Both Companies and Students are Unequal in Ability to Use AI Productively

It is relatively hard to determine whether artificial intelligence is helping some students learn better or stunting their development of learning skills.

But that might not be unexpected. 

In the business world, higher-performing firms seemingly are better at adopting and extracting value from new technologies, including AI. 

A small minority of high-performing companies capture the bulk of AI-driven returns (top five to 20 percent account for the large majority of gains), according to one analysis. Other studies tend to agree.  

Students who are already good at standardized tests may also be better at using AI as a learning technology. 

When educational researchers evaluate AI tools (such as automated writing assistants, generative AI tutors, or adaptive platforms), they frequently report positive short-term gains in student output, task completion speeds, or test scores. 

However, critics and recent consensus reports argue that these metrics may conflate performance with learning. That’s akin to “teaching to the test,” where the objective is to improve student test scores by focusing on improving test performance, rather than other learning objectives. 

Study	Design / students	What it found	Why it matters
Contractor & Reyes, 2026 – Experimental Evidence on the Learning Impact of Generative AI	Randomized experiment with undergraduates; AI vs. no AI; unaided tests immediately and one week later	AI increased immediate knowledge-test scores by 0.27 SD, with gains persisting one week; essay effects depended on whether AI was used for augmentation versus automation	Strong evidence that AI can improve actual unaided performance, but how students use AI matters greatly. (IZA)
Fischer, Rau & Rilke, 2025/26 – AI Tutoring Enhances Student Learning	RCT, 334 university students	AI tutoring raised test performance 0.23 SD; effects were largest among students with lower baseline knowledge and stronger self-regulation	Direct evidence for your hypothesis: baseline ability changes the size of the AI effect. (Scale)
Harvard AI tutor RCT, 2025	194 undergraduates; AI tutor vs. active-learning classroom	AI group learned substantially more in less time; researchers measured pre-test knowledge and post-test learning	Shows why pre-testing matters: the relevant outcome is change from baseline, not simply the final score. (PubMed Central (PMC))
Noy et al./PNAS high-school mathematics study	High-school students using GPT assistance in math practice	Unrestricted/basic GPT assistance could reduce performance on an unassisted exam; students nevertheless believed they had learned more	Particularly important: perceived learning and measured learning diverged. (DOI)
ChatGPT cognitive-crutch RCT, 2025	120 undergraduates; ChatGPT-assisted vs. traditional study	AI group scored 57.5% vs. 68.5% on a surprise retention test 45 days later	Shows why immediate test performance can be misleading: easier performance during learning can come at the cost of retention. (ScienceDirect)
Oreopoulos et al., 2026 – AI tutoring/mastery math	>6,000 middle-school students; randomized AI vs. conventional computer-assisted learning	AI students made fewer attempts but were more accurate conditional on attempting; strongest delayed-learning evidence occurred when AI was embedded in a mastery structure	AI may alter the learning process, not simply raise scores. Structure matters. (National Bureau of Economic Research)
Oreopoulos & Low, 2026 – Khanmigo	Two-year cluster RCT in 18 Tennessee middle schools	AI-tutor assignment raised math achievement about 0.06–0.08 SD per year, with larger effects for full-year active participation	Real-world AI effects can be modest because actual usage is much lower than theoretical availability. (National Bureau of Economic Research)
Saloojee et al., 2026 – medical students	RCT, final-year medical students, ChatGPT available during clinical exams	No significant improvement in clinical performance; prior academic performance predicted scores	A useful reminder that strong students don't automatically benefit more from AI; task and implementation matter. (PubMed)
Wu et al., 2026 – meta-analysis of 35 experimental/quasi-experimental studies	Meta-analysis	Finds substantial variation in ChatGPT effects across subjects, educational levels, instructional approaches and knowledge types	The average AI effect conceals considerable heterogeneity. (Nature)
Deng et al., 2025 – meta-analysis of experimental studies	Systematic review/meta-analysis	Overall positive effects on academic performance and higher-order thinking, but also reduced mental effort	Illustrates the central ambiguity: better outcomes can coexist with less cognitive effort. (DOI)
Nickow, Oreopoulos & Quan, 2020 – tutoring meta-analysis	Meta-analysis of tutoring experiments	Tutoring produced a large average learning effect (~0.37 SD), but effects varied substantially by program/context	Important pre-AI benchmark: individualized assistance has long produced heterogeneous effects, so AI shouldn't be expected to have one uniform effect. (National Bureau of Economic Research)
Kraft, Schueler & Falken, 2024 – tutoring generalizability	Meta-analysis of 265 RCTs	Effects fell to roughly one-third to one-half of the original estimates when studies were restricted to settings resembling large-scale standardized-test interventions	A major warning about extrapolating impressive experimental effects to real-world populations. (ERIC)

Up to a point, AI might help learning by minimizing friction. On the other hand, learning arguably often requires encountering friction and overcoming it. So if AI minimizes cognitive struggle, it might also negatively affect learning. 

The other unavoidable problem is that some students are just better equipped to benefit from AI tool use. Students who are already high performers, highly self-regulated, or technologically fluent tend to extract more value out of AI tools. 

Struggling students might rely too much on “getting the answers” without developing thinking, research or other skills. In other words, “output” might not reflect student learning so much as AI answers. 

And it might be a reasonable assumption that better-equipped learners will also tend to be those that learn most when using AI as well. 

Teachers might agree that an in-class essay exam using blue books is a more-reliable test of what a given student might know, though obviously also favoring better writers, compared to any out-of-class essay, which can be a better test of what a given AI engine knows and expresses. 

The point is that better-performing students are also likely to be better-performing users of AI. 

Study / Report Focus	Authors / Organization	Key Finding Regarding Performance vs. Learning	Source Link
Children’s and Adolescents’ Learning with Educational Technology	American Psychological Association (APA) (2026)	Warns that metrics like time-on-task, completion rates, and immediate output are frequently mistaken for learning. Highlights that AI can improve immediate performance while simultaneously reducing long-term independent knowledge retention.	APA Report on Educational Tech
Measuring Student Trust and Over-Reliance on AI Tutors	Educational Cybernetics and Studies (2025)	Investigates how moderate trust improves confidence, but over-trust triggers automation bias—where students accept AI-generated answers without critical evaluation, inflating performance metrics while depressing deep cognitive skill-building.	ECSE Article
Assessing the Quality of AI-Generated Exams: A Large-Scale Field Study	ArXiv Educational Technology Research (2025)	Evaluates AI-generated assessment items across ~1,200 students using Item Response Theory (IRT), establishing that while AI can efficiently construct valid psychometric tests, student baseline engagement heavily dictates performance outcomes.	arXiv Study
Improving Student Learning with Hybrid Human-AI Tutoring	Educational Research Working Paper Series (2023–2026)	Explores the nuance of baseline abilities, finding that while lower-achieving students show measurable proficiency gains with hybrid support, unmonitored independent AI usage risks widening the gap due to varying digital literacy and self-regulation skills.	arXiv Working Paper

I’m not sure how we compensate for learning prowess in general, anymore than we seem to systematically produce high-performing firms.

Diffusion Models Can Create New Images Without Infringing on Training Sources

A study of language models, especially those designed to create images and video, might suggest that copyright infringements by diffusion-based models might be impossible to prove. 

The authors say “locating a part of the training data that can be held responsible for a generated sample, can become impossible if a model is trained on a sufficiently large corpus of data.”

In other words, a large training set is all that is needed to render copyright traceable to any single source impossible. 

The study methodology relies on omitting an attributed unit from the training set (one image or all images by the same creator), the induced counterfactual sample cannot be directly attributed to that same unit. 

source: Nature Communications

In the above example, a generated image created from a smaller data set is more attributable to a single training image.

A large data set, on the other hand, creates a new image that shows no trace of a single training image.  

Model / family	Main generation	Diffusion?	Basic architecture / approach
GPT-4 / GPT-4o / GPT-5 family	Text, multimodal	No*	Primarily autoregressive Transformer for language
Claude 3/4 family	Text, code, multimodal	No*	Autoregressive Transformer
Gemini 1–4 family	Text, multimodal	No*	Primarily autoregressive Transformer
Llama family	Text, code	No	Autoregressive Transformer
Mistral / Mixtral family	Text, code	No	Autoregressive Transformer, including MoE
DeepSeek family	Text, code, reasoning	No	Autoregressive Transformer/MoE
Gemini Diffusion	Text/code	Yes	Text diffusion Transformer; generates blocks of tokens and iteratively refines them (Google DeepMind)
DiffusionGemma	Text/code	Yes	Non-sequential diffusion Transformer based on Gemma research (Google DeepMind)
Stable Diffusion 1.x/2.x/XL	Images	Yes	Latent diffusion, historically U-Net-based
Stable Diffusion 3/3.5	Images	Yes	Diffusion Transformer / MM-DiT
FLUX.1 / FLUX family	Images	Yes	Diffusion/flow-based Transformer
DALL·E 2	Images	Yes	Diffusion
DALL·E 3	Images	Yes	Diffusion-based image generation
Imagen 2/3/4	Images	Yes	Diffusion-based image generation; Google explicitly describes Imagen as a diffusion model (Google DeepMind)
Midjourney	Images	Yes*	Widely understood to use diffusion/related denoising techniques, though Midjourney does not disclose its architecture in detail
Adobe Firefly	Images/video	Yes*	Diffusion-based generation among its models; architecture varies by model
Sora	Video	Yes	Diffusion Transformer (DiT); starts from noisy video patches and denoises them (OpenAI)
Sora 2	Video	Yes*	Diffusion-based video generation; details of current implementation are less fully disclosed
Veo 2/3/3.1	Video	Yes*	Diffusion/Transformer-based video generation; Google describes Veo as drawing on its diffusion-model research (Google DeepMind)
Runway Gen-2/3	Video	Yes*	Diffusion-based video generation
Kling	Video	Yes*	Diffusion/Transformer-based video generation
Luma Dream Machine	Video	Yes*	Diffusion/Transformer-based video generation
MusicGen	Music/audio	No	Autoregressive Transformer over discrete audio tokens
AudioCraft / related models	Audio/music	Mixed	Different models use different architectures
Genie 2	Interactive worlds/video	Yes	Autoregressive latent diffusion—an interesting hybrid (Google DeepMind)

“Our findings provide a compelling case that attribution, the task of locating a unit of data within the training set that can be held responsible for a generated sample, is practically impossible on contemporary generative diffusion models,” say authors Zheng Dai and David Gifford.