分岔路口 相同目標，不同選擇

Every organization here that used AI to support its agents kept the tools in use. Bots that answered customers first had mixed results. Of the five that cut agent jobs because of AI, Klarna, Commonwealth Bank and KB Kookmin each reversed part of the cut; Salesforce and MONETA say their AI kept the work going.

Both ways mostly kept running. The one that was shut down, New York City's MyCity, generated answers to legal questions from businesses, and some of its answers broke city law.

Drive-thru voice ordering had mixed results: McDonald's ended its test with IBM and later tried another system, and Taco Bell began rethinking where AI should take orders alone. The phone-ordering cases here are few and still running.

All four retailers that used AI to pick markdowns kept it. AI for regular prices had mixed results: Amazon, Instacart and Greystar ended theirs after a regulator, a study or a lawsuit questioned what customers or tenants were charged, and Zillow closed its home-buying business after paying model-set prices it could not recover.

Most organizations that put AI-generated images in front of their audience withdrew or replaced them after complaints. AI-written copy mostly stayed, though two promotions were pulled. Those that used AI only for planning, with people making the final work, all kept going.

Three of the four that put synthetic voices in front of listeners pulled back after complaints: Prime Video removed some anime dubs, ScotRail replaced its disputed AI announcer, and OFF Radio Kraków ended its AI presenters within a week. AI translation of written text mostly stayed in use; 11 bit studios had professionals redo AI translations left in its game.

Aldi, Sainsbury's and Amazon removed checkout-free technology from grocery stores, while Konzum Plus and Żabka still run such stores. Retailers that added AI to existing checkouts or carts kept it in use or expanded it.

Privacy regulators ruled against the four retailers that installed facial recognition across stores before any privacy review, although a tribunal later overturned the Bunnings finding. New Zealand's Privacy Commissioner reviewed the Foodstuffs North Island trial and found its safeguards acceptable.

Every police force here that used AI to predict crime or flag people stopped, most after audits or reviews questioned whether it worked. Tools for material already in hand and for paperwork mostly stayed in use, though West Midlands Police's Copilot error shows their output still needs checking.

All four police forces that tried Clearview stopped, and privacy regulators found three of them in breach. Street-camera scanning had mixed results: courts stopped it in South Wales and Buenos Aires, while London kept its cameras. Matching faces against the state's own records continued in two of three cases.

None of the risk-scoring tools here is still in use. Note-taking and search tools mostly stayed; in Victoria, the privacy regulator ordered child-protection staff off generative AI after a caseworker used ChatGPT to draft a court report.

The three hospital projects built on a general advisor (MD Anderson, Rhön-Klinikum and Birmingham's Babylon service) all ended; two newer ones that check doctors' notes or suggest diagnoses in telemedicine report working. Narrow tools mostly stayed; Epsom and St Helier's ended when its pilot funding ran out. Note drafting with tools the hospital provided stayed in use; the one failure was a doctor using a public chatbot.

Four of the five that used AI to watch for cheating dropped or restricted it, after doubts about accuracy, privacy or fairness. Marking with teachers checking, and tutoring, mostly stayed in use; Korea's AI textbooks lost their status as textbooks after one semester.

Every driverless bus service here on public routes ended or was suspended, often when funding, the supplier or passengers fell away. Inside closed sites such as ports, mines and airports, driverless vehicles mostly kept running. Of the five delivery services between places, three were stopped or cut back; driverless trucks on fixed routes kept running.

Algorithms that decided riders' and drivers' shifts, deadlines or account access led to fines in Italy and the Netherlands, and Meituan loosened its delivery times after a press investigation. Algorithms that planned routes and loads stayed in use.

Forecasts that fed routine ordering and production stayed in use everywhere here. The two failures came where a forecast backed a large commitment: Zillow bought homes at model-set prices and closed the business, and Sushiro ran short of promoted items and received an advertising order.

Every court sanction here followed citations that nobody checked before filing. Firms that rolled out a chosen tool reported hours saved, but Morgan & Morgan's lawyers were still fined over cases invented by the firm's own AI tool.

Models that scored individual welfare claimants were stopped or challenged over bias in the Netherlands, France, Sweden and Denmark; the UK's DWP says its model for benefit advance requests works with human review of flagged cases. Models that scored tax returns and employer contributions in Belgium, France and Spain stayed in use, while an audit questioned the governance of Australia's expense-claim models.

Both ways mostly stayed in use, by the employers' own accounts. The three scoring tools that ended judged people on something other than the job: Amazon's learned to mark down women, iTutorGroup's rejected older applicants, and Rikunabi sold predictions of whether students would turn down job offers.

Four of the six companies that trained models on customers' photos, recordings or chats faced regulators and had to pay, delete data or models, or shut the service. Generated training data and the companies' own records drew no such action; Lionsgate's studio-only video model ran into problems, with copyright questions unresolved.

All five organizations whose staff used public chatbots for work stopped or restricted that use. Organizations that gave staff their own secured assistant mostly kept it; Societe Generale retired SoGPT after employees found it lagged commercial assistants.

Seven of the nine publishers whose AI produced whole pieces withdrew, corrected or paused the work; Il Foglio ran its openly labelled AI edition for the planned month, and Network18 still publishes AI match reports under human supervision. Newsrooms where reporters gathered the facts and AI helped with drafts and tags kept the tools in use.

Seven of the eight that used AI to write text going to clients or readers had to withdraw or correct it, refund fees or answer a union arbitration; PwC reports its AI-drafted client findings worked. Firms that used AI to search and summarize research for their own staff kept it.