How AI fails, and what it costs you

AI systems fail in patterned ways, and knowing the patterns is more useful than knowing the underlying technology. Fabrication is the most familiar: confident, specific, wrong, and most common for citations, statistics, regulations, dates, and anything niche. Staleness is the quiet one: the tool's knowledge has a cutoff, so recent rule changes, prices, and product details may be confidently out of date. And bias shows up as reproduced assumptions — about who does which job, which names sound professional, which neighborhoods are desirable — because the patterns in its training data include ours. Business consequences follow directly. Fabricated compliance details create legal exposure. Stale pricing or policy text creates customer disputes. Biased phrasing in job postings, candidate screening, or tenant and lending communication creates discrimination risk that is both legally serious and genuinely wrong. Any use of AI touching hiring, housing, credit, insurance, or health deserves a heavier review gate and, in some jurisdictions, specific legal duties. The defenses are unglamorous and effective. Verify unsupplied specifics before they reach a customer. Supply current facts yourself rather than trusting recall for anything time-sensitive. Read output for who is assumed and who is excluded, especially in anything describing people. Where language matters most, use structured neutral templates rather than free generation. And keep a short log of the errors you catch: after a month it becomes a review checklist tuned to how this tool actually fails on your work, which is far more useful than a general warning to be careful. Staleness is more dangerous than fabrication in regulated work, because it is harder to spot. A rule that was correct last year reads as correct today, and nothing in the output signals that it has changed. Bias in business writing usually appears as an assumed customer rather than an offensive statement. Describing the product instead of the intended person removes most of the exposure without any loss of persuasiveness. A recruiting firm reviewed a month of AI-drafted job posts and found repeated phrases implying a preferred age range. Adding a single prohibition to the standing prompt eliminated the pattern, and their error log became the review checklist.

Key takeaways

• Business consequences follow directly.
• The defenses are unglamorous and effective.
• Precise, checkable, unsupplied specifics are where invented content appears most often and does the most damage.

Sign in to complete activities, save progress, and use the AI Coach.