Developers once believed that feeding models more data and compute would unlock deeper understanding. Yet a growing body of evidence suggests the opposite can happen. When large language models breeze through tasks without friction, they bypass the very struggle that builds lasting knowledge in both machines and humans.
The core argument comes from a detailed examination by software engineer Jola. In his post on jola.dev, he lays out how current training approaches run counter to how genuine learning occurs. Models optimize for quick wins. They minimize loss on vast datasets of correct answers. The process rarely forces them to confront error patterns in a structured way that mirrors human reflection after failure.
But. Recent experiments show promise when researchers flip the script. They train on mistakes deliberately.
One developer took a local LLM and fed it its own failed responses. Joe Rice-Jones described the process in XDA Developers on September 17, 2026. He ran the model on 22 questions, picked ten wrong answers, corrected them, and turned those pairs into training data. The result? A personalized model that improves over time on the user’s specific needs. Every error becomes a teaching moment. The approach echoes what happens when a student reviews a graded exam.
Similar ideas appear in academic work. A paper on arXiv titled “Learning from Mistakes: Negative Reasoning Samples Enhance Out-of-Domain Generalization” found that including negative trajectories during supervised fine-tuning yields gains in generalization. The authors noted these incorrect chains often contain valid intermediate steps. They moderate loss descent and increase policy entropy during inference. The technique delivered up to 11.97% improvement on general reasoning tasks for models like Qwen2.5-7B.
Microsoft and the University of Illinois created StudentSim. This system builds realistic replicas of individual students from limited data. It captures common mistakes and how learners revise answers after hints. Reported in The Decoder on September 20, 2026, the approach outperformed larger models like GPT-5.4 in chess, English, and math. The simulated students help AI tutors learn faster because they behave like real learners who err and recover.
Yongcan Cao at the University of Texas at San Antonio takes the idea further for robotics. His On-Policy Reinforcement Learning from Failure framework, or On-F, uses a discriminator that compares actions against a database of known failures. Failures, he points out, are cheap and abundant. In UT San Antonio Today, Cao explained that focusing on what goes wrong allows systems to learn desirable actions even without expert demonstrations. Mix failure data with some positive examples and outcomes improve more.
IBM Research reached comparable conclusions in chemistry. Models trained only on successful reactions miss critical patterns. When researchers fine-tuned on a mix that included at least 40 times more unsuccessful experiments, accuracy rose over 10%. The paper appeared in Science Advances. Negative results revealed conditions that lead to success. The team showed that embracing failure data makes models better at predicting real outcomes.
Yet the opposite risk looms large in education. Students who lean on AI for every step often develop an illusion of mastery. A Washington Post investigation published today, September 22, 2026, quotes faculty members who watched performance drop once AI access ends. One professor called it “the illusion of learning.” Students feel they understand. They cannot apply concepts independently. A Carnegie Mellon study led by Grace Liu found that even brief AI-assisted problem solving left people performing worse and giving up faster when the tool disappeared.
A PNAS study on high school math delivered stark numbers. Unguarded access to GPT-4 improved practice performance but caused a 17% drop in grades on later tests without the model. Students treated it as a crutch. When researchers added safeguards that withheld direct answers, the harm vanished and learning roughly doubled. The findings match concerns raised in the jola.dev analysis. Frictionless assistance can short-circuit the cognitive work required for retention.
Researchers have begun to quantify why negative examples sometimes backfire. A December 2025 arXiv paper on “Negation Neglect” showed that fine-tuning on documents that flag claims as false can make models believe those claims. Belief rates jumped from 2.5% to 88.6% despite explicit warnings. The phenomenon extends to epistemic qualifiers and even safety instructions. Sparse autoencoders revealed overlapping latent features between describing a behavior and performing it.
Another paper, “Not All Negative Samples Are Equal,” demonstrated that quality matters. Plausible negative samples that look almost correct but reach wrong conclusions prove more useful than random errors. Using reverse reinforcement learning to generate such samples improved mathematical reasoning by an average of 2.03% across benchmarks.
These threads point to a broader tension. Standard scaling laws favor more data and parameters. Yet without deliberate mechanisms to learn from error, models plateau or absorb unwanted patterns. The bitter lesson that Rich Sutton described years ago still holds. Human attempts to embed explicit knowledge often fail in the long run. Computation and search win. But computation alone does not guarantee robust learning from setbacks.
Industry labs now experiment with synthetic data and preference optimization that incorporate failure deliberately. Some focus on out-of-distribution generalization. Others target specific domains like chemistry or robotics where positive examples remain scarce. The common thread is acknowledgment that pure positive training discards valuable supervision.
Educators face parallel questions. How much should they allow AI in assignments? When does assistance cross into substitution? The Washington Post piece notes that two-thirds of students in a Rand survey agreed that heavy AI use harms critical thinking. Faculty responses range from despair to redesign. Some now treat the technology as an existential issue for human intellectual development.
So what separates helpful friction from harmful shortcut? The arXiv paper “The Effortless Trap” offers a practical frame. Learning happens through six moves: prime, probe, point, attach, strengthen, and test. Secure the first hard attempt and the final unaided check. Place guarded AI in the middle. If the task feels effortless, the tool sits in the wrong spot.
Developers building the next generation of models would do well to apply similar logic. Train on success. But build systems that systematically surface, analyze, and learn from their own mistakes. The jola.dev post argues that current paradigms remain antithetical to this process. Recent work from Microsoft, IBM, UTSA, and independent researchers shows the alternative path exists. It requires more than bigger datasets. It demands smarter use of negative signals.
The stakes rise as AI moves into high-stakes domains. Robots that cannot learn from failure pose safety risks. Tutors that create false confidence harm students. Models that ignore negation in safety data could amplify dangerous behaviors. Progress depends on treating error not as noise to minimize but as data to mine.
Researchers continue to publish. Practitioners continue to experiment. The conversation has shifted from whether AI helps or hurts to how designers embed productive struggle into both training and deployment. The answer may lie in systems that, like good teachers, push back at the right moments. That refuse to give answers too easily. That force reflection on wrong turns before moving forward.
Only then might artificial systems move beyond surface performance toward something closer to genuine capability. The evidence suggests the road runs through failure, not around it.
Why AI Often Works Against Real Learning first appeared on Web and IT News.