Ford Spent Billions Learning What a German Court Just Made Law: Your AI's Mistakes Are Your Mistakes.

"The AI made an error" is not a legal defence. A Munich court ruled as much in June, treating Google's AI-generated false claims as Google's own speech — not neutral search results, not a vendor's liability, not a tool's mistake. Ford had already paid for the same lesson in dollars: billions in quality costs and 350 engineers rehired after AI inspection systems shipped without the people who knew what good actually looked like.

These aren't abstract warnings about AI risk. They're specific liability cases, and the principle the court articulated applies to anyone whose name is on output that AI helped produce.

What happened

Ford's VP of hardware engineering was direct about it: "Mistakenly, we thought that by just introducing AI and adjusting the design requirements, that would produce a high-quality product." The result — Ford now leads US automakers in recalls in 2026, with 51 recalls covering more than 11 million vehicles (Hacker News, June 25, 601 points). The engineers who could evaluate whether the AI's quality judgements were actually correct had been removed. By the time the defects appeared, they were already in production.

The Munich Regional Court ruling adds legal weight to what Ford learned operationally. The court held Google liable for false claims in its AI Overviews, because the output is treated as Google's own content rather than a neutral tool's. Bruce Schneier's summary of the principle: "AI agents are agents of the person or organization that deploys them." If AI drafts a false claim in your report, proposal, or client communication, the accountability sits with you.

Timothy B. Lee offered the clearest reframe (via Simon Willison, June 26): "This is like saying there's no learning curve to being a manager because your employees will just do whatever you tell them to do." Managing AI output is a learnable skill. Delegating without it produces the same result as delegating to an employee you never manage.

What this demands on Monday

Three things, for anyone whose AI output reaches the outside world.

Keep the evaluators. Ford's error wasn't deploying AI for quality inspection — it was removing the engineers who could tell whether the inspections were correct. You can automate the generation. You cannot automate the evaluation without eventually producing something you can't afford to defend.

Review AI output with the same rigour as an employee's. Courts now treat them as equivalent. If you wouldn't send a new hire's report to a client without reading it, the same standard applies here — and if you would approve an employee's document after a five-second glance, that's a separate problem.

Write quality criteria before you prompt. "Looks good" is not a review threshold. Define what a correct output looks like before you start, so you're evaluating against something concrete rather than against your own optimism.

Rachel: the statistic that was technically there

Rachel is a senior analyst at a 25-person strategy consultancy. Her firm uses AI to generate first-pass research summaries from uploaded source documents.

Three months ago a summary reached a client containing a statistic that was technically in the source but presented out of context. It wasn't a fabrication. It was a selective reading the model produced without malice, and she hadn't checked it closely enough. The client caught it.

She now runs every AI summary against five questions before it leaves the firm: Is every number traceable to a source? Is every causal claim supported or qualified? Is the framing consistent with what the sources actually argued? Is anything stated with more certainty than the source warrants? Is there a missing caveat the reader would want?

Running the checklist takes less time than writing the apology email.

Nadia: the citations that had quietly expired

Nadia is a compliance manager at a 180-person financial services firm. Her team uses AI to draft regulatory update summaries that go to 40 client firms each week. The summaries are fast and read clearly.

After the Munich ruling she audited the last twelve. Three contained regulatory references that were accurate as of the model's training data and had since been superseded. The clients had received them with no flag attached.

She added one checkpoint: every regulatory citation is verified against the official register before the summary goes out. Twenty minutes per summary. The alternative — a client relying on outdated guidance, with her firm's name on the document — carries a consequence she now owns by name.

The one thing

Ford's VP said they thought AI would produce a high-quality product. A German court said that when it doesn't, the organisation that deployed it is responsible.

The AI will keep making errors. The only question is whether your process catches them before they become your errors.