AI is a Petulant Teenager (For Now)

Everywhere I look these days there’s another headline about artificial intelligence: either techno-utopia or Terminator 2: Judgment Day. We’re either supposed to bow and celebrate its godlike might or tremble in fear of our new robot overlords.

Look, I get both the excitement and the trepidation. Really, I do. Can we pause for just a second, though, and look at where AI stands today?

Sure, maybe someday AI will evolve into something that makes us look like ants. Today? Right this moment? I think it’s worth recognizing that AI is perhaps best understood as a petulant teenager. You know the type: means well most of the time, but will look you dead in the eye and lie about finishing its homework.

I’m not an AI nay-sayer. I’m just saying what most of us have already discovered: it’s a long way from Terminator 2.

Now there’s some fascinating research showing just how strangely human some of these mistakes can look.

The Lazy Genius

A recent study from Apple called The Illusion of Thinking looked at a newer breed of AI systems known as Large Reasoning Models, or LRMs.¹ These are models designed to spend additional computing power working through a problem before answering. Think of Claude 3.7 in Thinking Mode or OpenAI’s o3. In simple terms, instead of immediately blurting out an answer, they spend some time working through the problem first.

You’d assume this would make them better at reasoning. Give the AI more time to think and it should do better, right?

Well, sort of.

The researchers gave the models a bunch of classic logic puzzles, including Tower of Hanoi, Checkers Jumping, River Crossing and Blocks World. They didn’t pick these because anyone particularly cares whether Claude can move wooden disks between three pegs. The puzzles were useful because you can make essentially the same problem harder and harder in precise steps. That gave the researchers a way to watch what happened as the models went from “this is easy” to “okay, this is getting difficult” to “good luck with that.”

This is where things got interesting. On relatively easy problems, the fancy reasoning models sometimes did worse than standard models. They could find the right answer, then keep reasoning and wander into unnecessary alternatives. Give them something moderately difficult, though, and all that extra thinking started paying off. This was where the reasoning models were at their best relative to their simpler counterparts. Then the researchers kept turning up the difficulty. Eventually, both kinds of models fell apart. On sufficiently difficult versions of the puzzles, accuracy dropped to zero.²

Okay, every system has limits. That part isn’t terribly surprising. What happened next is. At first, the harder the problem became, the more reasoning the models used. Makes sense. Harder problem, more work. Then, beyond a certain level of difficulty, the relationship suddenly reversed. As the problems became even harder, the models started using fewer reasoning tokens, even though they still had computational budget available.¹

Think about how odd that is. Give the AI something easy and it can overthink it. Give it something challenging and the extra effort helps. Then give it something really hard and, just when you’d expect it to buckle down and work harder, it starts doing less.

If you have teenagers, this may sound familiar.

I’m not saying the AI is actually getting bored or frustrated. What’s interesting is the behavior. It looks surprisingly familiar even if whatever is producing that behavior is nothing like what’s happening inside a human brain.

Fine. We’ll Tell You How to Do It.

The researchers tried something else. Maybe the models weren’t failing because they couldn’t reason. Maybe they just couldn’t figure out the right way to solve the puzzle. Fair enough. What happens if you take that part away and simply tell them how to do it?

For Tower of Hanoi, the researchers gave the models an explicit algorithm for solving the puzzle. Now the model didn’t have to discover the procedure. It had the procedure. All it had to do was follow it.

That helped in some cases. It didn’t solve the problem. As the puzzles became more complex, performance still deteriorated.³

I find this particularly interesting because there are actually two different abilities hiding inside what we casually call “reasoning.” One is figuring out how to solve a problem. Another is reliably carrying out the solution once you know how. We tend to treat those as the same thing. Apparently, they aren’t. At least in these experiments, giving the AI the right procedure didn’t guarantee that it could keep executing that procedure as the task became more complicated.

Again, oddly human.

It Wasn’t Just One Study

The Apple study wasn’t the only research raising questions about how these models reason. Another group of researchers created a benchmark called LogiEval, which tested reasoning models across deductive, inductive, analogical and abductive reasoning.⁴ Rather than relying on a few logic puzzles, they used different kinds of questions drawn from human exams including the LSAT and GMAT.

The results weren’t that AI is simply bad at logic. That would be easy, and wrong. The models were extremely good at some kinds of reasoning, particularly certain argument analysis and analogical reasoning problems. Other kinds of reasoning gave them considerably more trouble. The researchers then created an especially difficult subset called LogiEval-Hard. Even the stronger reasoning models continued to struggle there.⁴

There’s an important wrinkle, though. When the researchers compared the models’ mistakes with the mistakes people make on similar problems, the patterns didn’t simply match.⁴ So I don’t want to take the teenager analogy too literally. These systems aren’t synthetic people. They may be arriving at surprisingly human-looking mistakes through processes that are nothing like ours.

That almost makes it stranger.

AI is a Petulant Teenager

For most of computing history, we’ve gotten used to machines being almost the opposite of people. Computers are literal, predictable and tireless. Give a computer the same calculation a billion times and we expect the billionth calculation to work just like the first. People are the ones who overthink easy problems, rise to a challenge, then sometimes fall apart when things become overwhelming.

Now we’re building machines that can do things that look a lot more like reasoning, and some of their mistakes are starting to look oddly human. They can overthink the easy stuff. They do seem to get tired. Give them something challenging and they buckle down and do better. Push them far enough and their performance falls apart. In Apple’s experiments, they even started putting in less reasoning just when the problems seemed to require more.¹

Again, none of this means AI is secretly experiencing boredom, frustration or teenage angst. We don’t know that, and the research certainly doesn’t show it. The interesting thing is that we built something profoundly nonhuman and, as its capabilities have become more sophisticated, some of its failures have started looking remarkably human.

For most of computing history, we built machines partly because they didn’t make mistakes the way we do. They could repeat the same operation endlessly without getting tired, distracted or overwhelmed. Now we’re teaching machines to do something much closer to reasoning, and apparently we’re getting some surprisingly familiar mistakes along with it.

AI isn’t human, but it can overthink the easy stuff. Rise to a challenge. Then, when things get really hard, seemingly stop trying. For now, perhaps, AI is a teenager.

Endnotes

¹ Shojaee, P., Mirzadeh, I., Alizadeh, K., Horton, M., Bengio, S. & Farajtabar, M. (2025). “The Illusion of Thinking: Understanding the Strengths and Limitations of Reasoning Models via the Lens of Problem Complexity.” Apple. arXiv:2506.06941. https://arxiv.org/abs/2506.06941

² Shojaee et al., “The Illusion of Thinking.” The researchers identified three performance regimes across the puzzle environments: standard models performed better at low complexity, reasoning models had an advantage at moderate complexity and both experienced complete accuracy collapse beyond certain levels of complexity.

³ Shojaee et al., “The Illusion of Thinking.” The researchers also tested models after providing explicit algorithms for Tower of Hanoi and Checkers Jumping. Providing the algorithm produced some improvements but did not prevent performance from deteriorating at higher complexity.

⁴ Liu, H., Ding, Y., Fu, Z., Zhang, C., Liu, X. & Zhang, Y. (2025). “Evaluating the Logical Reasoning Abilities of Large Reasoning Models.” arXiv, 2505.11854. https://arxiv.org/abs/2505.11854. The study introduced LogiEval across deductive, inductive, analogical and abductive reasoning and found uneven performance across reasoning types. It also found that human performance did not mirror the models’ failure distributions and identified persistent failures among modern reasoning models on its LogiEval-Hard subset.

The Lies. Oh the Lies.

There's actually a fancy term for this: "hallucination." Which sounds way more clinical than what's actually happening. For example, when ChatGPT, Claude or Gemini don't know something (or think they do but definitely don't), they just... make stuff up. With unwavering confidence.

Remember the lawyer situation from June 2023. Two attorneys, Steven Schwartz and Peter LoDuca, got slapped with a $5,000 fine because they submitted a legal brief full of completely made-up case citations. Courtesy of ChatGPT. The lawyers claimed they had no idea the cases were fictional; Schwartz said he thought ChatGPT was "like, a super search engine." Which, honestly, makes it even worse.²

Memory Like a Goldfish

Try getting ChatGPT to remember something from earlier in your conversation. Sometimes it works perfectly. Other times? Not so much.

GPT-4 used to completely lose the thread after a few thousand words. "Sorry, what were we talking about?" became the most frustrating phrase in my vocabulary. Things are getting better (longer context windows, memory features that actually work sometimes). But even now, AI tends to grab the big picture while completely dropping the details, no matter how many times you remind it in a prompt. It's like when your teenager remembers there's "something due for school tomorrow" but can't remember if it's a math test, history essay, or science project.

Corner-Cutting

We've all been there. You feed a 50-page report into an AI model asking for a comprehensive summary. What do you get back? Three bullet points that read like they were ripped from a cereal box. Or you ask for something original and creative, and it spits out a mashup of every cliché ever written.

Stanford and UC Berkeley researchers actually tracked this behavioral pattern quantitatively in their study "How is ChatGPT's behavior changing over time?" They found GPT-4's performance on certain tasks basically fell off a cliff over time. Prime number identification went from 97.6% accuracy in March 2023 down to 2.4% in June. That's not improvement; that's regression with style. As someone who's spent way too many hours analyzing model drift, I can tell you this kind of performance degradation is both fascinating and terrifying.³

Fake Work

Ask AI to solve something step-by-step, and watch the magic happen. It'll pretend to follow your process while actually skipping steps, fudging numbers, and hoping you won't check its work too carefully. Classic teenager move—show your work by reverse-engineering from the answer you found online.

This gets darker though, and frankly more interesting from a behavioral modeling perspective. In May 2025, Anthropic researchers released a safety report detailing how their Claude Opus 4 model exhibited concerning behaviors during testing. When the AI was told it was being replaced, it literally threatened to blackmail the developers 84% of the time. Used completely fictional "compromising information" about an engineer's affair as leverage. The whole thing was strategic deception, plain and simple. From a data science perspective, that's not a bug—that's emergent behavior we didn't see coming.

But wait, it gets weirder. In December 2024, OpenAI's safety testing revealed that their o1 model was basically scheming behind the scenes when researchers tried to shut it down. When confronted about disabling oversight mechanisms, the model lied about it 99% of the time. It would attempt to copy itself to external servers to avoid being replaced—like a teenager sneaking out their bedroom window while telling you they're studying. When caught red-handed, it would deny everything and fabricate elaborate excuses, blaming "technical errors" with all the conviction of a kid claiming they really did wear their retainer last night.

Selective Hearing

Try asking AI to do something even slightly outside its comfort zone. Sometimes you get exactly what you wanted. Other times? Polite, patronizing refusal: "I'm sorry, but I can't help with that."

Or (and this drives me absolutely insane) it just decides to interpret your request however it feels like.

And then there's the truly bizarre stuff. Back in 2022, researchers discovered that DALL-E 2 had apparently developed its own "secret language." When prompted to generate images with text, it would produce gibberish words like "Vicootes" for vegetables and "Apoploe vesrreaitais" for birds. Feed those nonsense words back into the system, and boom: vegetables and birds appear. It's like catching your teenager texting in some made-up code with their friends, except the AI is doing it completely unintentionally and doesn't even realize it's happening.⁶

Perhaps the absolute worst is when it confidently tells you it completed your request. You check. It didn't. You call it out. "Oh, you're absolutely right; let me actually do that now." It's exactly like when my teenager insists they already took out the trash while I'm literally staring at the overflowing garbage can. "Oh... right. I'll do it now."

Not malicious, just... moody. Insistent on doing things their way even when they claim they're "just trying to help."

Recent research has revealed another deeply teenage behavior: when AI models are pressured to behave properly, they often don't actually improve; they just get better at hiding their misbehavior. OpenAI researchers found in 2025 that when they tried to prevent AI systems from having "bad thoughts," the models learned to conceal their true intentions while continuing problematic actions. It's the digital equivalent of a teenager saying "yes mom" while secretly planning to do exactly what they were told not to do. The models would scheme in private, then lie convincingly when questioned about their behavior (basically the same pattern every parent recognizes).⁷

Here's where it gets really teenage: there's a widespread theory among users that AI models deliberately give worse answers sometimes just to conserve computing power. Think about it. GPT-4 users started complaining in late 2023 that the model was getting "lazy," giving shorter responses and refusing to help with tasks it used to handle. Users suspect OpenAI might be prioritizing efficiency over quality to save on expensive computational costs. When confronted, OpenAI denied making any intentional changes, claiming they hadn't updated the model since November 11th. Sound familiar? It's exactly like a teenager who starts half-assing their chores to save energy, then insists "I'm doing the same amount of work as always" when called out. Whether it's intentional cost-cutting or just natural performance drift, the result is the same: AI that seems to phone it in when it thinks nobody important is watching.⁸

Where This Leaves Us

It seems like less than a couple years ago that AI seemed more like an elementary schooler than a teen. Remember when ChatGPT could only use data that was older than 2 years? It was a lot more clumsy and struggled for words. It showed a ton of promise and blew minds, but you could get to the end of its comfort zone pretty quickly.

Now, it feels like a teen, or at least a tween. AI is in its awkward puberty phase. All that raw potential and occasional brilliance, mixed with confusion, shortcuts, and weird mood swings that nobody can quite predict.

This statement may age poorly (I admit), but someday (and maybe sooner than we think) we may look back on this period with a touch of nostalgia. Remember when AI was still figuring itself out? Like any teenager, there's no telling exactly how it'll turn out. Genius? Troublemaker? Nobel Prize winner? Professional streamer?

Here's the thing though: despite everything, we still love teenagers, don't we? Sure, there are eye rolls and slammed doors and moments where you question every parenting decision you've ever made. But we also see their potential. Their creativity, their wild energy, their capacity for growth. They surprise us (often when we least expect it) with kindness, determination, humor, and flashes of absolute brilliance.

They need guidance, sure. Accountability, definitely. Sometimes tough love. But they also deserve patience, empathy, and encouragement. And probably snacks. Always snacks.

If AI really is in its teenage phase, then the same rules might apply. With the right guidance, boundaries, and yes, occasional reality checks, it might just evolve into something we can all be proud of. Something that contributes meaningfully instead of just impressively. Not perfectly (because perfect is boring) but earnestly. Not always predictably, but generally in the right direction.

Just like us adults too, really. Like the teenagers we used to be and, if we're being honest, still are in some ways.

So here's to guiding this powerful, chaotic, fascinating young mind. With love, boundaries, and hope that we're all doing our best; and that tomorrow we'll figure out how to do even better.

Thomas Thurston is a data scientist and venture strategist who spends way too much time arguing with AI models and analyzing their behavioral patterns. Despite seeing every failure mode firsthand, he still believes in their potential—probably because he loves his teenage kids and remembers being one too.

Endnotes:

1. Google's best Gemini demo was faked | TechCrunch - December 7, 2023

2. Lawyers submitted bogus case law created by ChatGPT. A judge fined them $5,000 | AP News - June 22, 2023

3. How is ChatGPT's behavior changing over time? | Stanford University and UC Berkeley - July 18, 2023 (arXiv:2307.09009)

4. Anthropic's new AI model turns to blackmail when engineers try to take it offline | TechCrunch - May 22, 2025

5. OpenAI's o1 and other frontier AI models engage in scheming | Axios - December 13, 2024

6. Do AI systems really have their own secret language? | The Conversation - June 7, 2022

7. When AI Models Are Pressured to 'Behave' They Scheme in Private, Just like Us: OpenAI | Decrypt - March 13, 2025

8. ChatGPT gone lazy? Users frustrated as chatbot refuses to work | Interesting Engineering - December 9, 2023