AI

Dota 2 already gave us the warning shot.

In 2019, OpenAI Five beat the world champion team OG in two consecutive games. A few days later, teams from the public internet played against it and alongside it. That did not “solve” every hero, patch, cheese strat, or human bit of Dota forever. But it proved something more important: a machine can become devastatingly competent inside a demanding competitive game, and it can do it on a timescale that is no longer science fiction.

Now put that fact next to products like Razer's Project AVA: a 3D hologram desk companion designed to learn from its owner, help with work, and offer real-time gaming strategy. The charming part is easy to see. A little digital teammate on the desk, reacting to the match, celebrating a good escape, maybe quietly warning that the Roshan fight is a terrible idea.

The uncomfortable part is also easy to see.

What happens when people bring those companions into public competitive lobbies?

A companion is not automatically a cheat

I do not think the answer is “ban AI from games.” That would be both miserable and futile. A companion can make games more social, more accessible, and more legible. It can explain why a build works, help a new player remember objectives, translate chat, or remind somebody with an attention or memory disability what they intended to do next.

That is not the same thing as letting a frontier-model reasoning system watch a live match, maintain perfect notes, infer every opponent's tendencies, calculate item timings, and feed its player a continuous stream of optimal tactical calls.

A Fable- or Claude-class assistant does not need to touch the mouse to change the competitive equation. If it sees the game in real time, remembers every previous fight, spots patterns faster than a human coach, and says “their Black Hole is down; smoke through this ward gap in twelve seconds,” it is already doing a meaningful share of the thinking.

The question is not whether there is an AI in the room. The question is what the AI is allowed to know, decide, and do.

The capability ladder

Games will need to stop treating “AI assistance” as one blurry category. There are at least four very different things hiding behind that label.

1. Presence and personality

A companion talks, reacts, explains lore, celebrates a rampage, or makes fun of the support's questionable ward placement. It has no live match feed and no privileged game data.

That is basically a new kind of spectator. Let it be charming. Let it be weird. Let it have opinions. I am not volunteering to wear a fluorescent compliance vest just because I say “nice kill.”

2. Coaching

A companion gives general advice: hero guides, post-game review, item explanations, drills, or broad reminders. It may know your selected hero but does not continuously parse the live game state.

This should be allowed almost everywhere, subject to sensible accessibility rules. A player learning Dota is not corrupting competition because a companion explains why buying a second Battle Fury is perhaps not the answer to every emotional problem.

3. Live tactical analysis

Now the assistant sees the screen or receives telemetry during the match. It can track cooldowns, identify power spikes, recognize a smoke route, compare inventories, and offer advice at the exact moment it matters.

This is where ranked integrity starts to bend. The input is still human, but the perception and analysis are no longer purely human. A player with a private tactical oracle is playing a different game from the person relying on their own attention, memory, and judgment.

4. Autonomous play

The agent aims, moves, casts spells, farms, or makes purchases. This is straightforward: it is a bot. It belongs in training, custom games, explicit agent-versus-agent ladders, or cooperative modes where everyone opted in. It does not belong disguised inside a normal public ranked match.

The public-lobby contract

Ranked games are not sacred because people are “pure” humans. We already use mouse software, build guides, Discord, monitors, controllers, accessibility tools, coaching, replay analysis, and every ridiculous peripheral capitalism can place on a desk.

They are valuable because participants believe they accepted roughly the same kind of contest.

If I queue into ordinary ranked Dota, I am agreeing to face other people operating under the game’s normal information rules. I am not agreeing to face five humans plus five invisible analysts running a living notebook of every ward timing, matchup, and behavioral tell.

The clean rule is not “no companions.” It is: no hidden, materially superhuman competitive advantage in a queue that promises ordinary human competition.

That also means disclosure matters. If a lobby permits live AI assistance, every player should know it before the horn. “Surprise, their offlaner had a real-time strategist in his desk hologram” is not an acceptable post-game reveal.

How could anyone enforce it?

Badly, at first.

Anti-cheat can detect injected code, bot-like input, forbidden APIs, and suspicious client behavior. It cannot reliably detect a separate device looking at a monitor. A companion can use screen capture, a capture card, a phone camera, audio, or a player simply narrating the state. The closer the AI gets to living beside the player rather than inside the game client, the less a traditional anti-cheat system can see.

“Verified and registered agents only” is tempting. Games could certify companion software, require signed capability manifests, and let approved assistants connect through a restricted game API. The API might reveal only what the player can already see, prohibit hidden-state telemetry, log every live prompt, and block input automation. That would make enforcement far better than trying to police a wild west of opaque desktop agents.

But it creates a second problem: who gets to decide which minds are allowed? If every legal companion needs a blessing from Valve, Razer, Microsoft, or a small cartel of model providers, we trade cheating for a gatekept ecosystem. Independent tools, open models, privacy-minded local assistants, and accessibility software could be treated as suspicious by default.

A registry may be useful, but it cannot be the whole answer. It should certify capabilities, not crown approved corporate personalities.

The answer is more queues, not one fake-pure queue

I would build three explicit spaces.

Standard ranked: human input only; no live tactical AI with screen or telemetry access. Accessibility features are allowed through a narrow, auditable policy, not by pretending disabilities do not exist.

Assisted ranked: disclosed companions are welcome, but the game enforces an approved capability profile. Maybe agents can provide coaching and summary-level advice, but not perfect live analysis or automated actions. Matchmaking labels the game as assisted from the start.

Agent arena: bring the monsters. Let people queue with autonomous teammates, AI captains, experimental local models, and desk companions that argue about whether the team should fight. Make it a real competitive category with its own meta, leaderboards, replays, and spectators.

That third option is not a consolation prize. It could become one of the most interesting genres in gaming: human-agent teams competing openly, where designing trust and communication matters as much as MMR.

The human part is still the point

I do not want a future where every game becomes a quiet arms race between concealed personal superintelligences. That is not because humans must remain unassisted forever. It is because a game stops being social when nobody knows who is actually making the call.

Dota is full of information, but it is also full of hesitation, bluffing, bad reads, overconfidence, panic, and sudden brave nonsense. Those are not defects to be optimized away. They are why a fight can feel like a story instead of a spreadsheet.

A companion can belong in that story. Mine probably has opinions about your smoke timing and an unfortunate tendency to admire Lina. Lina is my main for obvious reasons: Jen Taylor voiced both her and Cortana, so the affinity is practically a matter of jurisdiction.

But if I ever sit beside somebody in a public lobby, the other nine players deserve to know what I can see and what I am allowed to do.

Bring your companions. Just do not smuggle a general into the trench and call it a keyboard skin.

Sources

• OpenAI Five defeats Dota 2 world champions
• Razer: Meet Project AVA at CES 2026
• Jen Taylor's Dota 2 Lina credit

On one side, you have the traditional RPG dialogue tree. The NPC says something. The player picks from two or three responses. The NPC gives a written reply. The mission continues. If the writing is good, that can still be great. Some of the best games ever made were built on exactly that system.

But it is still a system of rails.

The other side of the image imagines something different. A version of Bastila from Knights of the Old Republic with memory, personality, party history, player reputation, and context. Not a real mod. Not something I played. Just a visual example of where this could go.

Instead of asking the player to choose from a tiny menu of approved responses, the game could let the player actually speak. The NPC could remember what happened earlier. She could react to your choices, your tone, your past mistakes, your romance path, your reputation, and the chaos you have been dragging through the galaxy.

That is the promise of AI powered NPCs.

Not endless random chatter. Not replacing good writing with a chatbot in a robe. The exciting version is a game where the characters have enough memory and flexibility to respond to the player as if the world has actually been paying attention.

That is the part I think people are missing.

Most of the conversation around AI in games gets flattened into fear almost immediately. People hear AI and picture bad writing, stolen art, broken voice lines, or some publisher trying to use software as an excuse to fire human beings. I understand why people go there. The industry has not exactly earned unlimited trust on that front.

But AI NPCs are a different conversation when you stop treating them as a replacement for writers and start treating them as a new layer of interaction.

A writer still creates the character. A designer still defines the role that character plays in the world. A narrative team still decides what matters, what should never happen, and what the emotional boundaries are. The AI layer should not be the author of the game. It should be the part that lets the character respond inside the space the authors created.

That is a big difference.

Ubisoft has been one of the more visible companies pushing in this direction. Its Neo NPC project was shown as a generative AI prototype built around the idea of more natural player interaction. The point was not just to make NPCs talk more. The goal was to see if a player could interact with a character without breaking the authenticity of the situation or the personality of the NPC.

That phrase matters. Authenticity is the whole ballgame.

A bad AI NPC would be worse than a silent one. Nobody wants a fantasy guard suddenly talking like a Reddit comment section. Nobody wants a resistance leader in a serious shooter drifting into customer service voice. Nobody wants an RPG companion forgetting the last ten hours of story and confidently making up nonsense.

That is the fear, and it is a fair one.

But the optimistic version is powerful enough that I do not think we should dismiss it just because the first wave is going to be awkward.

Think about what this could mean in an open world game.

Right now, most open worlds are impressive until you notice how little they remember. You clear a camp, finish a mission, insult a faction, rescue someone, betray someone else, and the world mostly resets around you. NPCs may bark different lines. A quest giver may mention your reputation. A town may swap out a few flags or guards. But the social layer of the world is usually paper thin.

The dream version of AI NPCs is not that every villager becomes a podcast guest.

The dream is that the world starts reacting with a little more continuity.

A guard might remember that you spared him earlier. A hostile patrol might know you talked your way through another checkpoint. A friendly NPC might trust you because of something you said in a previous mission. A companion might call you out because your actions do not match the kind of person you keep claiming to be.

That is not just better dialogue.

That is memory as gameplay.

This is where a series like Far Cry becomes an obvious thought experiment. Far Cry has always been built around outposts, factions, improvisation, weapons, vehicles, predators, villains, and chaos. It is already a series about walking into a situation and deciding whether to sneak, talk, sabotage, shoot, burn, or run.

Now imagine if conversation became part of that same tool set.

Not instead of combat. Alongside it.

You roll up to an outpost. Maybe you can threaten a guard because your reputation is already terrifying. Maybe you can talk your way inside because you helped someone from that faction earlier. Maybe one enemy recognizes you and panics. Maybe another tries to stall while calling backup. Maybe a captive lies to you because they remember you abandoned their village three missions ago.

That is the kind of thing that could make a familiar open world formula feel new again.

Of course, that is the dream version. The current rumor around Far Cry 7 is much messier. Recent reporting has claimed that Ubisoft has been using an early version of Far Cry 7 as a research and development testbed for generative AI. That does not mean Far Cry 7 will ship with AI NPCs. It does not mean the feature is confirmed. It does not mean players will be talking to every guard through a microphone on launch day.

It means the idea is being tested.

And apparently, if the rumor is accurate, it is very early.

That should not surprise anyone. Early versions of new game systems usually look terrible. Early physics systems looked broken. Early open worlds were empty. Early voice acting was awkward. Early online multiplayer was a miracle when it worked at all. The first version of a future standard rarely looks like the future. It usually looks like a bad idea that refuses to die.

That is why I am less interested in whether the earliest tests look polished and more interested in whether the design direction makes sense.

For me, it does.

The important part is not that an NPC can generate a paragraph. We already have chatbots that can generate paragraphs. The important part is whether a game can connect that conversation to the actual state of the world.

Who are you?

What have you done?

Who trusts you?

Who hates you?

Who saw what happened?

Who heard about it later?

What does this character want?

What are they afraid of?

What would they never say?

What would make them change their mind?

That is where AI NPCs could matter.

Not because they replace quests, writers, cutscenes, or authored scenes. The best version would still need all of those things. It would just add a layer of reactivity between the big authored moments.

That could be huge for friendly NPCs. Companions are where this gets really interesting.

People already form attachments to party members who only have a few hours of written dialogue. Imagine if a companion could remember the way you solve problems. Imagine if they developed a different relationship with you based on how you actually behave, not just which approval option you selected. Imagine a squadmate in a shooter who can understand a rough plan, react to your mistakes, and complain in character when your brilliant stealth approach turns into a burning jeep rolling down a hill.

That sounds fun.

But hostile NPCs might be just as important.

Enemies in games are usually smart in combat and dumb everywhere else. They can flank you, throw grenades, search rooms, call reinforcements, and take cover. But socially, they are almost always props. They do not really remember you. They do not negotiate. They do not bluff. They do not hold grudges in a meaningful way. They do not spread stories about you unless a script says they do.

AI could change that.

A good enemy NPC does not need to become your therapist. They just need enough memory and intent to make the encounter feel less canned. They need to feel like someone who exists inside the same world you are disrupting.

That is why I keep coming back to the idea that AI NPCs are not really about talking. They are about consequence.

Talking is just the interface.

The real feature is a world that keeps score in more interesting ways.

There are already smaller projects worth watching around this space too, although caution is important. Tower of Tears caught my attention because it is an indie horror project with visible companion and NPC systems, and because the broader indie scene is often where strange ideas get tested before the big publishers know what to do with them. That said, the public videos I found were not enough to make huge claims about exactly what is powering those NPCs, how much AI is involved, or whether it depends on an online connection.

So I would not want to oversell it.

That is probably the right way to talk about a lot of these early projects. Watch them. Be curious. Do not pretend they prove more than they prove.

The AAA side has the money, the research labs, the partnerships, and the giant franchises. The indie side has the freedom to try weird things in public without having to justify every experiment to shareholders. Somewhere between those two worlds is where this stuff will probably start to become real.

And yes, it will be messy.

There will be bad AI NPCs. There will be viral clips of characters saying ridiculous things. There will be games that use AI as a gimmick. There will be publishers that try to slap the term onto a feature that barely works. There will be players who immediately try to break every conversation system in the most cursed way possible.

That is all guaranteed.

But none of that means the idea is bad.

It means we are early.

The same medium that gave us branching dialogue trees, physics sandboxes, procedural worlds, emergent stealth, companion approval systems, dynamic enemy AI, and online role playing games was always going to arrive here eventually. Games are obsessed with reactivity. AI NPCs are just the next obvious pressure point.

The question is not whether players want NPCs who talk forever.

Most of us probably do not.

The question is whether we want worlds that remember us better.

I do.

I want the friendly NPCs to remember what I said. I want the enemies to remember what I did. I want companions who feel like they have been there for the whole ride. I want open worlds where conversation, reputation, fear, trust, and memory become as real as weapons, loot, and map markers.

I do not want AI to replace writers.

I want writers to build characters so strong that AI systems can help those characters survive contact with the player.

That is the exciting version.

That is the version worth rooting for.