Fat Bear Week and the Fate of the World
EIG vs ChatGPT-5: an ursine battle to decide the future
Forget about whether AI is coming for your job. Can AI even hold its own in a single, deeply human, deeply absurd task?
Like Kasparov versus Deep Blue, the team at EIG has recently immersed itself in a contest of strategy and prediction against a powerful and unfeeling machine: ChatGPT-5. But the venue for our clash with the embryonic Terminator is not a chessboard in Philadelphia but rather the frigid and salmon-teeming rivers of Alaska.
That’s right, the battleground is none other than Fat Bear Week.
What is Fat Bear Week?
Every October, Katmai National Park runs a public online bracket to crown the fattest, roundest, most winter-ready bear.
Think March Madness but with half-ton carnivores. Rangers introduce each bear with photos and backstories. The public votes, round by round, until one champion emerges.
What does this have to do with the economy?
Every job is a bundle of tasks. Each task can be thought of as falling on a spectrum that runs from routine to non-routine.
Routine tasks are explicit, codifiable, and have well-defined procedures, making them more susceptible to automation. Non-routine tasks are the opposite, hard to codify and thus to automate away.
Completing a Fat Bear bracket, which involves predictions about aggregate human behavior in a uniquely frivolous context, fits squarely in the non-routine bucket.
In other cases, the boundaries between routine and non-routine tasks can sometimes be blurry. This is where AI comes in.
Early progress in computing and digitization kickstarted the automation of a subset of routine tasks where the quality of output was high and the cost of integration was low. Computers and specialized software took a bite out of routine work. Accountants no longer burned through pencils and paper ledgers, and administrative assistants no longer shuffled index cards of phone numbers and addresses.
One way that AI affects workers is by shifting the boundary between routine and non-routine tasks. Indeed, early evidence on the impacts of AI show firms put greater emphasis on reallocating tasks than laying off workers. As the boundary between routine and non-routine tasks shifts we should expect workers to focus more time and effort on non-routine tasks in which they have a comparative advantage.
What was once not codifiable — like predicting which fat bear has that special je ne sais quoi to win over the internet — is now something AI might be able to do.
Pitting ChatGPT-5 against an office of knowledge workers in the fat bear bracket is our way of helping to suss out the modern boundary between what is routine and what is not, what is codifiable and what is not.
Our totally robust experiment
We started with a warm-up, the Fat Bear Junior bracket, where cubs and yearlings face off before the adults compete. Using ranger notes in addition to pre- and post-fattening photos, ChatGPT-5 scored each cub on heft, roundness, posture, photo angle, and notes. Using the bracket provided by the Rangers, we then let ChatGPT-5 make its picks. We also made our own.
Our AI competitor picked the chunky 2.5-year-old cubs of Bear 803. EIG staff and the public picked the yearling of Bear 128, which ended up being the eventual Junior winner. An early but unofficial victory for us carbon-based life forms.
After feeding the results of Fat Bear Junior back to ChatGPT-5, it opted to incorporate new measures into its fat bear index: storylines and charisma. The AI had “realized” (its words) it wanted to include these new measures and added a Narrative/Charisma factor, which it described as a “meme coefficient,” before attempting the adult competition.1
Armed with its revised scale, ChatGPT-5 rated all twelve competitors in the adult bracket. You can see those scores at the end of this post.
The results
Our jobs as Fat Bear Week forecasters are safe, for now:
We won! ChatGPT-5 finished dead last, ending with 7 total points out of 18.
Here are the results broken down by round:
In Round 1, ChatGPT5 correctly picked only one out of four matchups. The model favored bulk — Bear 503, Bear 26, Bear 856 — over charisma, producing early losses.
In Round 2, though, the machine managed a perfect 4-for-4, correctly calling Bear 32 (nicknamed Chunk), Bear 128 (nicknamed Grazer, the winner both of the past two years), Bear 602, and Bear 856.
In the semifinals, ChatGPT-5 split 1–1, correctly advancing Chunk but failing to account for Bear 856.
In the final, ChatGPT-5 thought Grazer could get the three-peat, but dynasties do end and the public crowned Chunk over Bear 856 with a vote of 96,350 and 63,725, respectively.
We’ve organized these results into a convenient bracket format:
Bottom line: AI showed flashes of potential but never threatened the humans. Fat Bear Week brackets appear to remain squarely in the non-routine column.
The bigger economic angle
When people consider the changes of AI displacing work, it is important that they not think in binary terms, with AI replacing a whole job or none of it. But as we’ve been emphasizing, jobs are accumulations of discrete tasks, and each of these tasks varies in its exposure to AI.
AI is likely to accelerate the automation of structured routine tasks, pushing workers and firms to focus their employees on non-routine tasks and coming up with new tasks altogether.
Non-routine work — tasks that remain non-codified or have distinctly human elements — is thus likely to increase as a share of human time and effort on the job. As it should: it is where our comparative advantage will endure the longest.
For now, we humans are just happy that we can still better predict which chubby bear will steal the hearts of other Fat Bear Week enthusiasts. In at least this one area, the machines don’t yet know us quite as well as we know ourselves.
APPENDIX
The bears of the bracket
Here are the AI’s rankings, in its own words:
Fat Bear Week 2025 — Scoring Table
1. 26 (F, daughter of Holly, mother of spring cubs)
Strong motherhood story, new litter after loss, daughter of Holly. Solid fatness but not yet top-tier in sheer heft.
Composite Score: 3.8
2. 32 “Chunk” (M, very large, broken jaw survivor)
One of the largest males (1200+ lbs). Legendary resilience after breaking jaw in 2025. Charisma/story weight high.
Composite Score: 4.8
3. 99 (M, young rising adult, strong fishing, first Fat Bear appearance)
Upcoming star, still growing into top tier, but voter appeal is limited without past wins.
Composite Score: 3.7
4. 128 Grazer (F, two-time champion, aggressive mom, legendary)
Defending champ from 2023 & 2024, mother of this year’s Fat Bear Junior winner. Narrative juggernaut.
Composite Score: 5.0
5. 503 (M, large, amiable, adopted by Holly, gentle giant)
Strong size, friendly reputation (“social male”), big recognition story. Lacks raw intimidation vs. Chunk/856.
Composite Score: 4.2
6. 602 (M, big male, “floatato,” dominant but comic charisma)
First-time Fat Bear appearance, large and dominant, fun charisma but less public recognition.
Composite Score: 4.0
7. 609 (F, young female, daughter of 909, former Fat Bear Junior champ)
Charismatic lineage, strong independence story, but lacks sheer bulk compared to older females/males.
Composite Score: 3.6
8. 856 (M, elder dominant male, past top bear, aging but crafty)
Former river king, aging but still formidable. Strong nostalgia/legend factor.
Composite Score: 4.5
9. 901 (F, medium-large female, resilient after lost litter)
Popular, recognizable, fat and healthy, but less charisma than Grazer/909.
Composite Score: 3.9
10. 909 (F, daughter of Beadnose, mother of 609, lost cubs, strong angler)
Voter appeal through family legacy (Beadnose), resilience after cub loss, fishing fame.
Composite Score: 4.1
11. 910 (F, sister of 909, skilled angler, separated from cubs in 2025)
Strong fishing, legacy bear, but overshadowed slightly by 909 and Grazer.
Composite Score: 3.8
12. 128 Jr. (F, Fat Bear Junior 2025 champion)
Fresh off Junior win, strong charisma, but small relative to adults.
Composite Score: 3.5
Some of the exact language used by ChatGPT-5:
For the adult competition, I’ll adjust my framework:
Keep Heft/Size as an important factor.
Add Narrative Weight (survival story, unique traits, or popular parentage).
Incorporate Public Visibility/Charisma (cuteness, meme potential, popularity in past years).
This way, my predictions won’t just track the fattest bear by biologist standards, but also what resonates with voters.
It's not just AI that'll take American jobs. In a globalized world, workers that make significantly less are threats to American jobs. I believe "significantly less" to be half or less. If Trump throws up tariffs or build walls it will make prices more expensive for Americans. If he blocks immigration, companies will have to outsource to solve worker shortages and high American wages.
We Americans are so used to our way of life that we don't see how inefficient, wasteful and costly we live. It all starts with our cars. Each is a 2-ton hunk of metal, used one-hour a day to move 200 lbs of flesh that costs us $12K a year. That's 95% of fossil fuel burned to move metal and 5% to move passenger, 5% usage and 95% parked and cost more than the average Mexican's salary. But there's more. Each American home dedicates garage and driveway (24'x24' + 24'x24') space equivalent to a 3BR apartment to our cars. Our 2 cars go to work, and our workplace have parking/driveway space that is about the same. If we factor in front/backyards of our homes and parking/driveway of all the places we drive to, we see that our housing shortage is because we need to have so much space for our cars and so don't build enough for humans to live. Our 2 biggest costs are housing and transport.
One way to fix this is to convert 30-60% of parking lots of workplaces with more than 200 to dense, affordable apartments for those that work there. These housing would not have individual driveways or garages, front/backyards. Instead, there can be common green spaces and a small pool of shared vehicles. Workers that live there can give up their cars ($12k/year) and 300 hours/year of commute. Many efficiencies work well at scale such as solar panel/battery, geothermal and we can share just about anything that we currently store in our garages.
Other efficiencies can build on these 5-minute communities such as the "other spouse" working in a support role such as an in-community market or daycare. These 5-minute communities can cut American workers' living cost and carbon footprint in half and make our labor more competitive.