human intervention" (Fortune). That second worry is where I part ways with Anthropic and its CEO, Dario Amodei: I believe the danger is the person holding the tool.

We've seen this before. At Babel, God's concern wasn't the bricks. It was what united, ambitious people would do with them: "nothing that they propose to do will now be impossible for them" (Genesis 11:6).

So amid all the doomsday messaging, I'll say it plainly: I don't believe AI will become sentient, and I think we are overhyped about it. A system that recombines what we've written down is not a mind that wants things. The worst outcomes of AI progress will most likely come from malicious people misusing it, not from AI taking over the decisions. That is where our attention belongs.

AI will keep getting better at computation. It will aggregate and synthesize what we already know into new results, and some of those results will matter. I'm genuinely excited about faster proof-checking and faster research.

But in nuanced, original creative work, the kind Jubal did with his first song, I believe AI will remain very limited. I don't think it will reach the extent of human creativity, or come anywhere close.

It reads the scrolls of every age, it learns each line on every page; but only dust that God once breathed can dream a song no one conceived.

The machine can stand on the shoulders of every generation before us. Only humans were made to stand where no one has stood before.