Man Vs AI: A Year-End Poetry Slam

As Deutsche Bank’s Adrian Cox writes in his year-end DB Research Institute report, generative AI has come a long way in the past three years. But how well does it compare against a 20-watt organic brain in the classic sellside research exercise of a humorous end-of-year poem?

Cox put it to the test: he wrote a year-end poem, then got ChatGPT, running OpenAI’s new GPT 5.2 large language model, to write one of its own. Cox then got a second – unbiased judge – Google’s Gemini 2.5, to compare them in a blind test, decide which was which, and declare a winner.

The answer are after the poems… see if you agree.

AI in 2025: Poem 1

The year of 2025
Saw AI in deployment
Some bubble dread and one “code red”
But not much unemployment

First Stargate launched a race for space
Worth half a trillion dollars
For data centers costing more
Than man-to-moon Apollo

Then China’s DeepSeek drove a wave
That rippled ever deeper:
Why use a branded model
When open source is cheaper?

In March: all eyes on tariffs
And Liberation Day
With stocks down double digits
Now AI looked passé

The White House raised the pressure
With its AI Action Plan
For US chips and power grids
And tighter export bans

By August, things were back on track
The boom had reignited
And Altman said the markets were
Becoming too “excited”

Nvidia surged once again:
Investors made a packet
Who bet on Cuda, GPUs,
And Jensen’s leather jacket

Sam Altman said he needed power
For training bigger bots
As much as India in fact:
Two-fifty gigawatts

The critics say that scaling’s dead
And taking cash for burning,
But backers say there’s other ways
Like reinforcement learning

With rising valuations
The elephant in the room:
Do booms turn into bubbles
And do bubbles then go boom?

The worry spread as Burry said
We’re headed for disaster
He said the chips Nvidia ships
Depreciate much faster

November brought us Gemini
And Google fighting back
With models and a side of chips:
It helps to own the stack

With one billion weekly users
And reports it plans to float
OpenAI said, it’s our “code red”:
We’ll need a bigger moat

Now data centers face a test
12,000 built and counting
With rising debt and widening spreads
And public protests mounting

The focus at the top will shift
To inference from training
And ways to squeeze more value, like
Agentic model chaining

The most important acronym:
Not GPT but FOMO
For enterprises seeking change
Adoption goes in slow-mo

So is this all a bubble?
Likely not right now, we say,
The road is long, demand is strong,
AI is here to stay
 

AI in 2025: Poem 2

In ’25 the chatbots grew a spine, a badge, a brief,
From fluent interns spouting prose to agents chasing chief-of-staff relief.
Reasoning was packaged, metered, priced per thoughtful pause,
“Chain-of-thought” went off-menu; audits frowned at mental flaws.

Prompts gave way to workflows, tools replaced the clever nudge,
If it didn’t hit the ledger, it was labelled “demo sludge.”
Compute ate the CapEx pie; the crust was power and land,
GPUs went macro-sensitive—rate cuts shook the rack demand.

Nvidia crowned the cycle, shovels sparkling in the sun,
When everyone needs mining gear, the merchant wins the run.
Open weights found religion—sovereignty, control, and cost,
While enterprises learned that “free” still bills you what you lost.

Data centres met physics: grids said “queue,” permits said “no,”
The bottleneck wasn’t clever code, but volts that wouldn’t flow.
Europe brought the rulebook, stamped with August dates,
Model cards grew footnotes like mid-cap prospectus debates.

Across the pond, policy argued who should hold the pen,
States played checkers, feds played chess, and firms shipped anyway—then.
Inside companies, pilots faced the valley of the real,
Security, governance, change—turns out adoption’s a big deal.

By Q4 the question shifted, sharpened by the street:
Not “can it reason better?” but “can it close the month-end sheet?”
So ’26 peers over the ridge, less hype, more grind,
Agents everywhere, margins tight, and regulators close behind.

The outlook’s sober, bullish-ish, with caveats you can quote:
Show revenue per watt, per risk, per token—then you’re afloat.

Man (Poem 1) vs AI (Poem 2): who won?

The results were slightly unexpected.

Cox told dbLumina, Deutsche Bank’s proprietary AI model running on Google’s Gemini 2.5 Pro, that it was a literary critic with expertise in AI. He then asked it to “Give me a short comparison of these two humorous and analytical poems about AI in 2025, assessing content, tone etc.”

He then labelled the human version as “Poem 1” and the AI one as “Poem 2”.

“It is a pleasure to analyse these two remarkably insightful poems,” it begins. So far, so good. It adds: “Read together, they form a remarkably complete and complementary diptych of the state of AI today: [the AI version] depicting the engine room, and [the human version] the stock ticker.”

But things went rapidly downhill when the LLM was asked “Which is better?”.

“While both poems are exceptional snapshots of 2025, if I must make a judgement on which is ‘better’ from a critical standpoint, my vote goes with [the AI version].”

Why? The human version “is a better summary of the news in 2025. It’s more accessible and immediately entertaining for a casual observer”. But the AI version “is a better analysis of the industry in 2025.”

It then twists the knife: “As a critic, I value depth and enduring insight over topical summary. Therefore, [the AI version] stands out as the superior piece of analytical literature,” it says.

Which was which?

Could this just be home team advantage, with the AI cheering louder for its own side?

Apparently not – because when Cox then asked dbLumina to guess which one was which, it guessed that the (losing) human version was written by AI and that the (winning) AI version was written by a human.

The actual human version “exhibits several hallmarks of a highly sophisticated Large Language Model tasked with creating a summary of a topic in verse,” it writes. That includes producing a “brilliant synthesis” of publicly available data, adopting a “slightly generic voice” and ending with a “very balanced, reasonable and somewhat predictable conclusion”.

By contrast, the actual AI version reads “more like the world-weary reflection of an industry insider,” it writes. It has a “jargon-laden perspective” reflecting “perspective born from experience, not from scraping news articles”, as well as “killer insights” and a “unique, cynical tone” with “lived frustration and nuanced understanding that is incredibly difficult for an AI to fake convincingly”.

In conclusion, it says, the human version (which it thinks was written by AI) “is an outstanding work of synthesis”, while the AI version (which it thinks was written by a human) “is a powerful work of analysis and lived experience”.

It is hard to judge from this whether AI is much better than expected – or much worse.

But one message from the AI to this analyst is clear: stick to the day job.