Julian Pintat, a German translator from freelance English, has seen him gradually unravel his 15-year career. His expertise, specialized in high stakes fields such as medical technology and pharmaceuticals, has been re-released as an AI cleanup service. In recent work, I translated the operational manual for oil drilling rigs, and AI mistranslated “scale” (mineral accumulation) as both a musical scale and a device for measuring weight. He says that fixing these basic flaws, which now account for 95% of his work, often takes longer than translating from scratch. It's an irritating reality of halving his income, planning a life plan including marriage, and starting a family indefinitely. Professional translators have been feeling the effects of artificial intelligence for almost a long time, as Google's translation and subsequent Deepl exploded into the scene a few years before ChatGpt. “I'm a coal mine canary,” Pintat says.
AI is changing the face of work, and Pintat is one of many people looking at a very different future. There have been a lot of conversations about technology that some CEOs like Ford's Jim Farley and Amazon's Andy Jussy predict will be wiped out by AI in the coming years, but the first wave of AI adoption has already changed the contours of new roles and the contours of their work.
For some companies, AI is increasing efficiency. London-based law firm A&O Shearman has allowed employees to effectively multiply and take over a project that refuses, says David Wakeling, global head of partners and global AI groups. To help major US banks comply with European law, the company scanned its 20-year licensing agreement and built a tool that identified what needed to be revised. “Two years ago we had 20 lawyers in our room, and we were probably a few paralyzed,” Wakeling said. But the tool reduces 2,400 regulatory requirements to 900, and cuts project costs by half, even when it takes up time to build the tool. Still, he softens his optimism. He says that basic ready-made AI assistants probably don't add much value. He states that actual results often require a customized or specialized solution. “We need a lot of grease on our elbow,” Wakeling said.
Meanwhile, AIG CEO Peter Zaffino told Time in June that insurers will use AI to do underwriting tasks faster. It trains a system to become a “junior underwriter” who can do most of the underwriting, allowing “more experienced practitioners” to do the rest. “Part of the cultural change is reskilling positions in a new world that allow us to be more productive than in the past,” he said.
A MIT report published in August concluded that 95% of AI pilots could not provide a return on investment. Even the AI coding assistant, which has been maintained as an AI victory use case, is called suspicious by a small, preliminary July investigation published by Berkeley-based AI research group Metr. A sample of 16 experienced developers was 19% slower if they estimate that using AI was 20% faster on average. It's spurred by excitement and the fear of perhaps missing out. This creates a gap between what AI can do and actual performance, and for both businesses and translators, squeezes workers trapped between the promise of superhuman efficiency and the reality of often flawed mechanical output.
The rapid advances in AI for white-collar tasks could quickly erase that gap. According to a preliminary survey released by data company Mercor in late September, the performance of tests, devised by veteran experts across banks, law and consulting, doubled in just one year. This paper appeared shortly after another report written by Openai. This attempted to measure the ability of AI to perform real tasks by comparing machine and human performance in blind tests. The best models were found to be compared by human authors almost half the time. However, both reports note that such tests measure performance of well-painted tasks. So, for now, AI models can make human workers' alternatives poor, and implementation remains important.
read more: AI learns to work as a doctor, lawyer, consultant
“Generative AI is a really great demonstration,” says Kaitlin Elliott, head of corporate all-generated AI solutions at Morgan Stanley, but making it convenient is more difficult than it looks. The bank has built its own meeting transcription and summary tool that Elliott says will save time on grunts. They also have AI-driven search tools that make it easy for staff to view information. “In the beginning, we thought we could give it all the knowledge, so we thought we could give it a very accurate response,” she says, but in reality it required well-structured data and careful testing.
The challenge of implementing AI is to create a demand for a new kind of expertise. “We still need technical skills,” Elliott says. Elliott says that AI tools usually have automated tasks given to junior staff, but younger generations are expected to bring AI skills to the organization. “They're all AI adopters. They know how to use it effectively,” she says. That demand is also creating opportunities for companies that sell their know-how as services such as scale AI, best known for their data labeling business, but this has expanded to support companies, including Morgan Stanley, in AI applications. When employing AI, it's important to work backwards from the problem, says Felix SU, engineering director at Scale Enterprise AI ARM. Applying AI for your own backfires. Su shows one example of Scale's client who built four chatbots for slightly different tasks, forcing staff to always copy and paste. Su means identifying a solution means applying generative AI, but often it means using traditional machine learning or software engineering.
The improvements are coming to translation. Deepl has provided features such as creating custom glossary, and this year we have introduced a tool that asks follow-up questions to clarify ambiguities that could be useful in niche domains. AI-powered translators are running faster, according to Deepl CEO Jaroslaw Kutylowski, but the company only offers numbers that compare the tool to its AI rivals. With improved performance, translators can offset lower costs per word by translating higher volumes of text, but Kutylowski points out that AI tools allow translations to be brought in-house, rather than outsourced to experts. “I think we're just being robust here with this civilization's ladder,” he says, acknowledging that it makes a difference in how people do their jobs. “It's a change we have to go through,” he adds.
Elsewhere, businesses aim to lower the technical bar for businesses to adopt AI by creating off-the-shelf enterprise solutions. For example, most virtual conference products incorporate AI summaryzers. A&O Shearman says Wakeling is useful for general questions, but while the company uses its own AI tool, ContractMatrix, for niche questions, especially Wakeling is useful for general questions. Also a new kind of enterprise-centric tools that are drawn from internal documents, slack messages, and emails to answer queries.
Last month, Canadian AI Compery released North North, acquired its own tool last month. The company's co-founder Nick Frossst says he's not sweating the final meeting as he uses tools to make easy preparations for individuals based on the entire company's history in seconds. North is currently working on 90% of the typical support tickets, but human operators are still in the loop. It's not just used internally because RBC, Canada's largest bank, uses the platform. (Salesforce, whose Time Co-Chair and owner Marc Benioff are CEOs, is an investor in Cohere.)
While Openai, Anthropic, and Google Deepmind bosses all believe that they are just a few years away with what is known as artificial intelligence or AGI (a system that can automate most human work), Frosst's outlook is relatively conservative. He doesn't think he'll use anything similar to current technology to reach AGI. Still, he says that even without AGI, the impact on labor is disruptive and compares it to the Industrial Revolution. “When there was a major shift in the labour market, much of what was resolved was at the union level, the government level,” he says. “This is a cross-individual issue and needs to be addressed as a group.”
