A classic computer game that has gone by many names over the years relies on this trope. In the original version, you had to run from robots, although modern versions have used zombies, vampires, Eldritch Abominations... basically, whatever. Anyway, you and the robots both move one square per turn (like a chess king), and robots will chase you down. You have no weapon, but the robots will attack and annihilate each other before they ever turn on you! Thus, you have to rely on robots' tendency to kill each other before they kill you. It's even been done with Daleks.
Your wingmen in Aero Fighters Assault are generally more competent than those in Star Fox 64 (probably because the levels are so big and the draw distance so lousy that you'll almost never see them and they can kill enemies without actually having to kill enemies) but they're just as bad in regards to frequently asking for help and being generally helpless. They also will not even attempt to attack the stage bosses, and destroying the bosses is the one thing you need to do to finish 6 of the 8 levels. A harmless- though amusing- example in the dialogue: Volk will often follow up his "thanks for helping me" line with his enemy taunt, leading to him saying "Much thanks, Comrade; now you will die!" A more meta example in that the two levels where you have no wingmen have a significantly better frame rate- and thus much better controls- than the other levels of the game. The extra strain placed on the game by creating and keeping track of your wingmen slows the game down painfully in the bulk of the levels.
Akinator can be surprisingly daft at times, since he doesn't really understand the questions he's asking (they're all user-contributed); as a result, he may keep asking you almost identical questions, or directly opposite ones ("Is your character real?" followed by "Is your character fictional (does not really exist)?") Of course, he'll ask if they have black hair and then ask if they have blond hairnote Of course, some characters do have multiple hair colors..., plus the always entertaining "Asks if they're from one universe, is told yes, guesses a different universe". This is likely the only program to think Mass Effect and Halo are the same universe. And sometimes it will ask you if you've met/said hello to the character after you've confirmed it's a fictional character or dead historical figure. Apparently it believes you might be Thursday Next or a time traveler.
The Bot Wars expansion for the Battlestations board game implements this in a tabletop game. The player's opponents are rogue robots which begin the game with below-average skills, and can learn as the adventure progresses. However, the bots also have a huge mothership; ship size is a factor in the difficulty of piloting, the mothership is the highest size possible in the game, and the bots' skills are below average.. No matter how the GM runs the bots, unless he/she fudges the dice rolls, it is perfectly possible for the bots to wreck their own mothership in the opening scenario due to their incompetence in piloting it.
Buckshot Roulette: The dealer is poor at keeping track of the shells that have been fired. This can lead to baffling moments such as wasting a magnifying glass on the last round in the chamber, or even shooting itself with it for no reason.
In Danganronpa Another Episode: Ultra Despair Girls, the Monokuma enemies have a problem at getting around walls chasing after dancing Siren Monokumas. They just run around trying to get to it, but failing.
Demigod, an early MOBA, tends to inflict this on players when they go against the bots on higher difficulty levels. The opposing team will specialise in hit-and-run tactics, prioritise game-changers like Reinforcement Flags, and just generally give you a run for your money. Your allies, on the other hand, will position themselves directly between two enemy gun posts and pick on irrelevant minions, while being whaled on by the enemy, thus feeding your opponents both gold and experience. Since your opponents are now relatively stronger, and can afford to upgrade their defensive structures, this process becomes streamlined, resulting in ally deaths roughly every few minutes.
A few missions in Destroy All Humans! require the player to follow certain characters to specific places or hypnotize them to perform a certain task. One mission in particular has the player escort a character back to their saucer, and once they start running, they will not stop for ANYTHING. This includes a busy street, where this almost inevitably happens.
Perhaps the most glaring example occurs in Rockwell, where hiding beneath the bridge near the fairgrounds whilst being hunted causes the AI to fail hilariously at getting to the player. Most of them will pile up on the bridge itself, some even getting killed as more and more crash into each other while the enemies on foot occasionally walk off the edge and fall to their deaths. The rest that manage to get beneath the bridge where the player is hiding will drive their vehicles straight into the water.
The AI's pattern for enemies with multiple instances of the same countdown equipment (Sneezy, Rat King, and Buster, among others) always boils down to "fill in the leftmost equipment first". This can lead to these enemies making suboptimal plays if they partially completed the rightmost equipment on a previous turn — they won't finish it until they end up with a turn where they can complete all of the previous equipment or are blocked from doing so.
Sticky Hands will still use Pickpocket on you to steal your money even if you don't have any, and since he only has one dice, he'd be wasting his turn when he could've used it to charge Run Away instead.
The Haunted Jar will still use Blight on you even if you're not poisoned, usually when it can't use its Poison Cloud.
The Handyman uses the Spanner to get dice values greater than 6 so he can maximize his Hammer attack. However, if the Spanner's upgraded, he'll use it two more times with the resulting dice even if they'll give out the same value.
The Wolf Puppy won't bother to charge another Fury stack if he can't use Wolf Puppy Bite after activating Woof Woof Woof in the previous turn.
The A.I. doesn't know how to effectively use status effects in the Parallel Universe due to their different mechanics. The Cactus, for example, will still build up Thorns if he has only odds, even if the PU version of Thorns only works within the turn it was applied because it absorbs extra HP instead.
If an enemy's HP is 2 or less, they won't pick up a burning dice, knowing it would kill them. However, if the Rose Rule is active, they'll still pick it up at 3HP, thinking it'll still deal 2 fire damage to themself, but the extra 1 fire damage will do them in.
In Reunion, Buster can't do damage if he rolls all evens since he needs odds to use Incinerate. Despite this, he'll still use his evens to use Scorch, which increases the HP cost of each burning dice, even if he didn't burn any.
The Division 2 has some pretty good AI that is usually competent with regards to taking cover, using grenades to flush enemies out of cover, and providing support to one another. However, sometimes their priorities cause them to make tactical errors.
While the AI understands the concept of flanking and will attempt to use it, you will often see enemies deciding to run for a cover position that flanks the players but requires charging into the open and right past the player to reach, and they won't make any attempt to fire on the player while moving into position, giving the player ample time to gun them down.
Heavy elites, who are covered in a massive amount of body armor that has to be torn away before they'll take damage, are usually slow moving but can break into a sprint if they're being chewed up from a distance. They can also one-shot deployable items like turrets or healing hives with a melee attack. The issue is that they will immediately prioritize any deployable they spot that they have a path to, even if that path requires them to charge across a wide open area and climb several ledges to reach. This can easily be exploited by clever players by throwing a turret onto a high ledge that can be reached, and then tearing into the elite while he ignores them to go kick the turret.
Similarly, one class of heavy elite is armed only with a slegehammer, and attacks by charging players and delivering an extremely powerful blow. However, this attack can easily be dodged, meaning that killing this type of elite is no more difficult than repeatedly dodging and firing while he recovers from a swing. They're one of the only elite enemies that is easier to defeat if you don't use cover.
Elements: the Game: Some light spells have Revive Kills Zombie. On lower difficulties, the AI doesn't know this. Cue your digital opponent vaporising his own creatures.
The interactive game Façade (2005) can be beaten entirely by answering "yes" to every question, as proven by Brutalmoose in his video.
But if you want to end the game rapidly, you can just type in "melon" and watch as a greatly offended Trip kicks you out of the apartment. The game's speech recognition system detects when players say a vulgar word and makes Trip and Grace react appropriately. However, the game also tries to incorporate various slang terms as well, which can catch a player off-guard when one of these words is typed in a perfectly innocent context and Trip shoves them out the door. It's not the only example of a seemingly innocuous word causing Trip to prudishly force you out of the apartment, but "melon" is by far the most famous example.
This plagued computerised Go engines (especially when compared with computerised Chess engines) until a few years ago, with them being trounced by professional Go players even when given substantial advantages. However, the latest version of AlphaGo, an AI whose design represents a major breakthrough in computer Go, seems to be far stronger than any human player, going undefeated against top professionals. Part of the reason it took so much longer to build a professional-level Go program is that the search space is much larger compared to Chess — there are typically two or three hundred possible moves available to either player, and most games last over two hundred moves — so it's not feasible to analyze all possible sequences for more than a few moves ahead.
In Gruntz, this applies to both your gruntz and the enemy's. Your gruntz will walk in a straight line towards their destination, carelessly walking into any obstacle that is in their path. And both your gruntz and the enemy gruntz can be easily distracted by toys given to them, no matter what happens around them. If the toy is a mobile one, then it gets even better, since they will randomly drive into any instant-kill obstacle near them. On multiplayer levels, the enemy shows some remarkable amounts of stupidity, such as:
Walking straight into pyramids that are about to raise (instant kill), under rolling boulders (instant kill), onto bridges that are about to sink (instant kill)...
If they have giant straws and stand on a goo puddle, then they will glitch and just stand on the puddle instead of walking next to it to drink it.
They will break any bricks standing on their way to your gruntz, even if it means destroying their own fortress' defense.
The usual method to beat the last boss in Guitar Hero III invokes this. Basically, there's a certain point where a Whammy attack will kill him in one hit. Why is this? In that particular section, instead of using the whammy bar to recover, he just hammers the STRUM BAR until he kills himself. One critical flaw in an otherwise complete bastard.
In Happy Tree Friends: False Alarm, you just point and click things to help the Tree Friends avoid and/or survive obstacles and rescue them...except these animals really just don't want to co-operate. A Tree Friend will sometimes walk in the wrong direction or walk into the corner of something, clip out of bounds, and fall out of the worldnote When this happens, the Tree Friend will not die, forcing you to restart the level.. Sometimes they won't even react to fire like they are supposed to; normally they'll run away from it, but sometimes they won't react or will run to a direction you didn't intend to. Other times they react abnormally to a fire, they'll just run straight into it and take fire damage.
One animated board game for MS-DOS called Hexxagon was indeed a lot of fun. Pit red gems against chrome drops on a hexagonal board in deep space. Landing next to your opponent's pieces would transform them into your own. The hard Craniac was usually pretty good, but when it was running out of spaces to go, it tended to make stupid moves such as jumping pieces into spaces it could have cloned into and in the process of doing so, often opening up holes allowing its opponent to capture some of its pieces. These stupid moves usually cropped up when the Craniac was losing, so it rarely changed who won the game, although if you had been narrowly losing to it, such a move could turn the tide in your favour. On lower difficulties, the Craniac also tended to make stupid moves much more frequently, but in those cases, it was expected behaviour.
The AI in Hitman (2016) does not always handle hazards well. As seen in thisAchievement Hunter video, they will assume that the hazard has been neutralized (which is normally true), even if the corpses of dozens of other people making the same mistake are clearly visible in the hazard. To say nothing of how nobody thinks twice about stepping in the hazard in the first place...
The Hobbit (1982) is still something of a landmark in terms of immersive AI; the characters could act on routines, grow bored, or refuse orders. Unfortunately, the game was far too ambitious for its time, because instead of acting like the characters in the book, this meant everyone acted like suicidally insane weirdos suffering from Attention Deficit... Ooh, Shiny! Just for example, it's very possible to reach Laketown and discover that Bard the Bowman got bored of waiting, wandered off, and got eaten by wolves. And if he is still alive, it's not uncommon for him to point-blank refuse to kill the giant dragon barreling down on his hometown.
Old PC or video game versions of Jeopardy! in the early 1990s had the AI contestants buzz in and answer in complete gibberish. The answer pool was so small that pulling a wrong answer from that could clue another player in later. Other versions had no answer pool, resulting in the correct response or the same gibberish every time. Examples include ZWXYZ on the Game Boy and XXX on the Genesis versions. This is true for the NES versions as well (save for Super Jeopardy!), but the gibberish is the exact same length as the correct response, and often shows some letters in the response as well. For example, if a correct response is TVTropes, the AI would show something like *V@r#pes.
Exploiting the Artificial Stupidity of the guards in Lode Runner is very useful, with some levels relying on it. For instance, you can position yourself on a ladder so they climb upwards when you're directly below them.
The final boss of Magic: The Gathering – Battlegrounds has the ability to cast any spell in the game, any time he likes. Theoretically this means he should be able to spam you with giant monsters while countering any spell that you try to cast. Instead, he just sort of hangs around not doing much, and can be trapped in a loop by summoning the same low-level Mook over and over again. This is, of course, intentional, as that level of power would by simply impossible to oppose if he used it in anything like a sensible fashion, but it's rather unsatisfying to beat a boss that could curb stomp you at will for no other reason than that he was too dumb to actually do it.
Sometimes, the computer can come up with masterful combos and expert tactical plans. Other times: they sacrifice their last point of life to Pestilence in order to kill some Llanowar Elves, and summoning a Lord of the Pit and then doing nothing with it, meaning it eats all the computer's monsters and starts on the computer's life total. In particular, it will only attack if the creature is guaranteed to survive the creatures you have out or it has enough monsters to zerg you to death. This means that it doesn't, for example, fling expendable creatures at you to whittle down your forces, even if those creatures have a significant upkeep like sacrificing a creature.
Pestilence itself is a global enchantment that deals 1 damage to EVERYTHING ON THE FIELD (including BOTH players) for 1 black mana, repeated as long as you have black mana/creatures (Pestilence is destroyed if no creatures are on the field). Lord of the Pit is a creature with 7 power (deals 7 damage), 7 toughness (can take 7 damage), is flying (can only be blocked by other fliers or creatures with "Web"), and has FIRST STRIKE (His 7 damage can be spread however he likes first, and then any defenders still alive deal damage to him, when normally both sides deal damage at once). This came at the drawback of needing to sacrifice a creature every round, if you didn't, he did 7 damage to his controller. Many a game was won by simply wiping the creatures away, or stalling until your opponent ran out.
Force of Nature had 8 Power, 8 Toughness, and had Trample (and extra damage above and beyond a defending creature's toughness hit the player-usually extra damage is wasted). This was at the expense of 4 green mana every round, or Force of Nature did 8 damage to the player. Destroying land, however, is tough, but one card made it easier: Living Lands. Living Lands turns all Forests (which PRODUCE the vital green mana) into 1 power, 1 toughness creatures. The AI was more then happy to fling these pipsqueaks at you. You can see where this is going...Extra Stupid in that AI decks seemed to run both Force of Nature and Living Lands together. Only one deck had Force of Nature without Living Lands, and that one's a Green/White mix deck.
The AI had a real problem with Spirit Link, which causes a creature to grant you 1 life for each damage it deals. A classic trick is to play it on an opponent's creature, making it unable to harm you. The AI would completely ignore this and continue attacking with the creature. You could go even further, and stack up to four Spirit Links on the creature; the AI would continue to attack you with it without noticing that each attack was increasing your life by 3x the creature's power. This could extend to being outright risible when the AI began to sacrifice cards to buff the amount of damage it did!
The Windows program Mission Maker has extremely primitive AI. Make a character 'Seek and Destroy' the player, then get another character between them. The hostile character, instead of moving around, will kill the other character to get to the player.
For Monopoly video game adaptations, it runs into this trope with its AI:
The Nintendo Entertainment System has a noticeable flaw with its AI. It will get hung up on unrealistic trades that are not realistically possible. It's possible that it will try to trade Park Place or Boardwalk straight up to another AI player and it will usually get declined by that player. It will also waste several seconds proposing trades before either quitting or getting declined.
Ubisoft's Monopoly Plus has one major AI flaw — when trading, they fail to notice whether the properties you offer are already mortgaged. note Oddly, they do notice if you're trying to take a mortgaged property, and demand less accordingly, which suggests this error is more like a typo than a design failure. They also have the (rather more reasonable) behavior of always unmortgaging properties received in trades, whenever possible. Accordingly, it's possible to mortgage a property, sell it to an AI for full price, watch them spend their cash to unmortgage it, then buy it back for nearly the same price. Repeating this process allows you to bankrupt an entire table of AIs on turn 1.
The board game On The Underground features a rules-controlled tourist who will visit new destinations in London every turn. In the standard rules he will always favor riding a train over walking any distance, meaning that he will take a long and complex route around multiple Underground lines in order to reach a station that's right next to him. Since you score whenever he rides a line you own, accommodating his bizarre behavior becomes a significant part of the game. Arguably justified in that the abstract nature of the Tube map means that some tourists in London do actually do this since the true distances between stations are not clear - notorious errors being riding between Leicester Square and Piccadilly Circus (you can see one from the other) or changing from the Bakerloo Line at Baker Street to go to Great Portland Street (when staying on the Bakerloo Line would take you to Regent's Park, which is less than a minute's walk from Great Portland Street station).
Artificial Stupidity has been with us since the days ofPac-Man. Because of the way the AI (highly advanced for the time) was programmed without any RNG whatsoever, patterns were discovered that guaranteed that the ghosts wouldn't eat you, even up to level 256. One of the two big innovations on the Pac-Man formula that made Ms. Pac-Man so big was just randomizing the location to which each ghost went at the start, making such pattern play unreliable.
In the chess app Play Magnus, Magnus at low ages is programmed to refuse all draw offers, no matter how hopeless his position is. Even if any remotely competent player can checkmate him in a few moves, he won't take the draw. Even if a draw is the best he can hope for anyway, he'll insist that you keep playing. Word of God says that he does this to prevent players from accumulating a lot of points at the bottom levels with almost no work: if he accepted draws in clearly lost positions, you could just capture a couple of pieces in the first moves, and then offer a draw.
The AI in the Trainer Challenge section of the Pokémon Trading Card Game Online follows a number of rules that usually do a pretty good job of emulating real play, but sometimes result in them causing their own loss. Chief among them seems to be an absolute refusal to discard cards unless the text on a card they play forces them to. Because of this, they will never ever pay a Pokémon's retreat cost, which means that if they send in a Pokémon with no damaging move (and that is unable to evolve into a form with a damaging move), you can literally just sit there and pass turns until they run out of cards before you do. Which will almost certainly happen, since they will also play trainer cards that draw from the deck at the earliest opportunity, with especial fervor given to Professor Juniper/Sycamore/Researchnote User discards their hand and then draws 7 cards, exacerbating the process. Additionally, there are many Pokémon cards across the AI trainer's decks that they use incorrectly, or will stick to a set strategy that doesn't involve hurting you, even after they attach enough energy to them to use a damaging move. Examples are given below:
Zach's Skarmory and Rika's Klefki will only ever use Call For Familynote Searches deck for up to 2 Basic Pokémon and puts them on the bench, even if their bench is full.
Calvin's Floatzel will only ever use Rescuenote Shuffles 3 Pokémon from the user's discard pile into their deck, even if there's no Pokémon in the discard pile. Additionally, his Manaphy has no damaging moves.
Grayson's Plusle will only ever use Positive Handnote User discards 1 basic Energy card from their hand to draw draw 2 cards from the deck. Funnily enough, despite having a near-identical move selection, his Minun will only ever use its damaging move.
Juji's Hitmontop will only use Quick Drawnote Draw 1 card, Penelope's Deoxys will only use Close Encounternote Draw 2 cards. Can be used on the first turn of play, unlike most moves, and Otis's Yanmega will only use Windfallnote Shuffles the user's hand into their deck and then draws 6 cards.
In contrast to the rest, Kendall's Dedenne never uses its deck search move and is instead a nasty counter to your fully charged Mons with its Energy Shortnote 20 damage times the amount of Energy attached the opponent's Active Pokémon attack, but if you send in a 'mon with no energy attached, it has no way to hurt you. Similarly, Logan's Roserade will only use Whiplashnote flip a coin until you get tails; discard an energy from your opponent's Active Pokémon for every heads, flailing away at an energy-less monster for the same result.
Disappointingly for the final fight in the final league, Daniel's deck features a Solrock that will always use Solar Generatornote Searches the user's deck for up to 2 Special Energy cards, and reveals them to the opponent before adding them to their hand, even though he only has two special energy in his deck anyway. He also uses the same Dedenne card that Kendall has.
Mick is a tough opponent, with deadly annoyances in Simisear, Magcargo and Torkoal, and a brutal self-fueling Blaziken, and, though rarely played, the Camerupt EX that appears in his Platinum and City Championship decks can hit hard with its Tumbling Attack... But once it gets enough Energy to use Explosive Jetnote The user may discard as many Fire Energy attached to their Pokémon as they wish, dealing 50 damage times the number of Energy discarded, its threat is completely neutralized because, as stated above, the AI will not willingly discard Energies under any circumstance, and this particular card gives poor Mick the freedom of choice.
The reason for this appears to be that the AI is only ever prompted to attack the player when the AI attaches an energy card or a "Pokémon Tool" card to one of their own Pokémon. Therefore, if the AI has no Energy or Tools in their hand (and no Trainer cards that they can use to obtain them), they will not make an attack, even if their Active Pokémon has enough Energy attached to it to do so. The AI will also not attach more Energy to a Pokémon than it needs for its most costly move (even if it never uses the move, as demonstrated above), so in a scenario where the AI has Energy cards in their hand but all the Pokémon they have in play are "full"note and they have no Tools to use, they also won't attack.
Wheatley, also known as the Intelligence Dampening Sphere in Portal 2 is a deliberate In-Universe example, described by GLaDOS as "the product of the greatest minds of a generation working together with the express purpose of building the dumbest moron who ever lived", and "the moron they built to make me an idiot". It's actually not a totally straight example, as this A.I. was made to be like a stupid human (and programmed very well for that purpose), rather than simply a badly programmed computer.
If you've ever played a video game adaptation of a game show, you've probably encountered computer contestants that couldn't answer simple questions correctly. Press Your Luck for the Wii is one of the egregious examples, with computer opponents answering questions such as "What animal do we get milk from?", "What is 36 divided by 6?", or "How many months are in a year?" wrong.
In The Simpsons: Road Rage, buses constantly crash into anything in sight without any provoking them, typically you.
In Splinter Cell: Conviction, at one point you are confronted with an enemy helicopter gunship. It always shoots in front of Sam and never thinks to try and flank him.
In Star Control 2 any computer-controlled ship that attempts to fight an human-controlled Thraddash Torch will lost, as the AI will happily crash into the puffs left behind by the afterburners of that ship even if they are harmful, and will not even raise shields on ships that have them like the Utwig or the Yehat ones. Computer-controlled Ur-Quan dreadnoughts will happily launch fighters even against ships that can outrun them as the Pkunk Fury and will not use them against the Utwig ship, even if just two fighters lack enough firepower to recharge its shields (and using them against a lone fighter will drain the shields away.)
In some cases during combat, the AI in Star Control 3 will run away across the screen without even attempting to engage your ship, a pain in the ass if your ship cannot overtake the enemy one or lacks long-range weapons. This is so bad that the develpers defined one key to end with this behaviour and have the computer-controlled ship attacking you.
The Star Fox series has wingmen's "calling for a help" as a fixed pattern in every side scrolling stages. They can't help themselves and will go down if you don't help them. All-Range Mode, however, turns their stupidity up to eleven. One particularly notable example of how bad the wingmen's AI is in All-Range Mode is in the Star Wolf dogfights in Star Fox 64. Each Star Wolf pilot is programmed to target a specific member of your squadron. Each wingman will constantly plead for you to help him by shooting down the Star Wolf member who's on his tail. Once you do, he will blissfully fly around in a circle minding his own business and make no effort to help you as the remaining Star Wolf members continue to rip you and your other wingmen to shreds. In a way, this is a blessing, as you don't get points if your wingmen score the killing blow on something. In Sector Z, they become rather competent when it comes to taking down missiles, which is not a good thing for medal huntersnote and no, you can't just shoot down your wingmen; if any of them are down, you cannot earn a medal.
The AI for Talisman: Digital Edition ends up having fairly frequent moments where it shows that it is a different kind of special:
It will happily overload its inventory by crafting a raft at a woods space, then take a talisman from a player that attacks it before their next turn (despite already having one) and drop the raft it just spent the previous turn crafting, and then on its very next turn, craft another raft.
It will always drop the Amulet from its inventory even if it just picked it up off the ground, even if it is a character that usually doesn't have enough Craft to even hold on to a single spell.
It will always pick up every item it encounters that it legally can, including the previously mentioned amulet. When this puts it over the inventory limit, it may discard a much more useful item (such as a weapon)
Enemies in the original Trine are not clever when it comes to water, which kills them instantly, and falls, which do likewise if of sufficient distance. There's even one boss-level enemy - the second crystal troll you face - which can be lured into charging over a cliff edge, leading to instant death. The goblins in the sequel are a bit smarter, although they're still not good with environmental hazards or not shooting each other.
The first video game based on the Wipeout (2008) television series has downright idiotic AI. There are only a small handful of obstacles in the entire game that the AI is capable of navigating. Bump the difficulty past the easiest level and the AI will complete one obstacle each in the first and final rounds and wipe out on everything else through zero interference on your part. A heaping dose of The Computer Is a Cheating Bastard is used to offset these tendencies.
In Wonderland AdventuresNPCs will make no effort to avoid danger when you tell them where to go, and will always take the same path between tiles even if it kills them.
In Dot Mania mode (collect 80 Pac-Dots first to win), AIs have a tendency to run into ghosts and hazards even if there's an alternate path that they can take.
In Da Bomb mode (a 30-second bomb is held by one player, who must pass it to another; be the last one alive to win), AIs often make dumb decisions, like running to the back in a straight line into a wall after tagging a living player, thus making them easier to get re-tagged again.
In general, when they're not blatantly cheating when it comes to the rules, the multiplayer AIs tend to run back and forth or use portals for no reason. Both cases can leave them defenseless to ghosts and hazards.
almost inevitably happens.