Generative AI does not have a clean border of competence. It is superb at some tasks and confidently wrong at others that look almost identical, and nothing on the screen tells you which side of that line you are standing on. Researchers at Harvard Business School and Boston Consulting Group called this the jagged technological frontier, and in 2023 they measured what happens when professionals cross it without noticing.

Nearly 800 BCG consultants were randomly assigned to work with or without GPT-4. On creative product-development tasks inside the model’s range, the AI users finished more work, faster, and produced output that graders rated markedly better. A separate task was built to sit just outside that range: recommending where a company should invest by combining spreadsheet figures with interview notes, where the right answer hinged on a detail in the interviews that the numbers hid. There, consultants using GPT-4 were about 19 percentage points less likely to reach the correct recommendation than those working alone. The model’s answer was fluent and plausible, and many of them adopted it.

The lesson is not that the model is bad; it is that a good model lowers your guard. Fabrizio Dell’Acqua, the study’s lead author, had earlier found the same pattern with recruiters: people handed a more accurate AI screened résumés less carefully than people handed a weaker one, and made worse calls. Your judgment matters most exactly where the tool’s help feels most convincing, so decide in advance which kinds of work you will verify line by line.