How to Turn Research Into a Game Mechanic for PBL
A practical framework for adapting 1980s detective game mechanics to structure inquiry-driven learning missions across any subject.
The 1985 geography game ‘Where in the World Is Carmen Sandiego?’ integrated research directly into its gameplay, students can now reconstruct this core mechanism across science, history, or almost any subject. Applying this methodology within PBL Future Labs learning missions provides a practical framework for learners to design, test, and deploy their own inquiry-driven games as structured project deliverables.
Background
Where in the World Is Carmen Sandiego? came out on April 23, 1985. It was developed first for the Apple II and became the beginning of the much larger Carmen Sandiego series.
There was something odd inside the box. A computer game had been packaged with a physical copy of The World Almanac and Book of Facts.
Gary Carlston, one of Broderbund’s founders, had enjoyed searching through an almanac as a child and suggested building it into the game. Programmer Dane Bigham later recalled that including a reference book with a computer game was unusual. It was also heavy and probably annoying to ship.
Just like in a well designed research project, the almanac had a simple job.
Players worked as detectives for the ACME Detective Agency. A witness might mention a currency, language, flag, landmark or historical group connected with the suspect’s next destination. The player searched the book, compared the clue with the available flight routes and chose where to travel.
Choose badly and you lost time.
One clue preserved by the Computer History Museum said that the suspect wanted to study Saxon manuscripts and did not like seafood. The player had to connect the Saxons with England, notice that London was an available destination and decide whether that was enough evidence to go there. Another clue mentioned fly fishing on the Kenai Peninsula. That pointed towards Anchorage.
The manual opened by telling the player, “Wake up, kid. Got an assignment for you.” It also warned that there was limited time to solve the case. Suddenly, searching an almanac had some pressure around it. You were not reading a page about currencies because a teacher had assigned it. Someone had stolen something and was getting away.
This is close to how I think about a PBL Future Labs Learning Mission.
Start with the learner’s role and give them a problem. Put the curriculum inside the evidence they must examine. Give them resources that do something during the mission. Their decisions should have consequences, even small ones, and the final solution needs to show what they actually understood.
A teacher could begin with one curriculum standard and build a case around it. Students might work as marine biologists tracing damage through an ecosystem. They could become historians checking whether an artefact is genuine, journalists investigating an AI-generated claim, or engineers trying to keep clean water moving through a city.
AI can help build the case files, witness statements, maps, images, false leads and decision points. It can generate a lot of shiny rubbish too. The test is whether students need the intended knowledge to finish the mission. If they can skip the research and bluff their way to the final answer, the design has failed somewhere.
Carmen Sandiego managed this with an Apple II, an almanac and a fictional crime agency in 1985.
Research as a game mechanic
Bigham later observed that older players often guessed the destination, while younger players were more likely to search the almanac. Younger players had less existing geographical knowledge, but the reference book gave them a way to continue rather than simply fail the clue.
This made looking something up part of the action. A gap in knowledge produced the next move.
A 2014 study by Matthias Gruber, Bernard Gelman and Charan Ranganath examined how curiosity affected memory. Participants answered trivia questions while undergoing functional magnetic resonance imaging. Information connected with questions that created high curiosity was remembered more successfully, including incidental information presented during the same period.
Carmen Sandiego used a similar structure, although it was designed decades before that study. The player encountered a clue, realised they lacked a piece of information and searched for it because the suspect was getting away. The location fact had a job inside the story.
The warrant system
The warrant was another important learning mechanic.
Finding the thief’s location was not enough. Players also gathered personal information from witnesses, such as the suspect’s hair colour, hobbies or preferred vehicle. They entered these details into the Crime Computer, narrowed the suspect list and requested an arrest warrant.
When only one suspect matched the evidence, the computer issued the warrant. The player then had to catch the correct person before the deadline. Reaching the final destination without the correct warrant could still cause the case to fail.
This structure separated two forms of evidence. Geographical clues identified where the thief had travelled. Witness descriptions identified who the thief was. Players had to manage both at once, and sometimes the clues were slightly silly.
That mechanic has a direct use in classrooms where students can generate plausible explanations with AI before checking whether the explanation is accurate. A warrant forces the student to commit to a conclusion. The game can then reject that conclusion when the evidence is missing, contradictory or too thin.
It also gives the teacher somewhere useful to look. The important evidence is not only the final answer. It is the route the student followed, the clues they accepted and the point where their reasoning went wrong.
Building a science version in Google AI Studio
The following example is based on NGSS MS-LS2-1, which concerns the effects of resource availability on organisms and populations within an ecosystem.
The case begins with a villain stealing a keystone species from a North American ecosystem. The player follows the resulting population changes through four locations, identifies the missing species and explains what happened after it was removed.
Paste the following instructions into Google AI Studio:
Build a single-file HTML game. Keep the code as simple as possible:
one HTML file, plain JavaScript, no frameworks, no libraries, no
build tools. A beginner must be able to read it.
THE GAME
"Ecosystem Detective" — a text chase game for Year 7 science.
A villain stole one keystone species from a North American ecosystem.
The player visits 4 ecosystems, interviews witnesses, then names the
missing species.
THE UI (keep it minimal)
- A story panel showing the current scene text
- 3 buttons for choosing a witness
- After location 4: a text input to name the species, a text area to
explain the chain of effects, and a "File Warrant" button
- Dark noir styling: dark background, light text, one accent colour
THE AI
Every scene, clue, and warrant judgement comes from the Gemini API
(gemini-2.5-flash). Send the full game rules as the system
instruction and the conversation history with each call.
System instruction for the AI:
- Never name the missing species. Clues only describe population
effects: prey booms, predators starve, plants stripped.
- Each scene: 3 short sentences describing the ecosystem, then 3
witnesses to pick from. Year 7 reading level, noir detective tone.
- Chosen witness gives one clue, then the villain leaves a short
taunting note hinting at the next location.
- Warrant is correct if the species is right AND at least 2 effects
are linked in order. If wrong, say which clue was misread, allow
one retry, then reveal the answer.
- Reply in JSON: { "scene": "...", "witnesses": ["...","...","..."],
"gamePhase": "playing" | "warrant" | "won" | "lost" }
Keep total code under 250 lines. Comment each section in plain
English so a teacher can follow it.References
The Strong National Museum of Play. “Where in the World is Carmen Sandiego?”.museumofplay
Gruber, Matthias J., Bernard D. Gelman, and Charan Ranganath. “States of curiosity modulate hippocampus-dependent learning via the dopaminergic circuit.” Neuron 84, no. 2 (2014): 486–496. PMID: 25284006.pubmed.ncbi.nlm.nih
Nature. “How curiosity enhances learning.” Nature 514 (2014): 143.nature
Phil