Tools for the Next Generation of AI Engineers

Students already use generative AI for homework. What would it take for them to become designers of those tools, rather than only recipients of answers? Stefania Druga opens with that question of participation. Multimodal AI could expand what children make and how they learn; she also describes cases in which students prefer AI assistance to human tutors. Her goal is to put children inside the engineering process.

Druga cites Gen Z as 70% of AI users, but that denominator should not be carried forward as an established fact. The related Salesforce survey measures the reverse: 70% of surveyed Gen Z respondents reported using generative AI. The practical design question remains how to turn young people's use of AI into an opportunity to understand and shape it.

Scratch provides a starting point: free, open-source visual programming developed at MIT, where programs are assembled from blocks. Druga describes its worldwide reach as more than 100 million children. During her master's work in the MIT lab, she began Cognimates in 2015 to extend that accessible programming environment into AI.

Cognimates extends Scratch so children can build games, train AI models, and program hardware with the same Lego-like blocks. One early example comes from Tiana, who wrote her first robot program and then played with it for half an hour. Her game was hide-and-seek: she ran around the room while the robot turned and, whenever its detected-person count exceeded zero, said “I see you.”

The interesting bug is visible without understanding the person detector's internals. The program checks the condition only once. To keep playing, it needs a loop around the detection and response. Expressed as Scratch blocks, that repeated behavior is:

when green flag clicked
forever
  if <(number of people) > (0)> then
    say [I see you]
  end
end

An AI capability becomes something a child can reason about through ordinary control flow. The detector supplies a value; the child's program decides what to do with it and whether to keep checking. Scratch's block library, coding area, and stage make that connection quick enough to discover through play.

In an embedded video, a child describes playing rock-paper-scissors through a camera: make a hand gesture on cue, then compare it with the computer's move. The child explains the computer's improvement through an analogy with practice—people do not know everything at first either, but get better by trying. The project gives the learner a concrete behavior to observe and discuss.

The adults describe different starting attitudes. Technically confident parents readily encouraged participation; another parent and teacher acknowledges that AI can feel frightening while still treating understanding it as an essential skill. A child's analogy makes the educational purpose especially clear: their father once took a car apart to see how it worked. Programming systems that adults usually build offers a similar way to inspect technology from the inside.

The projects from 2015–2016 used basic image and text classifiers. Alongside custom training, Cognimates supplied extensions for sentiment analysis, image classification, voice assistants, micro:bits, and robots. Children who found voice assistants too limited could program them to remember personal details and preferences. On the training page, they could drag and drop image examples to build their own classifier.

One child wanted a unicorn-versus-narwhal game. After collecting training images, the learner could select that custom model in a program, hold a drawing up to the camera, and inspect both the predicted class and its confidence. A narwhal prediction with very low confidence creates a useful next question: what differs between this drawing and the examples used for training? In Druga's example, most training images were cartoons, so the learner's next move was to add more hand-drawn examples. The important feedback is not merely whether the label is right; it is the connection between the model's behavior and the data the child supplied.

That cycle—supply examples, inspect a prediction, revise the examples—makes AI and data literacy tangible. The same tools supported projects across subjects:

• Science: turn a webcam into a microscope to examine food.
• Games: build camera-based rock-paper-scissors.
• Literature: analyze speech for resemblance to a writer's style or another category of language.

Across public schools, private schools, and community centers, Druga observed children forming hypotheses about models and robots, testing them, and revising their understanding. The design requirement is to make that scientific process easy to enter, with projects engaging enough that children want to continue.

To examine whether making changed children's perceptions, Druga asked questions before and after their programming and model-training experience. Were voice assistants, smart toys, and robots intelligent? Trustworthy? Likable? Friendly? This work preceded ChatGPT and Gemini. Druga reports a significant change in intelligence attribution after children learned to train and program models, although she supplies no effect size or statistical details here. She also reports that Cognimates was used worldwide and translated into 50 languages.

When the pandemic left families learning at home, the design problem expanded from children's individual projects to parents and children coding together. An early robot tutor illustrates the interaction Druga wanted to preserve. Jibo asks the learner to write a program that lets it learn their name: first find the green-flag block, then find the ask block. The robot offers feedback as the learner works. The object being programmed participates in a conversation about its own program.

A robot is an expensive prerequisite, so Druga moved toward a browser-based pair-programming companion. Before implementing an assistant, she ran design studies with families from varied backgrounds across ten US states. In this Wizard of Oz study, Druga herself supplied the apparent AI's chat responses. Children initially interacted as though they were talking to an assistant and were told afterward that a person had been responding. This let the study investigate what support families wanted before committing to a system design.

Children wanted help generating ideas, not an assistant that completed the project for them. A bear game or an interest in soccer could become the starting point for brainstorming. One twelve-year-old described the difficulty of beginning a project and immediately running out of ideas. Once a project existed, questions could help elaborate it: in Pong, how should the ball move, and how could it move faster? These prompts give the learner something to decide and try.

Encouragement mattered too. Participants described the frustration of getting code to work and the satisfaction of having someone acknowledge success. Parents reported that participating children spent twice their usual time programming. That comparison came from parents' accounts, not an instrumented measurement of programming time.

The companion did not remove the need for an adult. Parents still helped when the interaction became distracting or failed to manage turns between siblings. More complex Scratch concepts also proved difficult to explain: clone creates additional instances of an object, while broadcast introduces communication between scripts. Those limits helped define what the implemented assistant would need to support and what its evaluation should test.

Druga next assembled an evaluation benchmark, describing more than 100 Scratch program cases used to compare models. The tasks followed the needs identified in the family study: explain code, explain it through learning exercises, debug it, and generate ideas. She calls the results promising but gives no scores in the talk. The benchmark is a test of assistance on these tasks, not a measurement of children's learning gains.

The next step was to build Cognimates Copilot and test it with young people. Druga reports a newly completed study involving 18 participants from 11 countries, working in different languages. The implementation is straightforward: a code editor sits beside an AI chat; messages go to a web server, which returns responses from AI services. She was evaluating several models, including one fine-tuned on Scratch projects. The system could also generate images for games.

A session with a child from Mexico took place entirely in Spanish. Across the sessions, Druga observed conceptual support, design support, positive encouragement, and help navigating the platform. She also observed failures and occasions when children declined the assistant's suggestions. Those refusals matter: useful assistance includes leaving room for the learner to disagree.

A student from Jamaica who had never programmed in Scratch reached a functional program with Copilot's support. Much of the help was practical orientation: where to find a loop block and how to create variables. For experienced users, support shifted toward assets, code improvements, additional features, and new levels.

The same chat interface can serve these different needs, but the substance of its help must change with the learner's experience.

Children also used the assistant for backgrounds, character names, and plot ideas. In one example, a student pasted an image of their code, requested an idea, and then rejected the proposed movement. The assistant replied, “No worries. It's your game. Have fun.” The child remains the author, including when they reject the tool's advice.

That principle shapes how help is offered. Druga describes the prototype's default as asking questions instead of giving an answer. If a stuck student asks the same question three times, it offers a hint. This balances support with the challenge that makes the programming activity educational. It also treats the assistant as a motivator and starting point for Scratch's cold-start problem: arriving at the editor without knowing what to make.

Customization is part of that balance. Some children wanted voice interaction; others wanted only typing. Some always wanted three ideas; others preferred one. A fixed response style would make the same assistant helpful to one learner and intrusive to another.

Participants wanted assistance beyond a chat sidebar. They asked for agents that could move blocks alongside them and demonstrate programming: watch an agent build Pac-Man, or watch five agents collaborate on an Asteroid game. They also wanted generated assets placed directly on the stage instead of delivered only through chat. These were requests for future behavior, not capabilities demonstrated in the study.

Druga's proposed next phase was to integrate the agent throughout the interface, add sound generation, and potentially let it react to a camera stream. As capabilities expand, the assistant also needs to explain its limits. She gives examples of explanations referring to unavailable training images or a prompt that restricts answers to Scratch. The design requirement is understandable disclosure of the actual constraint; those illustrative messages do not establish why a particular model failed. She then turns to a live demonstration of the prototype.

The live session starts with a greeting. Copilot asks what the learner wants to work on in Scratch, and an audience suggestion supplies the project: a racing game. Asking for ideas produces a turbo-boost button and extra speed. This demonstrates the first step out of the blank editor—a specific feature that a learner can accept, change, or ignore.

Druga then moves from ideation to an existing Scratch project. She points to OCR implementations as examples of the sophisticated programs people build in Scratch. Browsing is slow, so she loads a project she has already downloaded. She draws input on its stage and asks the project to recognize it, describing the results as two and eight. The recognition belongs to the loaded Scratch program; the adjacent chat is there to help the learner understand and work with the project.

For a complex program, the intended explanation workflow is:

1. Capture a screenshot of the relevant blocks.
2. Attach that image to the chat.
3. Ask the assistant to explain the code.

The two-screen setup prevents Druga from capturing the intended screenshot during the live demonstration. She stops the attempt without showing an attachment or a resulting explanation. The digit-recognition interaction is visible; the screenshot-to-explanation sequence remains incomplete onstage.

Druga closes by describing the project as free and open source, inviting contributions, feedback, and sharing with young people as more features arrive. The invitation extends the same participatory model used in the studies: the people who use these tools should help determine what they become.

She also invokes the EU AI Act as a reason AI literacy can no longer be treated as optional. In her account, a requirement had passed earlier that month and called on providers and deployers to support sufficient literacy among staff and product users. Her educational conclusion is to begin that work early, while children are learning how to investigate and build with AI.

Her final pointer is to papers and studies on her website covering AI literacy, AI education, mathematical misconceptions, and science. The work connects those subjects through a practical learning experience: children make something, encounter a surprising result, and gain enough understanding to decide what to try next.