AI Suspects for Detective and Murder Mystery Games

In most detective games the interrogation is a dialogue tree, and the real puzzle is guessing which option the writer meant. Neocortex lets you build AI suspects the player questions in their own words: each one knows what happened, hides what it has to, gets nervous when the questions land close, and only reaches a confession when the player has earned it.

What it looks like

Port Vesper is a rain-soaked harbour city in an original noir game. Augie Strand, owner of the Blue Heron jazz club, was found dead in his office after closing. The detective has two suspects in two rooms. Delphine Ashcombe, the club's singer, is calm, precise and gives away nothing. Teddy Kask, the bartender, is friendly, talks too much and keeps wiping a glass that is already dry.

Teddy knows more than he says. His knowledge base holds the night as he saw it: he stepped into the alley for a smoke at one in the morning, saw Delphine leave by the back door, and heard nothing from the office. It also holds what he hides: he owed Augie money, and he tore his own name out of the club's ledger before the police arrived. The player found that torn page in the alley drain twenty minutes ago.

Detective: You said you locked up at midnight and went straight home. Teddy: That's right. Straight home. Ask anyone. Detective: Then explain why this ledger page with your name on it was in the alley drain. Teddy: Okay. Okay. I took the page. I owed him, that's all. I didn't touch him, I swear it.

Teddy's first answer comes back with a relaxed emotion tag, his second with fear. The game plays it on Teddy's face: eyes down, hands still at last. That reply also returns the flowState alibi_broken, and the case board pins Teddy's real timeline next to Delphine's. He has admitted the debt, but the End node marked confession is still waiting, and the player now has a reason to go back to the singer.

How Neocortex makes it work

• Knowledge base. Each suspect gets a PDF or TXT document with what they saw, what they hide and what would make them admit it. Free includes 1 upload, Starter 5, Pro 20 and Team 100, which covers a large cast with a document each.
• Character design. Personality and backstory in the visual editor decide how each suspect lies. Delphine deflects with short, polite answers. Teddy over-explains and contradicts himself.
• Emotions. Every line arrives with an emotion tag. Drive faces, posture and voice from it, or show a subtle tell in the UI for players who are paying attention.
• Conversation Flows. Lay out the interrogation with a Start node, Checkpoints like alibi_given and alibi_broken, and End nodes like confession or lawyer_requested. Each reply returns a flowState your case logic reads. Flows are on every plan.
• Events. When the player slaps evidence on the table, send it as an event, like "the detective shows the torn ledger page". The suspect reacts to the evidence, not just the words.
• Session memory. Keep each suspect's sessionId. When the player comes back with new evidence, resume the player's session and the suspect knows exactly what they already claimed.
• Voices. Streamed speech gives each suspect a voice of their own, and player questions can be spoken through speech-to-text. Custom voices, cloned or designed from a description, are on Pro (5) and Team (20).

Ideas to build

The cosy village mystery. Someone poisoned the prize marrow at the Little Ashby flower show. Six villagers, each with a small secret that has nothing to do with the crime: an affair, a forged recipe, a borrowed lawnmower. Their knowledge bases make every one of them act guilty about something.

The courtroom witness. In a legal drama, the player cross-examines a witness who saw the accident from a bus stop. The flow has a Checkpoint for contradiction_found and an End node for testimony_changed, and your trial system scores the cross-examination when it lands.

The android who might be malfunctioning. In a sci-fi station, a service android was the only witness to an airlock death. Its knowledge base holds corrupted logs it believes are true. The player has to work out whether it is lying, broken, or covering for someone.

The final reveal at dinner. Gather every suspect at one table in group chat (Pro and Team plans, up to six characters) for the accusation scene, and let them talk over each other as the detective lays out the case. Flows don't run in group chat, so do the individual interrogations one-to-one first and let your game track the reveal.

Building it

Detective games suit any platform Neocortex supports: the Unity SDK, the Unreal Engine plugin, the Godot addon (beta), or the Web API for browser and narrative games built in any engine.

1. Write the truth first. Build the full timeline of the night before writing a single character. Every suspect's story has to fit inside it.
2. Split it into knowledge base files, one per suspect: what they saw, what they hide, and what pressure or evidence would make them admit each part.
3. Build the interrogation flow with the beats your case depends on, and name the keywords after what your case board needs to know.
4. Design each suspect: personality, how they lie, how they crack, a voice.
5. Try to crack them in the dashboard. If a suspect confesses on the second question, tighten the backstory. If nothing works, write down what should.
6. Wire it up: evidence as events, emotions to animation, flowState to the case board.

Be explicit in the knowledge base about conditions. A line like "Teddy only admits the ledger if shown the torn page" gives the player a fair puzzle, and keeps the solution in your hands, not left to chance.

Suspects are one of our AI NPC, ambient AI and voice command use cases. For the people who send the detective on the case, read AI quest giver NPCs, and for a narrator who sets the scene, see an AI game master and narrator. The knowledge bases and emotional intellect pages explain both features in detail. Start with one suspect and one secret worth hiding.

Checked October 2026.

Frequently Asked Questions