What is Spark Program and how does it use generative AI to develop your abilities?

Spark Program is yeshcube’s research and development line for generative artificial intelligence applications that help people think, practice and create without replacing their intellectual work.

The program starts with a design question: what happens when AI stops delivering the result directly and begins to intervene through questions, objections, constraints and feedback? Spark turns that question into specific applications and evaluates whether participation, judgment and autonomy remain when the person stops using the tool.

The problem: completing a task is not the same as developing an ability

Generative AI can draft, summarize, argue or propose alternatives in seconds. That speed can improve the immediate result, but it does not in itself demonstrate that a person understands the task better or can repeat it independently.

When the system takes on operations such as selecting, connecting, synthesising or deciding, cognitive offloading occurs: some of the effort moves to the tool. This can be useful, as it is with a calculator or search engine, but it changes which abilities the person practices during the activity.

The available evidence shows that design matters. A field experiment with nearly one thousand students found that access to a general generative assistant improved performance during practice but could harm later unaided performance; a version with pedagogical safeguards mitigated that effect (Bastani et al., 2025). In professional work, research involving 319 people and 936 use cases found that greater confidence in AI was associated with less self-reported critical-thinking effort, while cognitive work shifted toward verifying and integrating responses (Lee et al., 2025).

These findings do not establish that generative AI develops or diminishes abilities in every setting. They do support the question that organizes Spark: how the system intervenes matters as much as what the model can do.

The Spark Program hypothesis

Spark investigates whether AI designed as a critical interlocutor can improve the quality of a process without taking it over.

Instead of writing a conclusion, it can ask for a justification. Instead of choosing an idea, it can test it against defined objectives. Instead of removing a contradiction, it can flag it so the person decides how to resolve it.

Technology provides context, challenge and feedback. The person retains the final wording, the decision and responsibility for the result.

Three design principles

Activation rather than substitution

The application identifies the cognitive operations that matter for a task and avoids fully automating them. It can detect a gap, put a question or present another perspective; it should not turn the person into a passive supervisor of an already completed result.

Productive friction

Removing every difficulty speeds up execution, but it may also remove practice. Spark keeps obstacles that require arguing, revising or choosing when that difficulty serves the objective.

Friction does not mean making an interface deliberately awkward. It must be understandable, proportionate to the person’s level and easy to remove once it no longer adds value.

Progression toward autonomy

Support works as scaffolding: it appears when it helps someone move forward and decreases once they can apply the strategy. The objective is to reduce the help needed to solve comparable tasks with independent judgment. The number of interactions with AI does not measure that progress.

What a Spark application can develop

Spark can be applied to activities where the process matters as much as the result:

• Writing and building a distinctive voice.
• Critical thinking and evaluation of arguments.
• Creativity and exploration of alternatives.
• Problem-solving.
• Decision-making.
• Learning and practicing procedures.

Each application must define a specific ability. “Improving creativity” is not enough as an evaluable objective: the project must specify which behaviors it expects to observe, in which population, for which tasks and against what comparison.

How an application is designed and validated

A Spark application organizes development into five steps:

1. Define the ability. Specify which operation it aims to activate and how it appears in a real task.
2. Assign the role of AI. Determine what it may question, challenge or suggest and which decisions always remain with the person.
3. Design the friction. Introduce feedback or constraints that prompt revision without adding arbitrary load.
4. Reduce support. Adjust assistance as autonomy increases.
5. Evaluate transfer. Test whether the person sustains the practiced strategies on new tasks and without the tool.

The final stage is essential. A better text, a faster decision or a higher score during use measures assisted performance. Evidence of ability development also requires unaided performance, transfer and persistence over time.

What Spark Program measures

Evaluation is adapted to each application, but it should distinguish at least five dimensions:

• Quality of the result during use.
• Intellectual participation and decisions retained by the person.
• Amount and type of help required.
• Performance on a comparable task without assistance.
• Persistence of improvement and transfer to new situations.

Comprehension, usability, dropout, errors and acceptance are also recorded. Internal product metrics can guide design, but they do not replace independent evaluation of the ability a program claims to address.

PunkVoice, the first commercially available application

PunkVoice applies the Spark model to professional writing on LinkedIn. Its assistant, Sheena, checks each draft against the account’s profile, pillars and publishing history. It detects clichés, raises objections and flags inconsistencies; it neither writes nor publishes on the person’s behalf.

PunkVoice holds ERL-3, initial transfer: it is commercially available and used outside the research setting. Its next evidence challenge is longitudinal: establishing whether people progressively require less assistance and sustain outside the application the editorial strategies practiced within it.

Its operation, data regime and validation status are explained in what is PunkVoice.

Which projects can join Spark

Spark Program is open to internal or external applications that:

• Define a specific, evaluable human ability.
• Keep authorship and meaningful decisions with the person.
• Use AI to question, challenge, guide or provide feedback.
• Incorporate verifiable progression toward less dependence.
• Measure transfer and unaided performance, not only usage or satisfaction.
• State their limitations, data practices and models used.

Applications whose main proposition is producing the final result with the least possible human participation fall outside the program, even when they use the same kind of generative model.

Current status and collaboration

Spark Program is active and has one transferred application, PunkVoice. Current work combines its follow-up with identifying new settings where AI can act as a tool for thought, practice and creation.

The program is seeking research groups in human-AI interaction, learning, creativity, writing, metacognition and ability assessment; organizations able to provide testing settings; and technology projects aligned with its principles.

Discover PunkVoice →

Discover the Allies program →

Contact yeshcube →