An EU Court Adviser Says Chatbot Answers Can Infringe Press Rights, Alexa for Shopping Gets 100+ Account Actions, and Hiker Rescues Show the Cost of Confident Wrong Answers

Three stories from the past week test where a chatbot's responsibility starts: for the content in its answers, for the actions it takes on an account, and for advice that is wrong in a way that matters outdoors. We could not open every original article from our environment, so details come from search summaries and secondary coverage, and we flag where sources are thin.

1. An EU Court Adviser on Chatbot Answers and Press Rights

On November 3, Advocate General Maciej Szpunar issued a non-binding opinion in Case C-250/25, Like Company v. Google Ireland, according to multiple news reports. The case came from a Hungarian court and concerns Gemini and the EU press publishers' right. As reported, his view is that a publisher's right may be infringed when a chatbot retrieves an article, reuses it to write an answer and delivers it to the user, depending on circumstances. Coverage says the assessment could turn on whether answers keep recognizable parts of the article's original expression often enough to substitute for visiting the publisher's page, and on whether access was unlawful, such as getting around a paywall.

Two details matter for builders. Copyright protects expression, not the facts a story reports, so a short factual answer is in a different position from a near-copy of the prose. And reports say he treated training on copyrighted material as permitted under the text and data mining exception, and found no evidence Gemini was trained on the claimant's articles. Google reportedly said Gemini learns language patterns rather than copying content and gives publishers control over use of their material. The opinion is not binding, though judges often follow them, and the Court of Justice has not ruled. Read the opinion itself before acting on any summary.

If your chatbot does retrieval or web search, the risk sits in the output step, not the model. A practical guard is to cap how much of a retrieved page can appear in an answer and to prefer paraphrase plus a link:

// Flag answers that copy long runs of a retrieved source.
function longestSharedRun(answer: string, source: string, n = 12): boolean {
  const words = source.split(/\s+/);
  for (let i = 0; i + n <= words.length; i++) {
    if (answer.includes(words.slice(i, i + n).join(' '))) return true;
  }
  return false;
}
// If true: regenerate with "paraphrase, do not quote" or show a link instead.

2. Alexa for Shopping Moves from Finding Things to Changing Things

Amazon added more than 100 new actions to Alexa for Shopping for U.S. customers, according to several outlets summarizing its announcement. Reported examples include creating and editing lists and registries, pausing or changing the frequency of Subscribe & Save deliveries, updating delivery instructions, and asking about billing, refunds and gift card balances. Coverage ties the timing to holiday shopping. We could not open Amazon's own post, so check it for the full list and any confirmation steps.

The shift is from an assistant that answers to one that edits an account. That changes the engineering. A wrong answer can be corrected in the next message; a cancelled subscription or a changed address may not be noticed until a package goes missing. For any action that changes state, show what will change before doing it, make it undoable where you can, and log who or what triggered it. Reads and writes should be separate tool permissions, not one broad one.

3. Hiker Rescues and the Confidence Problem

Several outlets reported rescues of hikers who had relied on chatbots for route planning, including two Lithuanian hikers in Poland's Tatra Mountains who reportedly followed a ChatGPT shortcut, and three hikers on Mount Shasta who used Gemini. Google reportedly said it could not reproduce the Shasta advice, and OpenAI pointed to its guidance that ChatGPT can sound confident when wrong. Researchers quoted in coverage say chatbots have no built-in way to signal uncertainty, so a bad route reads like a good one. These are secondhand reports, and the incident details vary by source.

The lesson is about domains where the cost of a confident error is physical. If your product can be asked about safety-critical topics, add a visible check: state which facts the answer depends on (current snow, trail closures), say plainly that the model cannot see them, and point to a live source. A system prompt that only says "be careful" does little; a rule that routes these questions to a template with a live-conditions warning does more.

What to Take Away

Responsibility is widening in three directions: to the text you reproduce, the state you change and the physical outcomes your advice can cause. Cap quoted material, gate writes with confirmations, and add live-source warnings in high-stakes domains.

Suggested visual: a chart sorting chatbot features by reversibility, from reading to account changes. Suggested code sample: a confirm-before-write wrapper for tool calls.

— Maya