New Study: Why AI Overreliance Hurts Skills and How to Fix It

Anthropic’s latest study shows a 17% drop in skill mastery for users who rely heavily on AI assistants.

Welcome to The Median, DataCamp’s newsletter for January 30, 2026.

In this edition: Big Tech earnings diverge over AI ROI, OpenAI launches Prism, and Google rolls out Project Genie. Plus, why AI coding assistants can stunt skill development and practical steps to fix it for learners, educators, and engineering leaders.

This Week in 60 Seconds

Big Tech Earnings Diverge Over AI ROI

While Microsoft, Meta, and Apple all beat analyst revenue expectations this week, market reactions varied widely depending on the perceived returns from their AI investments. Meta’s stock surged 10% after the company demonstrated that AI is directly boosting its core advertising business, providing a clear narrative of immediate value. Conversely, Microsoft shares plunged 10% as investors balked at massive AI-driven capital expenditures coupled with a slight deceleration in Azure cloud growth, signaling growing skepticism about the timeline for returns on infrastructure spending. Meanwhile, Apple posted a record quarter of $143.8 billion driven by the iPhone, yet saw only a tepid stock rise as the market remains cautious about its long-term AI strategy, despite a new collaboration with Google and promises of a more personalized Siri in the pipeline.

OpenAI Launches Prism to Accelerate Scientific Workflows

OpenAI has officially released Prism, a free, AI-native workspace designed to modernize scientific writing and collaboration. Built on the foundation of the acquired Crixet platform, Prism integrates the new GPT-5.2 model directly into a cloud-based LaTeX environment, allowing researchers to draft papers, debug equations, and manage citations without toggling between fragmented tools. The workspace supports unlimited collaborators and features “Thinking” capabilities that can reason over a paper’s entire structure.

Google Rolls Out Project Genie for Infinite World Creation

Google has officially launched Project Genie, an experimental research prototype available to AI Ultra subscribers in the U.S., offering the first public access to its Genie 3 world model. The tool allows users to create, explore, and remix infinite interactive environments in real time, using “World Sketching” to define physics and character perspectives before dynamically generating the path ahead. Although the current iteration is limited to 60-second generations and lacks full fidelity in physics simulation, Google positions the release as an important advancement in its AGI roadmap, shifting focus from static 3D captures to general-purpose systems that can simulate and navigate the complexity of the real world.

OpenAI to Retire GPT-4o and Legacy Models in February

OpenAI has announced the definitive retirement of GPT-4o, GPT-4.1, and several mini variants from ChatGPT on February 13, 2026, consolidating its consumer lineup around the GPT-5 series. While the company had previously restored GPT-4o access to appease users attached to its specific conversational warmth, OpenAI states that recent customization features in GPT-5.1 and GPT-5.2, such as adjustable tone and enthusiasm controls, have rendered the older model obsolete. With daily usage of GPT-4o dropping to just 0.1%, the move allows OpenAI to reallocate resources toward developing a less restrictive version of ChatGPT designed for adults, a shift supported by newly implemented age prediction tools to safeguard younger users.

Anthropic Study Finds AI Coding Assistants May Hinder Skill Development

Anthropic has published a new randomized controlled trial challenging the assumption that AI tools purely enhance developer productivity, finding that novice engineers who used AI to learn a new library scored 17% lower on retention quizzes than their manual counterparts. The research highlights a specific degradation in debugging skills, as users who delegated code generation to the AI failed to engage cognitively with the material. Surprisingly, the study also found that the AI group was not significantly faster on average, as many participants spent excessive time crafting prompts rather than writing code, effectively trading potential skill mastery for a wash in productivity. We’ll unpack the results in the Deeper Look section below.

---

Unlock 600+ Courses

Join millions advancing their careers with data and AI skills. Get full access to 600+ courses, career tracks, and certifications, including Python, AI, Power BI, and SQL. Available for individuals and teams for a limited time. Save 50% now!

---

A Deeper Look at This Week’s News: The Right Way to Learn with AI Assistants

This week, Anthropic released a study that challenges the prevailing narrative of AI as a pure productivity engine. The findings reveal that while AI can help novices complete tasks, it can actively inhibit the deep learning required to understand why the code works, potentially creating a generation of engineers who can generate software but cannot debug or maintain it.

Study design

To isolate the impact of AI on learning, researchers recruited 52 software developers (primarily junior engineers) and asked them to learn a new tool: Trio, a Python library for asynchronous programming that none of them had used before.

The participants were split into two groups. The control group had to complete the coding tasks using only official documentation and web search. The treatment group, however, had access to an AI assistant (powered by Claude) embedded directly in their coding environment.

Source: Anthropic

The goal wasn’t just to finish the code. After the session, all participants were given a knowledge quiz (without AI help) to test their conceptual understanding, code-reading ability, and debugging skills.

The results

On the post-task knowledge quiz, the AI-assisted group scored 17% lower (approximately two letter grades) than the control group that coded manually.

The erosion of skills was most acute in debugging, suggesting that while AI can help novices generate code, it actively degrades their ability to fix it when it breaks.

Counterintuitively, the AI group was not significantly faster than the manual coders.

The p-value was statistically significant only for the quiz score variable. Source: Anthropic

The study identified six distinct interaction patterns with the AI that determined success or failure. Users who engaged in “AI delegation” (copy-pasting without reading) or “iterative debugging” (repeatedly pasting errors back into the chat) had the lowest quiz scores.

Conversely, users who adopted “conceptual inquiry” (asking how concepts work before coding) or “hybrid explanation” (asking the AI to generate and explain code) maintained learning outcomes comparable to manual coders. This suggests that cognitive engagement (not just code production) is the variable that matters.

Source: Anthropic

Practical takeaways for learners

The participants who retained the most knowledge were those who refused to let the AI do the thinking for them. If you are using AI to upskill, the goal is to keep your cognitive engine running even while the AI types the syntax.

Based on the study’s high-scoring interaction patterns, here is how to structure your learning sessions:

1. Avoid AI delegation: The lowest retention scores came from users who simply asked for code and pasted it without review. If you treat the AI as a vending machine for code, you bypass the neural circuits required for mastery.

2. Ask for a hybrid output: One of the most effective patterns involved asking the AI to generate code and explain the underlying logic in the same prompt. This forces you to read and process the output before using it.

3. Ask follow-up conceptual questions: If you do ask the AI to write a script, follow it up immediately with a conceptual question (e.g., “Why did you use an async context manager here instead of a standard loop?”).

4. Don’t debug blindly: The study found that repeatedly pasting error messages into the chat until the code works is a trap. Instead, ask the AI to explain the cause of the error before asking for the fix.

Practical takeaways for educators

A student can now turn in a perfect assignment (mirroring the “AI Delegation” group) yet fail a conceptual quiz. Here are concrete ways to adapt your curriculum based on the study’s specific findings:

1. Invert the assignment: The study showed that AI-reliant users experienced the largest decline in debugging skills. Instead of asking students to write code from scratch (which they will just delegate to AI), give them a pre-written, working AI script that contains subtle logic errors or security vulnerabilities. Grade them on their ability to find, fix, and explain these bugs.

2. Grade the chat: For any AI-assisted assignment, students could submit their chat logs. Award points specifically for prompts that ask why or how, and deduct points for iterative debugging (blindly pasting error messages).

3. Require code comments: To force the “generation-then-comprehension” pattern (where the student must audit the AI’s output), require detailed comments for every meaningful line (or block) of AI-generated code.

4. Use conceptual quizzes: The study used a post-task quiz to expose the learning gap. Use syntax and conceptual quizzes immediately following a coding project. A student who ships complex features but fails basic conceptual questions is a red flag for AI over-reliance.

Practical takeaways for engineering leaders

For engineering managers and CTOs, this study serves as a warning against measuring AI success solely by velocity. If your junior engineers are using AI to close tickets faster but are failing to learn the underlying systems, you are effectively borrowing against your future technical leadership. Here are a few practical takeaways:

1. Beware of false velocity: The study found that while the “AI Delegation” group was the fastest, they also had the lowest comprehension. Be wary of metrics that reward pure speed (like ticket volume), as they may encourage the exact behaviors that stunt skill growth.

2. Audit the process: The study shows that correct code does not imply understanding. Ask junior engineers to explain why a specific pattern was used or to walk through the logic of an AI-generated block.

3. Incentivize hybrid workflows: Encourage your teams to use AI tools that explain code rather than just auto-completing it. Discourage the AI delegation pattern where the tool acts as a black box.

4. Redefine hiring and promotion skills: As syntax becomes commoditized, the value of a developer shifts to high-level architecture, system design, and debugging. Your hiring and promotion rubrics should evolve to weigh these skills more heavily than rote coding speed.

---

Industry Use Cases

NVIDIA Open-Sources Earth-2 to Democratize AI Weather Forecasting

NVIDIA has launched the Earth-2 family of open models, aiming to replace the costly, supercomputer-dependent physics simulations traditionally used for weather forecasting with faster AI alternatives. The new stack includes “Nowcasting,” a generative model that predicts local storms at kilometer resolution in minutes, and a medium-range tool capable of forecasting over 70 variables up to 15 days out. Early results are significant: the Israel Meteorological Service achieved a 90% reduction in compute time while improving rainstorm prediction accuracy, and energy major TotalEnergies is deploying the models to enhance decision-making for power grid operations. Read more in this article from NVIDIA.

Animators and Google DeepMind Create AI-Hybrid Film for Sundance

Pixar alum Connie He and Google DeepMind have bridged the gap between manual animation and generative video in Dear Upstairs Neighbors, a short film premiering at the Sundance Film Festival. Facing the challenge of animating complex, shifting abstract expressionist styles that are prohibitively difficult to render traditionally, the team fine-tuned Google’s Veo and Imagen models on specific character designs to ensure consistency. By utilizing novel video-to-video workflows, animators could feed rough sketches or 3D block-outs into the model, using AI as a styling tool that preserves the human performer’s timing and intent while upscaling the final output to 4K. Read more in this article from Google DeepMind.

Anthropic Partners with UK Government to Power Public Services

Anthropic has been selected by the UK’s Department for Science, Innovation and Technology (DSIT) to pilot a dedicated AI assistant for Britain’s primary digital services platform, aimed initially at modernizing employment support. This agentic system, powered by Claude, goes beyond simple Q&A by actively guiding citizens through complex processes, such as finding jobs or accessing training, while maintaining context across sessions so users don’t have to repeat information. The project utilizes a “Scan, Pilot, Scale” framework, with Anthropic engineers embedding directly with civil servants to build internal government expertise and ensure the system can eventually be maintained independently. Read more in this article from Anthropic.

Tokens of Wisdom

Writing software comes with a lot of responsibility, especially if you share it with people. You are responsible for what that software can do and what that software can expose. If you’re building on top of other people’s data, that’s a really big weight that you have to think through very, very deliberately.

—Matt Palmer, Developer Relations at Replit

In our latest DataFramed podcast, we sat down with Matt Palmer to discuss vibe coding, the rise of non-developers building impactful tools, and Replit’s role in simplifying the coding process. If you want to learn vibe coding, we partnered with Replit to create a course on Vibe Coding With Replit.