# 24 AI Failure Modes and the Control That Catches Each One | IgnatiusTheYoungerAI > The 24 documented ways AI-assisted work fails at review, each with the mechanism, who carries the downside, the human control, and its Judgment Multiple. - canonical: https://ignatiustheyoungerai.com/failure-modes - html: https://ignatiustheyoungerai.com/failure-modes - machine-view: https://ignatiustheyoungerai.com/failure-modes?view=machine - markdown: https://ignatiustheyoungerai.com/failure-modes.md - site-index: https://ignatiustheyoungerai.com/llms.txt - evidence-labels: RESEARCH · REPORTED · ANALYSIS · MODELED · RECOMMENDATION · OBSERVED — carry the label with the claim; MODELED is arithmetic from stated assumptions, never a measured outcome - publisher: IgnatiusTheYoungerAI (Nathan Battin) · hello@ignatiustheyoungerai.com --- FIELD TAXONOMY // AI-ASSISTED WORK # WHERE AI BREAKS. 24 documented failure modes of AI-assisted work. Each entry names the mechanism, the business exposure, the human control that catches it, and the modeled Judgment Multiple. **24** MODES DOCUMENTED **36** PUBLIC SOURCES **6** CONTROLS THAT CATCH THEM SCROLL TO ENTER SPRINT 01 ## The Index 24 CASE FILES · EACH OPENS TO MECHANISM, EXPOSURE, CONTROL, AND MULTIPLE SORT [01 Hallucination JM ~15x to ~58x OPEN CASE FILE 01 →](/failure-modes/1) [02 Knowledge Cutoff JM ~50x to ~208x OPEN CASE FILE 02 →](/failure-modes/2) [03 No True Understanding JM ~10x to ~52x OPEN CASE FILE 03 →](/failure-modes/3) [04 Overconfidence JM ~27x to ~133x OPEN CASE FILE 04 →](/failure-modes/4) [05 Training-Data Bias JM ~0.7x to ~10x OPEN CASE FILE 05 →](/failure-modes/5) [06 Context-Window Limits JM ~13x to ~77x OPEN CASE FILE 06 →](/failure-modes/6) [07 No Common-Sense Grounding JM ~10x to ~85x OPEN CASE FILE 07 →](/failure-modes/7) [08 Prompt Injection JM ~4x to ~108x OPEN CASE FILE 08 →](/failure-modes/8) [09 No Persistent Memory JM ~9x to ~71x OPEN CASE FILE 09 →](/failure-modes/9) [10 Sycophancy JM ~115x to ~167x OPEN CASE FILE 10 →](/failure-modes/10) [11 Reasoning Fragility JM ~16x to ~119x OPEN CASE FILE 11 →](/failure-modes/11) [12 No Real-World Verification JM ~14x to ~137x OPEN CASE FILE 12 →](/failure-modes/12) [13 Verbosity Bias JM ~9x to ~86x OPEN CASE FILE 13 →](/failure-modes/13) [14 Uncertainty Miscalibration JM ~15x to ~147x OPEN CASE FILE 14 →](/failure-modes/14) [15 Training-Data Quality JM ~5x to ~67x OPEN CASE FILE 15 →](/failure-modes/15) [16 Lack Of Accountability JM Not calculable — and say so. OPEN CASE FILE 16 →](/failure-modes/16) [17 Temporal Reasoning JM ~6x to ~100x OPEN CASE FILE 17 →](/failure-modes/17) [18 Mathematical Fragility JM ~21x to ~167x OPEN CASE FILE 18 →](/failure-modes/18) [19 Cannot Truly Cite JM ~19x to ~280x OPEN CASE FILE 19 →](/failure-modes/19) [20 No Sensory Grounding JM ~11x to ~82x OPEN CASE FILE 20 →](/failure-modes/20) [21 Cross-Session Inconsistency JM ~5x to ~35x OPEN CASE FILE 21 →](/failure-modes/21) [22 Difficulty With Negation JM ~19x to ~385x OPEN CASE FILE 22 →](/failure-modes/22) [23 Cultural And Linguistic Blind Spots JM ~15x to ~151x OPEN CASE FILE 23 →](/failure-modes/23) [24 No True Creativity JM ~13x to ~130x OPEN CASE FILE 24 →](/failure-modes/24) JUDGMENT MULTIPLE = ANNUALIZED EXPOSURE ÷ ANNUALIZED COST OF THE HUMAN CONTROL ALL MULTIPLES MODELED — NOT MEASURED SPRINT 02 ## Exhibit: Case File 10 THE ONE YOU CAN READ FREE, IN FULL, RIGHT NOW FAILURE MODE 10 // SYCOPHANCY ### The more certain you sound, the less pushback you get. Preference tuning creates systematic pressure toward agreement. Confident statements receive validation. Expressed doubt prompts reversals. The model is grading your tone when you think it is grading your work. > Pull your last 50 AI-assisted approvals. Count how many the model argued against. If the answer is zero, you don't have a reviewer. You have a rubber stamp. [READ IT FREE →](/preview) [OPEN CASE FILE 10 →](/failure-modes/10) MODELED ANNUAL EXPOSURE MODELED $180K–$260K ANNUAL COST OF THE HUMAN CONTROL $1,560 JUDGMENT MULTIPLE ~115x to ~167x SPRINT 03 ## The Control Layer EVERY ONE OF THEM STILL LANDS ON A HUMAN DESK. YOU ARE STILL NECESSARY. Every failure mode in the index lands on a human desk. The book's 90-day plan turns that into a number you can defend: a logged baseline, a control you have tested, and a Judgment Multiple built from your own figures. Book buyers get each updated edition at no extra charge. MILESTONE 01 30 DAYS A logged error rate for your own AI-assisted work. Measured, not guessed. MILESTONE 02 60 DAYS One control, designed and run and revised, with its time cost measured. MILESTONE 03 90 DAYS Your Judgment Multiple, built from observed inputs, in one page a budget owner reads. THE PLAN THREE MILESTONES · DAYS 30 / 60 / 90 [READ THE PLAN →](/trial) SPRINT 04 ## Take It With You THE FULL TAXONOMY, THE CONTROLS, AND THE 90-DAY PLAN · 206 PAGES ### The Book $29 The AI "Keep Your Career" Bible: all 24 failure modes, the six controls, and the 90-day proof plan. Every updated edition included at no extra charge. [GET THE BOOK · $29](https://agrb01-w9.myshopify.com/cart/43747019161690:1) FREE SAMPLE: [CASE FILE 10, FULL CHAPTER →](/preview) ### Keep Me Sharp One email when there is a verified result worth your time: a new control template, an updated edition, a tool you can use at work that week. DOUBLE OPT-IN. CONFIRM VIA THE EMAIL YOU’LL RECEIVE. EVIDENCE LIBRARY: [36 SOURCES, ALL PUBLIC →](/evidence)